# Numonic AI

> Numonic is an AI-native digital asset management platform. It captures the full generation envelope around every AI asset — prompt, model, seed, workflow graph, lineage, and compliance metadata — so creative teams can find, reproduce, and prove the provenance of work made with ComfyUI, Midjourney, Stable Diffusion, and other generative tools.

## Product & company

- [Numonic — AI-native digital asset management](https://www.numonic.ai/): Home: the AI creative system of record — capture, organise, and prove the provenance of generative AI assets.
- [Pricing](https://www.numonic.ai/pricing): Seat-based pricing. Every plan (Free included) gets the full product — library, lineage, prompt versioning, C2PA exports, REST API and MCP server. You pay for scale, governance, and support, not for features.
- [Product overview](https://www.numonic.ai/product): How Numonic captures the generation envelope: metadata, lineage, semantic search, collections, collaboration, and privacy-safe export.
- [ComfyUI workflow management](https://www.numonic.ai/product/comfyui): Version ComfyUI workflows, preserve PNG metadata, and track team collaboration with compliance-ready provenance.
- [Midjourney library management](https://www.numonic.ai/product/midjourney): Organise Midjourney outputs with prompt, parameter, and evolution-chain tracking across thousands of generations.
- [Asset lineage](https://www.numonic.ai/product/lineage): Navigate the full chain of custody for every asset — from prompt through every variation, upscale, and edit.
- [Semantic search](https://www.numonic.ai/product/search): Find assets by meaning, visual similarity, prompt, model, or metadata across the whole library.
- [Solutions for solo creators](https://www.numonic.ai/solutions/creators): Tame thousands of AI outputs as a single operator, with the same API and MCP access as teams.
- [Solutions for studios](https://www.numonic.ai/solutions/studios): Multi-client asset organisation, lineage tracking, and audit trails for creative studios.
- [Solutions for enterprise](https://www.numonic.ai/solutions/enterprise): Governance, roles, approvals, audit retention, and roadmap controls (SSO, data residency) for large teams.
- [Solutions for compliance](https://www.numonic.ai/solutions/compliance): EU AI Act and California SB 942 disclosure, provenance signals, and C2PA-signed exports with a full audit trail.
- [Company](https://www.numonic.ai/company): About Numonic Labs and the team building the AI creative system of record.
- [Trust center](https://www.numonic.ai/trust): Security, privacy, and data-handling commitments across all plans.
- [Blog](https://www.numonic.ai/blog): Articles on AI asset management, creative workflows, provenance, and compliance.
- [Glossary](https://www.numonic.ai/glossary): Definitions for AI asset management, content provenance, and regulatory compliance terms.
- [Guides](https://www.numonic.ai/guides): In-depth pillar guides to organising and governing AI-generated content.
- [Free tools](https://www.numonic.ai/tools): Free interactive utilities: metadata inspectors, an EU AI Act disclosure generator, and library calculators.
- [Integrations](https://www.numonic.ai/integrations): ComfyUI node reference and integration guides for the generative tools Numonic supports.
- [API reference](https://www.numonic.ai/docs/api): REST API reference for Numonic — authentication, endpoints, and the OpenAPI spec at /api/v1/openapi.json.
- [MCP server reference](https://www.numonic.ai/docs/api/mcp): Connect an AI agent to Numonic over MCP — tool reference and the machine-readable spec at /api/v1/mcp-spec.json.

## Guides (pillar pages)

- [The Complete Guide to ComfyUI Asset Management](https://www.numonic.ai/guides/comfyui-asset-management): Master ComfyUI asset management: workflow versioning, PNG metadata preservation, team collaboration, and compliance-ready provenance tracking.
- [How to Organise AI-Generated Images: The Complete Guide](https://www.numonic.ai/guides/organise-ai-generated-images): The solo creator’s guide to taming thousands of AI outputs. Learn naming conventions, folder systems, tagging strategies, and multi-tool workflows.
- [AI Content Compliance for Agencies: From Governance Bottleneck to Competitive Advantage](https://www.numonic.ai/guides/ai-content-compliance): The agency guide to EU AI Act Article 50, California SB 942, IPTC 2025.1, C2PA metadata, and compliance workflows that accelerate creative velocity.
- [AI-Native DAM Architecture: A Pattern Language for Generative Asset Management](https://www.numonic.ai/guides/ai-native-dam-architecture): A guide to building DAM systems for AI-generated content — search, lineage, compliance, and curation patterns.
- [Organise Your Midjourney Library: The Complete Post-Export Guide](https://www.numonic.ai/guides/midjourney-organisation): Master Midjourney image organisation beyond the web app. Export strategies, metadata preservation, legacy back-catalogue recovery, style reference management, and compliance-ready workflows for 5,000+ image libraries.

## Glossary

- [AI Content Provenance](https://www.numonic.ai/glossary/ai-content-provenance): A verifiable record of an AI-generated asset's origin, including the model, prompt, parameters, and workflow used to create it. Provenance enables reproducibility, compliance with regulations like the EU AI Act, and proof of creative ownership.
- [C2PA](https://www.numonic.ai/glossary/c2pa): The Coalition for Content Provenance and Authenticity (C2PA) is an open technical standard for certifying the origin and history of digital content. It embeds cryptographic signatures into files to verify whether content is human-created, AI-generated, or edited.
- [AI Metadata](https://www.numonic.ai/glossary/ai-metadata): Structured information embedded in or associated with AI-generated content that describes how the content was produced. This includes model identifiers, prompt text, generation parameters (seed, steps, CFG scale), and workflow configurations.
- [Workflow Reproducibility](https://www.numonic.ai/glossary/workflow-reproducibility): The ability to recreate an AI-generated asset by replaying its original workflow with identical parameters. Reproducibility requires preserving the complete generation context: model version, seed value, sampler settings, and node configurations.
- [DAM (Digital Asset Management)](https://www.numonic.ai/glossary/dam): A system for organizing, storing, retrieving, and distributing digital files. Traditional DAM platforms manage photos, videos, and documents with metadata tagging, version control, and access permissions. AI-native DAMs extend this with provenance tracking and workflow capture.
- [EU AI Act](https://www.numonic.ai/glossary/eu-ai-act): The European Union's comprehensive regulation governing artificial intelligence systems. Article 12 requires providers of high-risk AI systems to maintain detailed logging and audit trails. The Act establishes transparency requirements for AI-generated content, including disclosure obligations.
- [ComfyUI Workflow](https://www.numonic.ai/glossary/comfyui-workflow): A node-based visual programming graph in ComfyUI that defines the complete image generation pipeline. Each node represents an operation (model loading, sampling, conditioning, post-processing), and connections between nodes define the data flow from input to output.
- [AI-Native DAM](https://www.numonic.ai/glossary/ai-native-dam): A digital asset management system designed from the ground up for AI-generated content. Unlike traditional DAMs adapted for AI workflows, an AI-native DAM automatically extracts metadata, tracks provenance, enables workflow reproducibility, and supports compliance requirements.
- [Prompt Metadata](https://www.numonic.ai/glossary/prompt-metadata): The text prompts, negative prompts, and associated generation settings captured alongside an AI-generated asset. Prompt metadata enables search, categorization, and reproduction of creative outputs across teams and workflows.
- [AI Audit Trail](https://www.numonic.ai/glossary/ai-audit-trail): A chronological record of all operations performed by or with an AI system, including inputs, outputs, configuration changes, and user interactions. Required by EU AI Act Article 12 for high-risk AI systems and increasingly expected for all commercial AI use.
- [IPTC 2025.1](https://www.numonic.ai/glossary/iptc-2025-1): The 2025.1 revision of the IPTC Photo Metadata Standard that introduces dedicated fields for AI-generated content. New fields include AISystemUsed, AISystemVersionUsed, and DigitalSourceType, providing a standardized vocabulary for recording how AI tools contributed to content creation.
- [Shadow AI](https://www.numonic.ai/glossary/shadow-ai): The use of AI tools by employees without organizational knowledge, approval, or governance. Similar to shadow IT, shadow AI creates compliance risks when staff use unapproved generative tools for client work without proper metadata tracking, disclosure, or IP verification.
- [AI Deployer](https://www.numonic.ai/glossary/ai-deployer): Under the EU AI Act, a natural or legal person that uses an AI system under its authority, as distinct from the AI provider who develops or markets the system. Most creative agencies are deployers, which carries specific transparency, monitoring, and record-keeping obligations.
- [SB 942 (California AI Transparency Act)](https://www.numonic.ai/glossary/sb-942): California Senate Bill 942, effective January 2026, requiring providers of generative AI systems to offer provenance tools and users of covered AI systems to disclose AI-generated content in commercial contexts. Applies to systems with over one million monthly users.
- [AI Content Disclosure](https://www.numonic.ai/glossary/ai-content-disclosure): The practice of transparently communicating to audiences that content was created with or by artificial intelligence. Disclosure requirements vary by jurisdiction—mandated by EU AI Act Article 50, California SB 942, and increasingly expected by platform policies and industry standards.
- [Content Credentials](https://www.numonic.ai/glossary/content-credentials): A consumer-facing implementation of the C2PA standard that displays a tamper-evident provenance record for digital content. Content Credentials show whether an image, video, or audio file was created by a human, generated by AI, or edited, along with the tools used.
- [Content-Addressed Storage](https://www.numonic.ai/glossary/content-addressed-storage): A storage model that identifies files by a cryptographic hash of their content rather than by file path or name. Two files with identical binary content produce the same hash and are stored once, eliminating duplicates regardless of how many times they are imported or where they appear in a library.
- [Ingest Pipeline](https://www.numonic.ai/glossary/ingest-pipeline): The multi-stage processing system that transforms a raw uploaded file into a fully indexed, searchable asset. Stages include content hashing for deduplication, tool-specific metadata extraction, thumbnail generation, and asynchronous enrichment such as embedding generation and session clustering.
- [Embedding Space](https://www.numonic.ai/glossary/embedding-space): A high-dimensional mathematical space where images and text are represented as numerical vectors such that semantically similar content occupies nearby positions. The foundation of semantic search — querying by meaning rather than exact keywords — with cosine similarity measuring relatedness between vectors.
- [Hybrid Search](https://www.numonic.ai/glossary/hybrid-search): A retrieval architecture that combines structured metadata queries (exact filters on tool, date, model) with vector similarity search (semantic meaning, visual resemblance) in parallel, then fuses the ranked results into a single list. Enables queries that are simultaneously precise and semantic.
- [Creative Session Clustering](https://www.numonic.ai/glossary/creative-session-clustering): The automatic grouping of AI-generated assets into discrete creative sessions by detecting boundaries from temporal gaps, parameter changes, and tool switches. Provides retroactive organizational structure without manual tagging — a session boundary is inferred when the generation rhythm pauses.
- [Metadata Inversion](https://www.numonic.ai/glossary/metadata-inversion): The architectural shift where AI-generated assets arrive with rich generation metadata already embedded by the creation tool, inverting the traditional DAM assumption that metadata must be added by humans after upload. The challenge shifts from metadata creation to metadata capture and normalization.
- [Metadata Normalization Pipeline](https://www.numonic.ai/glossary/metadata-normalization-pipeline): A three-stage processing layer that translates tool-specific metadata formats from different AI tools (ComfyUI JSON, Midjourney Discord strings, DALL-E API responses) into a unified canonical schema. Enables cross-tool search and comparison from a single query interface.
- [The Two Metadata Problem](https://www.numonic.ai/glossary/two-metadata-problem): The absence of a common metadata standard across AI generation tools. Each tool uses its own storage location, format, field names, and encoding — ComfyUI stores JSON in PNG chunks, Midjourney uses Discord message strings, DALL-E returns API response objects — making cross-tool search architecturally difficult.
- [Generation Lineage](https://www.numonic.ai/glossary/generation-lineage): The chain of parent-child relationships between AI-generated assets — from an initial generation through variations, upscales, and refinements. Unlike code version control with explicit commits, image lineage must be inferred from parameter mutations, temporal proximity, and visual similarity.
- [Collection Branching](https://www.numonic.ai/glossary/collection-branching): Creating lightweight copies of asset collections that maintain lineage to the original without duplicating underlying files. Analogous to Git branches for creative projects — enables parallel workstreams, versioned client deliveries, and safe experimentation while avoiding file multiplication.
- [Portfolio Distillation](https://www.numonic.ai/glossary/portfolio-distillation): The process of progressively filtering a large generative asset library down to a curated portfolio — typically the top 2% of output representing the best work. Uses layered signals including quality scoring, session highlights, and deliberate aesthetic selection.
- [Semantic Search](https://www.numonic.ai/glossary/semantic-search): A retrieval method that finds content based on conceptual meaning rather than exact keyword matches, by comparing vector embeddings in high-dimensional space. Enables finding images by visual or conceptual similarity even when query terms share zero keywords with the stored metadata.
- [Temporal Search](https://www.numonic.ai/glossary/temporal-search): Querying an asset library using time-based expressions — "what I made last Tuesday," "images from the brutalist architecture session," or "everything from the first three weeks of the project." Treats time periods and creative sessions as first-class navigation dimensions.
- [Cross-Tool Provenance](https://www.numonic.ai/glossary/cross-tool-provenance): The challenge of maintaining a continuous provenance chain when creative work flows across multiple AI and traditional tools — from ComfyUI to Photoshop to Midjourney — where each tool has its own metadata format and no awareness of what preceded it.
- [Metadata Stripping](https://www.numonic.ai/glossary/metadata-stripping): The removal of embedded metadata from image files before distribution. Creates a tension for AI-generated content: stripping protects creative process privacy and prevents prompt disclosure, but removes the provenance chain that regulations require and audit trails depend on.
- [LoRA (Low-Rank Adaptation)](https://www.numonic.ai/glossary/lora): A technique for fine-tuning large AI image generation models on a small dataset to learn a specific style, subject, or concept. LoRA files modify a base model's outputs and require their own metadata management, versioning, and provenance tracking as they directly affect generation character.
- [Model Context Protocol (MCP)](https://www.numonic.ai/glossary/model-context-protocol): An open protocol that allows AI agents and language models to interact with external tools and services through a standardized interface. In creative workflows, MCP enables AI assistants to search, organize, and manage assets in a DAM system directly — making the library agent-accessible.
- [Agent-First API Design](https://www.numonic.ai/glossary/agent-first-api): An interface design philosophy where API endpoints are built primarily for machine callers (AI agents) rather than exclusively for human users through a UI. The user interface becomes one consumer among many, changing how operations are scoped, errors are communicated, and tool contracts are specified.
- [Deduplication](https://www.numonic.ai/glossary/deduplication): The identification and elimination of duplicate assets in a library. Exact deduplication uses content hashing to detect byte-identical files stored under different names. Near-duplicate detection uses visual embedding similarity to identify perceptually identical images that differ at the binary level.
- [Automatic Curation](https://www.numonic.ai/glossary/automatic-curation): Using quality signals, behavioral patterns, and visual analysis to surface the highest-value assets from a large generative library without manual rating. Signals include technical quality scores, selection frequency, within-session diversity, and aesthetic assessments from embedding-based classifiers.
- [Style Reference (--sref)](https://www.numonic.ai/glossary/style-reference): Midjourney's --sref parameter applies a consistent visual aesthetic from a reference image or saved style code to new generations. Style references enable brand consistency across prompts by decoupling visual identity from prompt text, allowing teams to enforce approved aesthetics at scale.
- [Variation Lineage](https://www.numonic.ai/glossary/variation-lineage): The branching history of how an AI-generated image evolved through successive variations, upscales, and remixes from its original generation grid. Lineage tracking preserves the parent-child relationships between outputs so creators can trace any final asset back to its initial prompt and understand the creative decisions at each branch point.
- [Batch Export](https://www.numonic.ai/glossary/batch-export): Downloading multiple AI-generated images simultaneously as a ZIP archive from a platform like Midjourney. Batch exports introduce specific failure modes at scale — ZIP timeouts above 5,000 images, filename collisions causing silent overwrites, and inconsistent metadata embedding — that make post-export verification essential.
- [Prompt Library](https://www.numonic.ai/glossary/prompt-library): A structured, searchable collection of generation prompts with their parameters, output examples, and performance history. Unlike simple prompt lists in Notion or spreadsheets, a proper prompt library maintains bidirectional links between prompts and their generated outputs, enabling retrieval by visual result rather than text recall alone.
- [Style Drift](https://www.numonic.ai/glossary/style-drift): Gradual visual inconsistency that emerges when multiple team members generate AI images without shared style governance. Style drift occurs when creators use slightly different parameters, unapproved style references, or ad-hoc prompt modifications that individually seem minor but cumulatively produce a fragmented brand aesthetic.
- [trainedAlgorithmicMedia](https://www.numonic.ai/glossary/trained-algorithmic-media): An IPTC Digital Source Type standard value indicating that content was generated by an AI model trained on data. Midjourney embeds this value in downloaded images' IPTC metadata, providing a machine-readable disclosure signal. Unlike C2PA, this tag is unsigned and can be trivially removed, creating a compliance gap for regulations requiring tamper-resistant disclosure.
- [Metadata Recovery](https://www.numonic.ai/glossary/metadata-recovery): The process of reconnecting AI-generated images that have lost their generation metadata to their original prompts, parameters, and provenance records. Recovery techniques include Job ID matching against platform history, visual similarity search against known-good assets, and temporal correlation with generation logs.
- [Folder-based Organisation](https://www.numonic.ai/glossary/folder-based-organisation): An asset management approach where images are organized primarily by their storage location in a hierarchical folder tree. Folder-based systems force each asset into a single location despite having multiple relevant attributes (project, style, status, date), creating a single-axis retrieval constraint that breaks down above approximately 10,000 assets.
- [Delivery Package](https://www.numonic.ai/glossary/delivery-package): A structured collection of final AI-generated assets prepared for client handoff, including the image files in required formats, usage rights documentation, AI disclosure language for regulatory compliance, generation provenance records, and approval chain documentation linking each deliverable to its approved selection.
- [Provenance Gap](https://www.numonic.ai/glossary/provenance-gap): A break in the chain of generation metadata that occurs when an AI-generated asset crosses a tool boundary — for example, exporting from Midjourney to Photoshop to ComfyUI. Each tool transition risks losing provenance data because tools use incompatible metadata formats and most editing software discards generation-specific fields on save.
- [Aesthetic-texture matching](https://www.numonic.ai/glossary/aesthetic-texture-matching): Matching input texture richness to the target aesthetic — maximise pores/grain for cinema, restrain for self-tape/phone footage. A form of conditioning and guidance in cinematography.
- [Anti-cut prompting](https://www.numonic.ai/glossary/anti-cut-prompting): Placing explicit continuity language in the first sentence to suppress the model's default to insert cuts. A form of prompt craft in cinematography.
- [Model architecture](https://www.numonic.ai/glossary/architecture): A modern video-generation stack: the diffusion transformer, attention, the VAE, the text encoder, and the dual video/audio streams.
- [Asymmetric parameter allocation](https://www.numonic.ai/glossary/asymmetric-parameter-allocation): Dedicating ~14B parameters to the video stream and ~5B to the audio stream, linked by shared timestep conditioning. A form of model architecture in cinematography.
- [Attention budget](https://www.numonic.ai/glossary/attention-budget): The fixed total weight (1.0) the attention mechanism distributes across all tokens in a prompt. Also known as Zero-sum attention. A form of prompt craft in cinematography.
- [Attention mechanism](https://www.numonic.ai/glossary/attention-mechanism): The active navigator that aggregates, weights, and shifts across the entire vector field to determine the most probable next state. Also known as Attention. A form of model architecture in cinematography.
- [Audio-video coupling](https://www.numonic.ai/glossary/audio-video-coupling): Unified single-pass generation gives excellent per-take sync but ties vocal identity to each individual seed. Also known as Acoustic coupling. A form of failure mode in cinematography.
- [Base sampler](https://www.numonic.ai/glossary/base-sampler): The primary low-resolution DiT stage that builds structure/identity/motion; only ~17% of wall-clock yet does the creative work. A form of sampling and noise schedule in cinematography.
- [Bidirectional interaction](https://www.numonic.ai/glossary/bidirectional-interaction): Decoupled streams constantly exchanging information so visual cues map to auditory events with sub-frame precision. Also known as Cross-modal attention. A form of model architecture in cinematography.
- [Buffer generation](https://www.numonic.ai/glossary/buffer-generation): Generating extra seconds so the clean middle can bridge a blend, never inheriting tail-end trash pixels. A form of production workflow in cinematography.
- [Center of gravity](https://www.numonic.ai/glossary/center-of-gravity): The coordinate (the weighted mean) where the cumulative pull of all prompt tokens converges; the denoising target. Also known as CoG. A form of latent geometry in cinematography.
- [Chain-of-latents](https://www.numonic.ai/glossary/chain-of-latents): System-2 refinement of complex constraints across the latent path using an inference budget, mimicking human deliberation. A form of inference and reasoning in cinematography.
- [Chain-of-steps](https://www.numonic.ai/glossary/chain-of-steps): The finding that reasoning happens along the latent trajectory, with structural decisions made in specific denoising windows. Also known as CoS. A form of inference and reasoning in cinematography.
- [Chain-of-thought](https://www.numonic.ai/glossary/chain-of-thought): Forcing calculation of intermediate steps (e.g. how a voice should sound) before rendering the final frame or waveform. Also known as CoT. A form of inference and reasoning in cinematography.
- [Cinematic bias](https://www.numonic.ai/glossary/cinematic-bias): The DiT's trained default to introduce dynamic camera moves (pans, tilts, dolly pushes) mimicking cinematic training data. A form of model architecture in cinematography.
- [Classifier-Free Guidance](https://www.numonic.ai/glossary/classifier-free-guidance): The inference-time overshoot mechanism extrapolating between conditioned and unconditioned predictions to amplify prompt adherence. Also known as CFG. A form of conditioning and guidance in cinematography.
- [Close-up](https://www.numonic.ai/glossary/close-up): Framing tight on a subject's face or a detail. Also known as CU. A form of shot type in cinematography.
- [Cognitive surrender](https://www.numonic.ai/glossary/cognitive-surrender): The human habit of adopting AI outputs with minimal scrutiny, risking override of human intuition and deliberation. A form of inference and reasoning in cinematography.
- [Conditioning and guidance](https://www.numonic.ai/glossary/conditioning): How prompts and reference latents steer the trajectory: CFG, hard/soft conditioning, identity anchoring, state-continuity.
- [Context window](https://www.numonic.ai/glossary/context-window): The 1,000-token prompt memory limit; beyond it, earlier instructions lose their gravity. A form of embedding and tokens in cinematography.
- [Cross-feeding](https://www.numonic.ai/glossary/cross-feeding): Injecting a percentage of one shot's end-vectors into the next shot's high-noise onset to carry identity, lighting, and camera momentum. Also known as Inertial transfer. A form of latent geometry in cinematography.
- [Decoder](https://www.numonic.ai/glossary/decoder): Translates the fully denoised latents back into visible pixels and audible sound waves. Also known as VAE decoder. A form of model architecture in cinematography.
- [Decoupled latents](https://www.numonic.ai/glossary/decoupled-latents): Using modality-specific VAEs to compress signals into separate video and audio latents rather than one shared space. A form of model architecture in cinematography.
- [Deep-fried pixels](https://www.numonic.ai/glossary/deep-fried-pixels): Neon-glowing edges, crushed blacks, and crunchy texture from high CFG multiplying structural error late in a take. A form of sampling and noise schedule in cinematography.
- [Delta vector](https://www.numonic.ai/glossary/delta-vector): The small per-step calculation that subtracts a specific amount of randomness, nudging latents toward a likelier coordinate. A form of latent geometry in cinematography.
- [Denoising process](https://www.numonic.ai/glossary/denoising-process): The iterative method transforming random noise into a structured audiovisual sequence via per-step delta vectors in latent space. Also known as Denoising trajectory. A form of sampling and noise schedule in cinematography.
- [Diffusion](https://www.numonic.ai/glossary/diffusion): Generating structure from chaos by iteratively removing Gaussian noise to reveal an image or sound. A form of sampling and noise schedule in cinematography.
- [Diffusion Transformer](https://www.numonic.ai/glossary/diffusion-transformer): The 22B-parameter asymmetric hybrid marrying transformer scaling with diffusion fidelity; the shot engine's structural backbone. Also known as DiT. A form of model architecture in cinematography.
- [Drift](https://www.numonic.ai/glossary/drift): Unintended departure of identity/style across frames or versions (the visible symptom of entropy creep). Also known as Latent drift. A form of generation mechanic in cinematography.
- [Drift inheritance](https://www.numonic.ai/glossary/drift-inheritance): The Extend function treating a drifting tail as the new truth, so the next shot starts broken and collapses faster. A form of failure mode in cinematography.
- [Dynamic CFG scheduling](https://www.numonic.ai/glossary/dynamic-cfg-scheduling): Front-loading high CFG (7-10) during structural onset to lock identity, then decaying it as entropy creeps in. Also known as CFG decay. A form of conditioning and guidance in cinematography.
- [Embedding and tokens](https://www.numonic.ai/glossary/embedding): The numerical representations the model reasons over: text/latent/thinking tokens, positional and rotary embeddings, the context window.
- [Embedding](https://www.numonic.ai/glossary/embedding-vector): A numerical vector representation of words, images, or audio capturing inherent properties and semantic relationships. A form of embedding and tokens in cinematography.
- [Entropy](https://www.numonic.ai/glossary/entropy): A measure of unpredictability/disorder in the latent space; generation moves from maximum (noise) to low (structure). A form of sampling and noise schedule in cinematography.
- [Entropy creep](https://www.numonic.ai/glossary/entropy-creep): The cumulative slow-burn rise of disorder from compounding FP8/BF16 rounding errors that progressively dissolves the structured manifold, producing drift. A form of latent geometry in cinematography.
- [Environment lock](https://www.numonic.ai/glossary/environment-lock): Inability to transition between distinct spatial manifolds (interior/exterior, room/room) in a single take without collapse. A form of failure mode in cinematography.
- [Establishing shot](https://www.numonic.ai/glossary/establishing-shot): Opening wide that situates location/time of a scene. A form of wide shot in cinematography.
- [Failure mode](https://www.numonic.ai/glossary/failure-mode): Systemic video-generation failure states and hard walls: text warping, noodle hands, environment lock, drift inheritance, tail-end collapse.
- [Final descent](https://www.numonic.ai/glossary/final-descent): The last large sigma jump to 0.0 that removes haze and sharpens high-frequency detail (pores, fabric grain, hair edges). A form of sampling and noise schedule in cinematography.
- [Flicker and pop](https://www.numonic.ai/glossary/flicker-and-pop): Luminance mismatch between two stitched shots causing the seam to visibly jump; smoothed by the glue schedule. Also known as Flicker. A form of sampling and noise schedule in cinematography.
- [Gaussian noise](https://www.numonic.ai/glossary/gaussian-noise): Random uncorrelated values following the normal distribution; the clean-slate raw material diffusion sculpts. A form of sampling and noise schedule in cinematography.
- [Gemma 3 encoder](https://www.numonic.ai/glossary/gemma3-encoder): The instruction-tuned text encoder (3,840 hidden dims, 262,208 vocab) parsing prompts into the shared latent manifold. Also known as Gemma 3 12B, or Text encoder. A form of model architecture in cinematography.
- [Genetic blueprint](https://www.numonic.ai/glossary/genetic-blueprint): Initial latents treated as ground-truth DNA dictating how identity and lighting evolve across the entire sequence. A form of conditioning and guidance in cinematography.
- [Generation mechanic](https://www.numonic.ai/glossary/gen-mechanic): Latent-generation control surface that materially affects matchable output.
- [Global structure phase](https://www.numonic.ai/glossary/global-structure-phase): The high-sigma early steps that decide global geometry and bone structure rather than colour or texture. Also known as Rough sketch phase. A form of sampling and noise schedule in cinematography.
- [Glue schedule](https://www.numonic.ai/glossary/glue-schedule): A low-noise refinement schedule (0.25, 0.18, 0.12, 0.06, 0.0) that smooths a blend seam without changing geometry. Also known as Transition sigma schedule. A form of production workflow in cinematography.
- [Grammatical scoping](https://www.numonic.ai/glossary/grammatical-scoping): Using brackets for global tone and em-dashes to scope cues to specific tokens, preventing attention bleed. A form of prompt craft in cinematography.
- [Gravity](https://www.numonic.ai/glossary/gravity): The attractive pull tokens exert on the trajectory; early prompt content weights more heavily, so anchors go first. Also known as Front-loading. A form of prompt craft in cinematography.
- [The great leap](https://www.numonic.ai/glossary/great-leap): The large middle sigma jumps that fast-travel through the manifold once structure is set; over-sampling here breeds drift. A form of sampling and noise schedule in cinematography.
- [Guidance oversaturation](https://www.numonic.ai/glossary/guidance-oversaturation): Over-amplifying low-frequency signals via high CFG, producing deep-fried colours and unrealistic contrast. A form of sampling and noise schedule in cinematography.
- [Hard-conditioning](https://www.numonic.ai/glossary/hard-conditioning): The starting frame compressed into latents that act as the literal physical denoising state the model extends. Also known as Hard-conditioning latents. A form of conditioning and guidance in cinematography.
- [High-noise cluster](https://www.numonic.ai/glossary/high-noise-cluster): Tightly clustered early sigma steps (1.0 to 0.975) spending ~50% of compute to lock identity and fight identity popping. Also known as High-noise end. A form of sampling and noise schedule in cinematography.
- [Hybrid pipeline](https://www.numonic.ai/glossary/hybrid-pipeline): Treating the video-generation model as one stage among image-generation, TTS, NLE/VFX, and audio-post tools. Also known as Multi-tool stack. A form of production workflow in cinematography.
- [Identity anchor](https://www.numonic.ai/glossary/identity-anchor): The t=0 frame that establishes and holds subject identity; it degrades as the model applies its transition function over time. A form of conditioning and guidance in cinematography.
- [Identity popping](https://www.numonic.ai/glossary/identity-popping): Characters resolving as different people when insufficient high-noise compute fails to lock a stable identity manifold. Also known as Identity pop. A form of failure mode in cinematography.
- [Inference and reasoning](https://www.numonic.ai/glossary/inference): How the model deliberates: System-1 vs System-2 processing, thinking tokens, chains of latents/steps/thought, statistical inference.
- [Inference budget](https://www.numonic.ai/glossary/inference-budget): The compute allotment a System-2 pass spends deliberating over prompt constraints before diffusion begins. A form of inference and reasoning in cinematography.
- [Instruction dilution](https://www.numonic.ai/glossary/instruction-dilution): Adding descriptive text reduces adherence to primary intent because every word competes for finite attention weight. A form of prompt craft in cinematography.
- [KISS threshold](https://www.numonic.ai/glossary/kiss-threshold): Pruning to the load-bearing tokens; adherence is ~80% at 3 instructions, ~50% at 8, under 28% at 15. Also known as KISS. A form of prompt craft in cinematography.
- [Laconic prompting](https://www.numonic.ai/glossary/laconic-prompting): Using the minimum words to convey a command, keeping the weighted mean a sharp point rather than a blurry cloud. Also known as Laconic precision. A form of prompt craft in cinematography.
- [Latent blending](https://www.numonic.ai/glossary/latent-blending): Mixing the generated latents of two shots inside diffusion (e.g. at sigma 0.5) so features settle on a shared mathematical middle ground. Also known as Latent blend. A form of latent geometry in cinematography.
- [Latent geometry](https://www.numonic.ai/glossary/latent-geometry): The geometry of the latent space: manifolds, vectors, trajectories, and the forces that pull generation toward a target.
- [Latent space](https://www.numonic.ai/glossary/latent-space): The high-dimensional space in which the model represents data by essential semantic features rather than raw pixels; similar concepts sit close together. Also known as Compressed latent space. A form of latent geometry in cinematography.
- [Latent token](https://www.numonic.ai/glossary/latent-token): A VAE-produced token representing a patch's textures/colours/edges; the raw diffusion material. Also known as Visual token. A form of embedding and tokens in cinematography.
- [Latent trajectory](https://www.numonic.ai/glossary/latent-trajectory): The continuous mathematical path the initial latents follow as they transform into generated latents across denoising. Also known as Singular trajectory. A form of latent geometry in cinematography.
- [Latent vector](https://www.numonic.ai/glossary/latent-vector): The quantified essence of creative intent — simultaneously a point (state) and an arrow (direction/transformation) in the manifold. Also known as Numerical vector. A form of latent geometry in cinematography.
- [Load-bearing token](https://www.numonic.ai/glossary/load-bearing-token): The 2-4 essential tokens — the unavoidable truth of the shot — that anchor the scene's centre of gravity. Also known as Massive anchor. A form of prompt craft in cinematography.
- [Video extend](https://www.numonic.ai/glossary/video-extend): One-click extension setting the last frame as new hard-conditioning latents; fast but inherits and compounds drift. Also known as Extend. A form of production workflow in cinematography.
- [Manifold](https://www.numonic.ai/glossary/manifold): The topological surface holding the set of all mathematically valid variations of a given subject or environment. Also known as Latent manifold. A form of latent geometry in cinematography.
- [Mode collapse](https://www.numonic.ai/glossary/mode-collapse): The model defaulting to the most generic average interpretation (e.g. 'person walking') when contradictory cues cannot resolve. A form of conditioning and guidance in cinematography.
- [Mode concentration](https://www.numonic.ai/glossary/mode-concentration): Final-stage CFG amplifying within-mode contraction to lock in fine details and reduce variability. A form of conditioning and guidance in cinematography.
- [Motor-skill failure](https://www.numonic.ai/glossary/motor-skill-failure): Chaotic, rapidly occluding hand topologies the 3D temporal RoPE fails to track, so digits merge or multiply. Also known as Noodle hands. A form of failure mode in cinematography.
- [Multimodal guider](https://www.numonic.ai/glossary/multimodal-guider): A specialised guider splitting CFG into independent text-guidance strength and cross-modal alignment strength. Also known as Multimodal guidance. A form of conditioning and guidance in cinematography.
- [Multimodal integration](https://www.numonic.ai/glossary/multimodal-integration): Treating pixels, audio, and text as numerical vectors in one hidden space, enabling frame-accurate lip-sync. Also known as Multimodal fusion. A form of model architecture in cinematography.
- [Noise injection](https://www.numonic.ai/glossary/noise-injection): Adding timestep-dependent noise directly to the hard-conditioning latents as the raw material the model refines. A form of sampling and noise schedule in cinematography.
- [Patch-based processing](https://www.numonic.ai/glossary/patch-processing): The VAE dividing a frame into a grid of patches, each compressed into a latent token of textures, colours, and edges. Also known as Patches. A form of model architecture in cinematography.
- [Positional embedding](https://www.numonic.ai/glossary/positional-embedding): Encoding token order/position, since transformers are otherwise position-agnostic about sequence. A form of embedding and tokens in cinematography.
- [Prompt craft](https://www.numonic.ai/glossary/prompt-craft): The discipline of writing token-efficient prompts: managing the attention budget, instruction dilution, prosody, scoping, anti-cut language.
- [Prompt iteration](https://www.numonic.ai/glossary/prompt-iteration): Changing prompt wording, which fundamentally changes the manifold; the fix for blocking, framing, or grammar problems. A form of production workflow in cinematography.
- [Punctuation as prosody](https://www.numonic.ai/glossary/punctuation-as-prosody): Typographical symbols inside quotes act as literal acoustic tuning: ! shouts, ? lifts, ellipsis decays, dash pauses. A form of prompt craft in cinematography.
- [Quantisation](https://www.numonic.ai/glossary/quantisation): Precision tiers (FP16, FP8, GGUF Q8/Q4) trading VRAM for softening that prompts counter with texture cues. Also known as Quantization. A form of production workflow in cinematography.
- [Reachable space](https://www.numonic.ai/glossary/reachable-space): The set of mathematically valid variations of a subject's identity established from the initial frame. A form of latent geometry in cinematography.
- [Reasoning](https://www.numonic.ai/glossary/reasoning): High-dimensional statistical inference (geometric optimisation over embeddings), not conscious cognitive judgment. Also known as Latent logic. A form of inference and reasoning in cinematography.
- [Rotary Positional Embedding](https://www.numonic.ai/glossary/ro-pe): Rotates embedding vectors in a complex plane by an angle set by token position, encoding relative distance. Also known as RoPE. A form of embedding and tokens in cinematography.
- [Sampling and noise schedule](https://www.numonic.ai/glossary/sampling): The diffusion denoising process and the sigma/noise schedule that allocates compute across the noise range.
- [Scaling law](https://www.numonic.ai/glossary/scaling-law): Transformers handling far more parameters than U-Nets without instability, enabling the asymmetric 22B engine. A form of model architecture in cinematography.
- [Seed](https://www.numonic.ai/glossary/seed): RNG seed; reproducibility anchor and a strong Signal-1 match key. A form of generation mechanic in cinematography.
- [Seed variance](https://www.numonic.ai/glossary/seed-variance): Varying the random seed (4-6 takes) to explore different trajectories within the same manifold when performance nuance is off. Also known as Seed rotation. A form of production workflow in cinematography.
- [Semantic axes](https://www.numonic.ai/glossary/semantic-axes): Abstract directions in the vector space where each position corresponds to a feature such as warmth, vertical motion, or metallic texture. Also known as Axes of meaning. A form of latent geometry in cinematography.
- [Semantic lock](https://www.numonic.ai/glossary/semantic-lock): Using high-vector-proximity synonyms (e.g. 'parched' vs 'thirsty') to steer texture toward a neighbourhood without more words. A form of prompt craft in cinematography.
- [Semantic redundancy](https://www.numonic.ai/glossary/semantic-redundancy): Synonyms and filler that each consume a token and dilute the gravity of the primary instruction. Also known as Descriptive noise. A form of prompt craft in cinematography.
- [Semantic steering](https://www.numonic.ai/glossary/semantic-steering): Steering the denoising trajectory toward the prompt's weighted mean using mathematical anchors rather than literal understanding. A form of conditioning and guidance in cinematography.
- [Shared block structure](https://www.numonic.ai/glossary/shared-block-structure): Each transformer block running self-attention, text cross-attention, audio-visual cross-attention, and a feed-forward network in sequence. A form of model architecture in cinematography.
- [Sigma](https://www.numonic.ai/glossary/sigma): The noise level at a given step; 1.0 is total Gaussian noise and 0.0 is a fully denoised image or audio track. Also known as Noise level. A form of sampling and noise schedule in cinematography.
- [Sigma schedule](https://www.numonic.ai/glossary/sigma-schedule): A non-linear sequence of sigmas that reshapes where in the noise range the model spends its compute budget. Also known as ManualSigmas, or Noise schedule. A form of sampling and noise schedule in cinematography.
- [Soft-conditioning](https://www.numonic.ai/glossary/soft-conditioning): Using an input merely as an inspirational visual reference rather than a literal starting state (contrast hard-conditioning). Also known as Visual reference. A form of conditioning and guidance in cinematography.
- [Spatial compression](https://www.numonic.ai/glossary/spatial-compression): The VAE's 32x downsampling of physical geometry, why micro-structures like text and hands struggle to survive. Also known as Spatial downsampling. A form of model architecture in cinematography.
- [Spatial upscaler](https://www.numonic.ai/glossary/spatial-upscaler): The heaviest stage (~37% of wall-clock) lifting high-frequency texture from the base sampler's low-resolution foundation. A form of production workflow in cinematography.
- [Spurious subtitles](https://www.numonic.ai/glossary/spurious-subtitles): Overlaid-subtitle hallucination triggered by mentioning dialogue, captions, or text in the prompt. Also known as Subtitle hallucination. A form of failure mode in cinematography.
- [State-continuity](https://www.numonic.ai/glossary/state-continuity): The architectural rule that the model extends states rather than transforms them, carrying the initial manifold forward. Also known as Inertial force. A form of conditioning and guidance in cinematography.
- [Statistical inference](https://www.numonic.ai/glossary/statistical-inference): Calculating the most probable visual and acoustic patterns for semantic vectors from the training distribution. A form of inference and reasoning in cinematography.
- [Stitch and denoiser](https://www.numonic.ai/glossary/stitch-and-denoiser): Generating two shots and blending their overlap with a light denoise pass to purge drift across the seam. Also known as Manual latent blending. A form of production workflow in cinematography.
- [System 1 processing](https://www.numonic.ai/glossary/system1): Fast, pattern-based output relying on immediate statistical likelihood; the default mode for most takes. Also known as System 1. A form of inference and reasoning in cinematography.
- [System 2 processing](https://www.numonic.ai/glossary/system2): Uses an inference budget and thinking tokens to evaluate complex prompt constraints before the diffusion pass. Also known as System 2. A form of inference and reasoning in cinematography.
- [Tail-end collapse](https://www.numonic.ai/glossary/tail-end-collapse): Exponential autoregressive failure at a take's end where limbs morph and identity pops into a new person. Also known as Structural collapse. A form of failure mode in cinematography.
- [Temporal compression](https://www.numonic.ai/glossary/temporal-compression): Compressing video across time (8x stride) so one latent token packs several frames; the reason on-screen typography is destroyed. Also known as Temporal downsampling. A form of model architecture in cinematography.
- [Temporal RoPE](https://www.numonic.ai/glossary/temporal-ro-pe): Generalises RoPE to 3D spatiotemporal tokens (t,x,y); low-frequency temporal channels avoid phase wrapping over long clips. Also known as 3D RoPE. A form of embedding and tokens in cinematography.
- [Temporal text warping](https://www.numonic.ai/glossary/temporal-text-warping): 8x temporal compression loses high-frequency edge data, so on-screen letters drift or morph into illegible glyphs. Also known as Text warping. A form of failure mode in cinematography.
- [Temporal upscaler](https://www.numonic.ai/glossary/temporal-upscaler): Stabilising micro-motion and frame rate after spatial upscaling. A form of production workflow in cinematography.
- [Text embedding](https://www.numonic.ai/glossary/text-embedding): Gemma-3-encoded vectors capturing prompt semantics that guide the video and audio generation. A form of embedding and tokens in cinematography.
- [Text token](https://www.numonic.ai/glossary/text-token): A sub-word unit from the 262,208-token vocabulary; common words are one token, rare words split into several. A form of embedding and tokens in cinematography.
- [Theatrical projection](https://www.numonic.ai/glossary/theatrical-projection): Over-loud delivery on quiet dialogue caused by a stray exclamation mark or capitals inside the quote. A form of failure mode in cinematography.
- [Thinking](https://www.numonic.ai/glossary/thinking): Strategic allocation of compute via thinking tokens to resolve ambiguity, simulating trajectories before committing. A form of inference and reasoning in cinematography.
- [Thinking token](https://www.numonic.ai/glossary/thinking-token): Hidden tokens the encoder generates to deliberate over a prompt before generation, critical for phonetic timing. A form of embedding and tokens in cinematography.
- [Tiered production architecture](https://www.numonic.ai/glossary/tiered-production-architecture): Three workflow variants differing in upscaler steps — survey (Rapid), comparison (Fast), shipping (Final). A form of production workflow in cinematography.
- [Time budgeting](https://www.numonic.ai/glossary/time-budgeting): Calibrating dialogue to 2.5-3.0 words per second so the speech fits the shot's duration. Also known as Words per second. A form of prompt craft in cinematography.
- [Token mass](https://www.numonic.ai/glossary/token-mass): The attention weight/gravity a token carries; load-bearing nouns and verbs are heavy, connector words near-zero. Also known as High-mass token. A form of embedding and tokens in cinematography.
- [Trajectory guidance](https://www.numonic.ai/glossary/trajectory-guidance): Point splines drawn on the initial frame infused into latent space to force generated latents along a precise physical path. Also known as Motion tracking, or Point trajectories. A form of conditioning and guidance in cinematography.
- [Transformer](https://www.numonic.ai/glossary/transformer): The attention-based architecture processing all tokens simultaneously for global context, scaling hardware-friendly with more parameters. A form of model architecture in cinematography.
- [Variational Autoencoder](https://www.numonic.ai/glossary/vae): Compresses raw pixels/audio into latent tokens and decodes them back, using 32x spatial and 8x temporal downsampling. Also known as VAE. A form of model architecture in cinematography.
- [Vector adjacency](https://www.numonic.ai/glossary/vector-adjacency): Reasoning by computing the mathematical distance between instruction tokens and latent trajectories in hidden space. Also known as Semantic neighbourhood. A form of inference and reasoning in cinematography.
- [Vector arithmetic](https://www.numonic.ai/glossary/vector-arithmetic): Navigating meaning by moving along interpretable latent directions (age, lighting, velocity), inducing semantically coherent transformations. Also known as Semantic traversal. A form of latent geometry in cinematography.
- [Vector leap](https://www.numonic.ai/glossary/vector-leap): The mathematical distance moved between latent points; drastic state transitions need large leaps the model suppresses for coherence. A form of latent geometry in cinematography.
- [Vector magnitude](https://www.numonic.ai/glossary/vector-magnitude): The length of a displacement vector, representing the scale of the requested change. A form of latent geometry in cinematography.
- [Voice consistency](https://www.numonic.ai/glossary/voice-consistency): Vocal identity is generated fresh per seed/prompt and varies across takes — an architectural, not configurational, limit. Also known as Vocal identity. A form of failure mode in cinematography.
- [Weighted mean](https://www.numonic.ai/glossary/weighted-mean): The model's most probable statistical interpretation of the whole prompt, used as the navigational target during early denoising. A form of latent geometry in cinematography.
- [Production workflow](https://www.numonic.ai/glossary/workflow): Production techniques and pipeline stages: extend, stitch-and-denoiser, seed variance, quantisation, tiered architecture, upscalers.
- [Workflow archaeology](https://www.numonic.ai/glossary/workflow-archaeology): The counter-intuitive truth that base sampler is not the wall-clock bottleneck; upscale and decode dominate — profile first. A form of failure mode in cinematography.
- [Architectural](https://www.numonic.ai/glossary/architectural): Photography of buildings and structures emphasising line, scale, and perspective. A form of subject in still-image work.
- [Aspect ratio](https://www.numonic.ai/glossary/aspect-ratio): The width-to-height proportion of the frame (1:1, 3:2, 16:9). A form of composition in still-image work.
- [Butterfly lighting](https://www.numonic.ai/glossary/butterfly): A frontal key placed above the lens, casting a butterfly-shaped shadow under the nose. Also known as Paramount lighting. A form of lighting in still-image work.
- [Candid](https://www.numonic.ai/glossary/candid): An unposed, spontaneous capture of a subject. Also known as Reportage. A form of subject in still-image work.
- [Chiaroscuro](https://www.numonic.ai/glossary/chiaroscuro): Strong contrast between light and dark used for dramatic modelling. A form of lighting in still-image work.
- [Color temperature](https://www.numonic.ai/glossary/color-temperature): The warmth or coolness of a light source measured in Kelvin. Also known as Kelvin, or White balance. A form of lighting in still-image work.
- [ControlNet](https://www.numonic.ai/glossary/control-net): A conditioning model constraining generation to a pose, depth, edge, or scribble map. A form of generation parameter in still-image work.
- [Denoising strength](https://www.numonic.ai/glossary/denoise-strength): How much the input image is altered (0 = identical, 1 = ignored) in image-to-image generation. Also known as img2img strength. A form of generation parameter in still-image work.
- [Environmental portrait](https://www.numonic.ai/glossary/environmental-portrait): A subject portrayed within their characteristic surroundings to convey context. A form of subject in still-image work.
- [Fill light](https://www.numonic.ai/glossary/fill-light): A secondary, softer light reducing shadow contrast from the key. A form of lighting in still-image work.
- [Framing devices](https://www.numonic.ai/glossary/framing-devices): Foreground elements (doorways, arches, windows) that frame the subject. Also known as Frame within a frame. A form of composition in still-image work.
- [Golden ratio](https://www.numonic.ai/glossary/golden-ratio): A compositional proportion (~1.618) placing key elements along a spiral or division for natural balance. Also known as Golden spiral. A form of composition in still-image work.
- [Headroom](https://www.numonic.ai/glossary/headroom): The space between the subject's head and the top edge of the frame. A form of composition in still-image work.
- [Hero still](https://www.numonic.ai/glossary/hero-still): A final, presentation-grade single image. A form of still role in still-image work.
- [Inpainting](https://www.numonic.ai/glossary/inpainting): Regenerating a masked region of an image while preserving the rest. Also known as Masked generation. A form of generation parameter in still-image work.
- [Key light](https://www.numonic.ai/glossary/key-light): The primary, dominant light source shaping the subject. A form of lighting in still-image work.
- [Leading room](https://www.numonic.ai/glossary/leading-room): Space left in front of a subject's gaze or direction of motion. Also known as Nose room. A form of composition in still-image work.
- [Light quality](https://www.numonic.ai/glossary/light-quality): Small-source hard light gives sharp shadows; large-source soft light gives gradual shadows. Also known as Hard light, or Soft light. A form of lighting in still-image work.
- [Loop lighting](https://www.numonic.ai/glossary/loop-lighting): A key slightly off-axis casting a small loop-shaped shadow from the nose. A form of lighting in still-image work.
- [Macro](https://www.numonic.ai/glossary/macro): Extreme close-range photography rendering small subjects at life-size or greater. A form of subject in still-image work.
- [Negative prompt](https://www.numonic.ai/glossary/negative-prompt): Text describing what to exclude, steering generation away from unwanted features. A form of generation parameter in still-image work.
- [Practical light](https://www.numonic.ai/glossary/practical-light): A light source visible within the frame (lamp, window) that motivates the scene's lighting. Also known as Motivated light. A form of lighting in still-image work.
- [Product shot](https://www.numonic.ai/glossary/product-shot): A controlled studio capture of a product as the hero subject. Also known as Packshot. A form of subject in still-image work.
- [Rembrandt lighting](https://www.numonic.ai/glossary/rembrandt): A portrait key creating a small triangle of light on the shadowed cheek. A form of lighting in still-image work.
- [Sampler](https://www.numonic.ai/glossary/sampler): The numerical solver that denoises the diffusion latent at each step. Also known as Sampling method. A form of generation parameter in still-image work.
- [Scheduler](https://www.numonic.ai/glossary/scheduler): The function setting the noise level at each step. Also known as Karras. A form of generation parameter in still-image work.
- [Split lighting](https://www.numonic.ai/glossary/split-lighting): A key at ~90 degrees lighting exactly half the face while the other half stays in shadow. A form of lighting in still-image work.
- [Sampling steps](https://www.numonic.ai/glossary/steps): The number of denoising iterations; more steps refine detail at higher compute cost. Also known as Steps. A form of generation parameter in still-image work.
- [Textual inversion](https://www.numonic.ai/glossary/textual-inversion): A learned embedding token capturing a concept or style, invoked by keyword. Also known as TI. A form of generation parameter in still-image work.
- [3D-bound still](https://www.numonic.ai/glossary/three-d-bound-still): A still produced as an input/reference toward a 3D asset, not a final. A form of still role in still-image work.
- [Three-point lighting](https://www.numonic.ai/glossary/three-point-lighting): The standard key + fill + back/rim lighting arrangement. A form of lighting in still-image work.
- [Upscaler](https://www.numonic.ai/glossary/upscaler): A model that increases resolution and adds high-frequency detail. Also known as Hires fix. A form of generation parameter in still-image work.
- [Vignette](https://www.numonic.ai/glossary/vignette): Darkening (or lightening) toward the frame edges that draws the eye inward. A form of composition in still-image work.
- [Ambient occlusion](https://www.numonic.ai/glossary/ambient-occlusion): A baked map darkening contact creases and cavities to fake soft contact shadowing. Also known as AO. A form of surfacing in 3D work.
- [Background removal](https://www.numonic.ai/glossary/background-removal): Isolating the subject on a solid or transparent background before upload; busy or subject-coloured backgrounds corrupt reconstruction. Also known as Cutout. A form of input-image craft in 3D work.
- [Texture baking](https://www.numonic.ai/glossary/bake): Transferring high-poly detail (normals / AO / curvature) into maps for a low-poly target. Also known as Bake, or Baking. A form of surfacing in 3D work.
- [Baked-shadow dent](https://www.numonic.ai/glossary/baked-shadow-dent): A hard cast shadow in the input sculpted into the mesh as a physical dent; cheapest to prevent, brutal to fix. A form of 3d failure mode in 3D work.
- [Blockout](https://www.numonic.ai/glossary/blockout): A rough, untextured massing pass that establishes scale and layout before detailing. Also known as Graybox, or Greybox. A form of geometry in 3D work.
- [3D-safe framing](https://www.numonic.ai/glossary/bridge-rule): The rule that a 3D-input still should be deliberately un-cinematic — flat light, deep focus, subject square-on and whole, plain keyable background. A form of input-image craft in 3D work.
- [Confabulated geometry](https://www.numonic.ai/glossary/confabulated-geometry): Hidden, concave, or back-facing geometry the model never saw and invents, usually wrongly. A form of 3d failure mode in 3D work.
- [Decal](https://www.numonic.ai/glossary/decal): A crisp logo or text reprojected as a texture overlay rather than modelled as geometry. A form of surfacing in 3D work.
- [Decimation](https://www.numonic.ai/glossary/decimation): Reducing polygon count of a dense mesh while preserving silhouette, for real-time or LOD use. Also known as Decimate. A form of production hand-off in 3D work.
- [Deep focus](https://www.numonic.ai/glossary/deep-focus): Everything-sharp input focus (the opposite of shallow depth of field), preserving the edge cues the reconstructor needs. A form of input-image craft in 3D work.
- [Displacement](https://www.numonic.ai/glossary/displacement): A map that physically offsets surface geometry — true depth rather than faked normal detail. Also known as Heightmap. A form of surfacing in 3D work.
- [Diffusion Transformer](https://www.numonic.ai/glossary/di-t): A transformer-backbone diffusion model; a common backbone for 3D generation. Also known as DiT. A form of 3d generation method in 3D work.
- [Flat lighting](https://www.numonic.ai/glossary/flat-lighting): Soft, even, shadowless illumination on the input, so cast shadows are not sculpted into the mesh as dents. Also known as Shadowless. A form of input-image craft in 3D work.
- [Game-ready mesh](https://www.numonic.ai/glossary/game-ready): A clean-topology, low-poly, UV'd mesh that drops into an engine with minimal or no retopology (e.g. image-to-3D tool output). Also known as Gameready. A form of production hand-off in 3D work.
- [3D generation method](https://www.numonic.ai/glossary/gen-method): The generative method that produced a 3D asset — matchable provenance for a generated model.
- [Geometry](https://www.numonic.ai/glossary/geometry): How a 3D asset's form is represented: polygon mesh, point cloud, Gaussian splat, or rough blockout.
- [Hair lumping](https://www.numonic.ai/glossary/hair-lumping): Hair and fur reconstructed as solid bulk mass rather than strands or cards; hero work grooms hair natively instead. A form of 3d failure mode in 3D work.
- [Production hand-off](https://www.numonic.ai/glossary/handoff): Turning a raw generated sculpt into a production-usable asset: retopology, UVs, PBR, interchange, and provenance.
- [High-poly](https://www.numonic.ai/glossary/high-poly): A dense, high-detail mesh — a digital sculpt or the source before decimation. Also known as Highpoly, or Sculpt. A form of mesh in 3D work.
- [Hollow fill](https://www.numonic.ai/glossary/hollow-fill): The model filling openings and cavities that should stay hollow, producing solid where the subject was open. A form of 3d failure mode in 3D work.
- [Image-to-3D](https://www.numonic.ai/glossary/image-to3-d): Generating a 3D asset from a single reference image. A form of 3d generation method in 3D work.
- [Input-image craft](https://www.numonic.ai/glossary/input-craft): How to frame the source still for image-to-3D: the reconstructor imagines geometry from pixels, so ambiguity becomes hallucinated geometry.
- [Level of detail](https://www.numonic.ai/glossary/lod): A set of progressively simpler mesh versions swapped by distance to keep a real-time scene performant. Also known as LOD. A form of production hand-off in 3D work.
- [Low-poly](https://www.numonic.ai/glossary/low-poly): A mesh with a deliberately small polygon count, for real-time or stylised use. Also known as Lowpoly. A form of mesh in 3D work.
- [Large Reconstruction Model](https://www.numonic.ai/glossary/lrm): A feed-forward model that reconstructs 3D geometry from one or a few images in seconds. Also known as LRM. A form of 3d generation method in 3D work.
- [Mesh](https://www.numonic.ai/glossary/mesh): Polygonal surface representation made of vertices, edges, and faces. Also known as Polygons, or Polymesh. A form of geometry in 3D work.
- [Multi-view diffusion](https://www.numonic.ai/glossary/multi-view-diffusion): Generating several view-consistent 2D images with a diffusion model, then reconstructing 3D from them; risks view disagreement / ghosting. A form of 3d generation method in 3D work.
- [Multiview-to-3D](https://www.numonic.ai/glossary/multiview-to3-d): Generating a 3D asset from several overlapping views of one subject; more depth cues than a single image. Also known as Multiview. A form of 3d generation method in 3D work.
- [3D-native latent diffusion](https://www.numonic.ai/glossary/native3-d-diffusion): Diffusing directly in a 3D latent space (the prior is genuinely 3D, not 2D-projected), yielding cleaner watertight geometry. A form of 3d generation method in 3D work.
- [Neural radiance field](https://www.numonic.ai/glossary/ne-rf): An implicit radiance field learned from posed images for novel-view synthesis. Also known as NeRF. A form of 3d generation method in 3D work.
- [Non-manifold geometry](https://www.numonic.ai/glossary/non-manifold): Edges shared by more than two faces or otherwise unclean topology that breaks downstream tools until repaired. Also known as Non-manifold. A form of 3d failure mode in 3D work.
- [Normal map](https://www.numonic.ai/glossary/normal-map): A tangent-space map that fakes high-frequency surface detail on a low-poly mesh. Also known as Normalmap, or Normals. A form of surfacing in 3D work.
- [Object fusion](https://www.numonic.ai/glossary/object-fusion): Multiple objects in one input merging into a single fused mass; avoided by one-subject-per-run. A form of 3d failure mode in 3D work.
- [Orthographic turnaround](https://www.numonic.ai/glossary/orthographic-turnaround): A consistent front/side/back/top image set that lets the model recover otherwise-hidden geometry. Also known as Turnaround. A form of input-image craft in 3D work.
- [Over-densification](https://www.numonic.ai/glossary/over-densification): A dense, evenly-triangulated sculpt with no clean edge flow; the reason hero output must be retopologised. A form of 3d failure mode in 3D work.
- [PBR materials](https://www.numonic.ai/glossary/pbr): Physically based materials (albedo / metallic / roughness / normal) that light consistently across scenes. Also known as PBR. A form of surfacing in 3D work.
- [Photogrammetry](https://www.numonic.ai/glossary/photogrammetry): Reconstructing 3D geometry and texture from overlapping photographs. Also known as Photoscan. A form of 3d generation method in 3D work.
- [Point cloud](https://www.numonic.ai/glossary/point-cloud): An unstructured set of 3D points, typically from scanning, prior to meshing. Also known as Pointcloud. A form of geometry in 3D work.
- [Poly budget](https://www.numonic.ai/glossary/poly-budget): The target polygon / face count an asset class is allowed, set by its role (hero vs background prop). Also known as Facelimit. A form of production hand-off in 3D work.
- [Asset provenance](https://www.numonic.ai/glossary/provenance): The recorded chain from source still → model version → retopo → licence tier for a generated 3D asset. Also known as Lineage. A form of production hand-off in 3D work.
- [Quad remesh](https://www.numonic.ai/glossary/quad-remesh): Rebuilding a triangulated sculpt as an even quad mesh with better edge flow for deformation. Also known as Quadremesh, or Remesh. A form of production hand-off in 3D work.
- [Rectified flow](https://www.numonic.ai/glossary/rectified-flow): A flow-matching generative formulation (e.g. a rectified-flow 3D generator) that learns near-straight transport paths for fast, stable 3D shape generation. A form of 3d generation method in 3D work.
- [Retopology](https://www.numonic.ai/glossary/retopo): Rebuilding a dense or scanned mesh as clean, animation-ready topology. Also known as Retopo. A form of topology in 3D work.
- [Rig](https://www.numonic.ai/glossary/rig): The skeletal / control structure that lets a 3D model be posed or animated. Also known as Armature, Rigged, or Skeleton.
- [Riggable pose](https://www.numonic.ai/glossary/riggable-pose): An A-pose or T-pose with limbs separated from the torso, so the mesh does not fuse arm-to-body and can be rigged. Also known as A-pose, or T-pose. A form of input-image craft in 3D work.
- [Scale and orientation drift](https://www.numonic.ai/glossary/scale-drift): Generated meshes carry no inherent real-world scale or up-axis; both must be standardised on ingest. A form of 3d failure mode in 3D work.
- [Software-defined asset](https://www.numonic.ai/glossary/software-defined-asset): The MovieLabs 2030 direction where an asset is a metadata-rich, provenance-carrying package rather than a bare mesh file. A form of production hand-off in 3D work.
- [Gaussian splat](https://www.numonic.ai/glossary/splat): A radiance-field scene represented as oriented 3D Gaussians, for real-time novel-view rendering. Also known as 3DGS, Splat, or Splatting. A form of geometry in 3D work.
- [Subject isolation](https://www.numonic.ai/glossary/subject-isolation): One subject per generation; multi-object inputs fuse, so separate runs are reassembled in a DCC. A form of input-image craft in 3D work.
- [Surfacing](https://www.numonic.ai/glossary/surfacing): How a 3D surface looks: UV layout, physically based materials, textures, and detail maps.
- [Text melting](https://www.numonic.ai/glossary/text-melting): On-surface text and logos smearing into illegible geometry; treated as a reprojected decal, not modelled as shape. A form of 3d failure mode in 3D work.
- [Text-to-3D](https://www.numonic.ai/glossary/text-to3-d): Generating a 3D asset from a text prompt. A form of 3d generation method in 3D work.
- [Texture](https://www.numonic.ai/glossary/texture): An image map applied to a surface — the base colour / albedo and its companion maps. Also known as Albedo, or Basecolor. A form of surfacing in 3D work.
- [Three-quarter view](https://www.numonic.ai/glossary/three-quarter-view): A front-plus-one-side hero view that yields the most depth from a single input image. A form of input-image craft in 3D work.
- [Topology](https://www.numonic.ai/glossary/topology): The edge/face layout of a mesh; clean topology deforms and textures predictably. Also known as Edgeflow. A form of geometry in 3D work.
- [Transparency lumping](https://www.numonic.ai/glossary/transparency-lumping): Glass, liquid, and other refractive surfaces sculpted as solid opaque lumps rather than thin-walled transparent shells. A form of 3d failure mode in 3D work.
- [Triplane](https://www.numonic.ai/glossary/triplane): A compact 3D representation storing features on three orthogonal feature planes; the output of many large reconstruction models. Also known as Tri-plane. A form of 3d generation method in 3D work.
- [OpenUSD](https://www.numonic.ai/glossary/usd): The Universal Scene Description interchange standard; layering and metadata carry an asset's lineage as first-class data. Also known as USD, or USDZ. A form of production hand-off in 3D work.
- [UV mapping](https://www.numonic.ai/glossary/uv): The 2D parameterisation that maps textures onto a 3D surface. Also known as UV, UVs, or Unwrap. A form of surfacing in 3D work.
- [Watertight geometry](https://www.numonic.ai/glossary/watertight): A closed, manifold mesh with no holes — required for boolean ops, 3D printing, and clean simulation. Also known as Watertight. A form of production hand-off in 3D work.
- [Acoustic creep](https://www.numonic.ai/glossary/acoustic-creep): Voice turning metallic and prosody degrading over successive extends as the phonetic manifold is lost. Also known as Audio latent drift. A form of audio continuity in audio production.
- [Ambience](https://www.numonic.ai/glossary/ambience): The continuous background environmental sound bed that establishes place. Also known as Atmos, or Room tone. A form of stem in audio production.
- [Audio extend](https://www.numonic.ai/glossary/audio-extend): The model carrying the acoustic state (room tone, mid-sentence) into a continuation; risks acoustic creep. Also known as Acoustic relay. A form of audio continuity in audio production.
- [Audio stitch and denoiser](https://www.numonic.ai/glossary/audio-stitch-denoiser): Latent-blending audio vectors for a phase-aligned cross-fade that prevents the DC-offset click of a hard cut. Also known as Audio latent blend. A form of audio continuity in audio production.
- [Background](https://www.numonic.ai/glossary/background): The most distant sonic layer (ambience, walla) sitting behind the mix. A form of audio plane in audio production.
- [Bed](https://www.numonic.ai/glossary/bed): A sustained underlying layer of music or ambience over which foreground elements sit. Also known as Ambient bed, or Music bed. A form of audio form in audio production.
- [Audio continuity](https://www.numonic.ai/glossary/continuity): Carrying the acoustic state across generated segments without an audible seam or identity break.
- [Cue](https://www.numonic.ai/glossary/cue): A discrete bounded segment of music or sound scored to a specific moment or scene. Also known as Music cue. A form of audio form in audio production.
- [Designed SFX](https://www.numonic.ai/glossary/designed-sfx): A synthesised or layered non-literal effect created for impact (whoosh, riser, sci-fi tone). Also known as Sound design. A form of sfx kind in audio production.
- [Dialogue](https://www.numonic.ai/glossary/dialogue): The recorded or generated spoken lines of on-screen or off-screen characters. Also known as DX, Dialog, VO, Voice-over, or Voiceover. A form of stem in audio production.
- [Diegesis](https://www.numonic.ai/glossary/diegesis): Whether a sound originates inside or outside the story world.
- [Diegetic](https://www.numonic.ai/glossary/diegetic): Sound whose source exists within the story world and can be heard by the characters. Also known as Source audio. A form of diegesis in audio production.
- [Foley](https://www.numonic.ai/glossary/foley): Performed everyday sounds (footsteps, cloth, prop handling) recorded in sync to picture. A form of stem in audio production.
- [Foreground](https://www.numonic.ai/glossary/foreground): The closest, most prominent sonic layer, usually dialogue or a key effect. A form of audio plane in audio production.
- [Audio form](https://www.numonic.ai/glossary/form): A bounded unit of audio content: a cue, a bed, or a transition between sections.
- [Hard SFX](https://www.numonic.ai/glossary/hard-sfx): A specific sync-to-picture effect tied to a visible on-screen action (door slam, gunshot). Also known as Spot effect. A form of sfx kind in audio production.
- [Midground](https://www.numonic.ai/glossary/midground): The intermediate sonic layer sitting between foreground performance and background ambience. A form of audio plane in audio production.
- [Music](https://www.numonic.ai/glossary/music): Composed musical material scoring a scene, either diegetic or non-diegetic. Also known as Score. A form of stem in audio production.
- [Non-diegetic](https://www.numonic.ai/glossary/non-diegetic): Sound outside the story world (score, narration) that the characters cannot hear. A form of diegesis in audio production.
- [Phonetic timing](https://www.numonic.ai/glossary/phonetic-timing): The natural rhythm/timing of speech the model computes via thinking tokens before rendering the waveform. Also known as Prosody. A form of audio synthesis in audio production.
- [Audio plane](https://www.numonic.ai/glossary/plane): The depth layer a sound occupies in the mix (foreground, midground, background).
- [Point of audition](https://www.numonic.ai/glossary/point-of-audition): The listening perspective the mix adopts — whose ears we hear from; the audio analogue of point of view. Also known as POA.
- [Sound effects](https://www.numonic.ai/glossary/sfx): Non-speech, non-music sounds representing the actions and events in a scene. Also known as SFX. A form of stem in audio production.
- [SFX kind](https://www.numonic.ai/glossary/sfx-kind): Whether an effect is a literal sync-to-picture hard effect or a synthesised designed effect.
- [Silence](https://www.numonic.ai/glossary/silence): The deliberate absence of sound used for dramatic or rhythmic effect (cf. the field manual's 'allowable silence' in the speech budget). A form of stem in audio production.
- [Stem](https://www.numonic.ai/glossary/stem): A grouped audio mixdown of one category (dialogue, music, or effects) kept separate for mixing and delivery. Also known as Submix.
- [Audio synthesis](https://www.numonic.ai/glossary/synthesis): How the model renders speech and sound: computing phonetic timing then translating it to waveforms.
- [Transition](https://www.numonic.ai/glossary/transition): A sonic bridge (crossfade, segue, stinger) connecting two sections. Also known as Crossfade, or Segue. A form of audio form in audio production.
- [Vocoder](https://www.numonic.ai/glossary/vocoder): The component matching typographical prosody triggers (! ? ... dash) to specific frequency and amplitude waveforms. A form of audio synthesis in audio production.
- [Walla](https://www.numonic.ai/glossary/walla): Indistinct background crowd murmur or chatter suggesting a populated space. Also known as Crowd. A form of stem in audio production.

## Blog

- [AI Storage Doesn't Die, It Compresses and Gets Model-Coupled](https://www.numonic.ai/blog/ai-storage-compresses-not-disappears): A new 2.07-billion-request study on AI image storage settles the "will regeneration replace storage" debate. The answer changes what teams should actually optimize for.
- [How to Audit an AI Model's Self-Reports](https://www.numonic.ai/blog/audit-ai-model-self-reports): Three techniques for testing what a language model actually knows, none of which use its own testimony as the measurement. Repeatable in an afternoon.
- [AI Copyright Rulings Test One Thing: Can You Prove Your Contribution](https://www.numonic.ai/blog/ai-copyright-rulings-burden-of-proof-lineage): German courts ruled on AI-generated art and copyright. The real question they are asking is not "was AI used" — it is whether you can reconstruct what you did.
- [What a Frontier Model Can't See About Itself](https://www.numonic.ai/blog/frontier-model-self-knowledge-limits): Three layers of machine self-access: what a model reports, what it represents but withholds, and what leaves no trace at all. They need different fixes.
- [The Six-Month Gap in a Model's Knowledge Cutoff](https://www.numonic.ai/blog/claude-opus-5-knowledge-cutoff-gap): A stated knowledge cutoff describes the training corpus, not what the model can recall. I measured the gap on Claude Opus 5. It was six months wide.
- [We Made Our AI Coding Agent Judge Its Own Cost](https://www.numonic.ai/blog/ai-agent-judge-its-own-cost): We kept overspending by running our best AI coding agent at maximum effort on routine tasks. Here's the self-judgment pattern that fixed it.
- [AI 2040’s Missing Layer: Content Provenance in the Basin of Sanity](https://www.numonic.ai/blog/ai-2040-plan-a-missing-content-layer): AI 2040 designs epistemic infrastructure but skips the content layer. Why provenance is the missing piece—and what the EU enforces on August 2, 2026.
- [The AI Multiplier Isn’t the Model: It’s the Harness](https://www.numonic.ai/blog/ai-multiplier-harness-not-model): The teams compounding fastest with AI aren’t the ones waiting for the next model. They’re building the scaffolding that turns each release into shipped work.
- [The Babbage Trap: When to Stop Mining for Gold Because You Suspect There’s a Path to Diamonds](https://www.numonic.ai/blog/babbage-trap-strategic-pivots-ai): A founder’s framework for strategic pivots in the AI age, drawn from Charles Babbage, the Kelly Criterion, and the cautionary mirror of HMS Victoria. When the diamond seam is real but the gold mine is not yet finished.
- [How We Used zlib Compression Ratios to Stop NotebookLM from Drowning in Codebase Noise](https://www.numonic.ai/blog/notebooklm-context-drowning-zlib-fix): NotebookLM ignored half our codebase. A zlib compression heuristic and XML-bound atomic packs turned a blind LLM into an expert architectural reviewer.
- [EU AI Act Article 50: What Tools Can Actually Do for Your Studio (and What Is Just Marketing)](https://www.numonic.ai/blog/eu-ai-act-article-50-mandatory-vs-marketing): Article 50 of the EU AI Act becomes enforceable on August 2, 2026. Here is how to read a vendor compliance page like a regulator would, the three terms vendors blur, and an honest map of what Numonic can and cannot do today.
- [Casting Shakespeare With AI: A Public R&D Diary](https://www.numonic.ai/blog/studio-memory-shakespeare-thought-experiment): Why studio memory matters more than generation. Notes from a CTO and CEO running a public AI-video R&D experiment with a four-actor pillar troupe.
- [The Creative Agentic Stack: Six Layers for AI-Native Media Infrastructure](https://www.numonic.ai/blog/creative-agentic-stack): The generic agent stack misses three layers creative teams actually need. Where we think AI-native media infrastructure has to go next, and an invitation to AI video and film teams building through it with us.
- [EU AI Act Article 50: A CTO’s Reading at T-100](https://www.numonic.ai/blog/eu-ai-act-article-50-t-minus-100): On August 2, 2026, Article 50 of the EU AI Act becomes enforceable. Here is what is settled, what is still in negotiation, and what creative studios should do in the next 100 days.
- [GPT Image 2 in ComfyUI: Reasoning-Driven Generation, Real Cost Data, and Why Every Output Needs Provenance](https://www.numonic.ai/blog/gpt-image-2-comfyui-complete-guide): How to use GPT Image 2 in ComfyUI via Comfy Cloud Partner Nodes — real credit/timing data, dense-text rendering, edit fidelity, and full-metadata capture.
- [Automate the Boring Parts of Your Creative Workflow](https://www.numonic.ai/blog/pipeline-builder-creative-automation): Build asset processing pipelines visually or as text. Twenty-nine stage types, eight templates, cron scheduling, and event triggers for AI creators.
- [Workflow Time Travel: See How Your ComfyUI Parameters Evolved](https://www.numonic.ai/blog/workflow-time-travel-comfyui): Numonic fingerprints your ComfyUI workflows and shows how parameters drifted across iterations. Stop guessing what changed between run 30 and run 45.
- [Your Prompts Deserve a Library, Not a Clipboard](https://www.numonic.ai/blog/prompt-library-ai-optimizer): Stop losing your best prompts. Numonic extracts every prompt from your AI-generated assets into a searchable, filterable library with thumbnails and version history.
- [The Three Layers of Content Trust: Identity, Creation, and Distribution](https://www.numonic.ai/blog/three-layers-content-trust): Content provenance needs three layers working together: who published it (identity), how it was made (creation), and where it travels (distribution). Most solutions only cover one.
- [AI as a Governance Partner: A Year of Building with Claude Code](https://www.numonic.ai/blog/ai-governance-partner-claude-code): Three years of AI-assisted development, from chat interfaces to governed in-codebase partnership. Here is what happens when you treat AI as a governance partner, not just a code generator.
- [New Feature: Asset Inspection Lightbox with 800% Zoom and Comparison Modes](https://www.numonic.ai/blog/asset-inspection-lightbox): Inspect AI-generated images at full resolution without downloading. Zoom to 800%, compare variations with synced viewports, and review metadata in a fullscreen lightbox.
- [The Agency Revenue Case for AI Governance](https://www.numonic.ai/blog/ai-governance-agency-revenue-case): AI governance is a profit center, not a cost center. Data on 2-3x pricing premiums, 2.3x client lifetime value, and 85-89% efficiency gains for governed agencies.
- [New Feature: Sessions View with AI Creative Reflections](https://www.numonic.ai/blog/sessions-by-creative-session): Numonic now groups your AI-generated images by creative session and lets you generate per-session AI reflections. No folders. No tagging. Just your creative flow, understood.
- [From Chaos to Catalog: How Numonic Auto-Builds Your Model Library from ComfyUI Workflows](https://www.numonic.ai/blog/from-chaos-to-catalog-auto-model-library): Stop manually tracking LoRAs and checkpoints. Learn how Numonic auto-detects models from ComfyUI workflows and builds a browsable, searchable library.
- [The Agentic Creative Workflow Is Here. Is Your Asset Management Ready?](https://www.numonic.ai/blog/agentic-creative-workflow-asset-management): AI agents now orchestrate ComfyUI, Stable Diffusion, and Midjourney from the terminal. One of them can batch-generate hundreds of images in minutes. Here is what changes and how to build infrastructure that keeps up.
- [App Mode Is ComfyUI's Platform Moment—and a Signal for Infrastructure](https://www.numonic.ai/blog/app-mode-is-comfyui-s-platform-moment-and-a-signal-for-infrastructure): ComfyUI App Mode transforms a node editor into a distribution platform. We analyze what this means for the generative AI stack and the infrastructure layer beneath it.
- [Bulk Export Failure Modes and Defences: Why Your Midjourney Archive Breaks](https://www.numonic.ai/blog/midjourney-export-bulk-reliability): ZIP timeouts, incomplete downloads, and filename collisions quietly corrupt Midjourney archives. Here is a reliability playbook for bulk exports at scale.
- [Organise Midjourney Images After Export: Inside Midjourney vs Everywhere Else](https://www.numonic.ai/blog/midjourney-10000-images-organisation): Midjourney's web app organises images inside the platform. But creative work happens outside it. Here is how to bridge the export gap at 5,000+ images.
- [Your Midjourney Prompt Library Has Outgrown Notion](https://www.numonic.ai/blog/midjourney-prompt-library-beyond-notion): Notion is where prompt libraries start. But prompts without visual context, asset linking, or version lineage are just text. Here is what a governed prompt library looks like.
- [Save Midjourney Prompts Like a Pro: From Snippets to Managed IP](https://www.numonic.ai/blog/midjourney-save-prompts-pro): Most MJ users save prompts in notes or spreadsheets. But prompts are now embedded in your downloads — and they travel with every file you share. Here is how to manage prompts as IP.
- [Finding That One Midjourney Image From Three Months Ago](https://www.numonic.ai/blog/midjourney-prompt-search-history): You remember the idea but not the prompt, the folder, or the date. MJ's search works inside the platform — but your exported files are unsearchable. Here is how to close the retrieval gap.
- [The Agency Guide to Midjourney: From Brief to Deliverable](https://www.numonic.ai/blog/midjourney-agency-workflow): Midjourney is excellent for concepting. But without governance, agencies lose iterations, leak prompts to clients, and have no audit trail. Here is a 6-phase production workflow.
- [AI Disclosure for Midjourney Exports: What Embedded Metadata Is Not, and What Compliance Looks Like](https://www.numonic.ai/blog/midjourney-eu-ai-act-compliance): Midjourney's IPTC Digital Source Type tag is a real standard — but unsigned metadata is a statement, not a proof. Here is what Article 50 requires before 2 August 2026.
- [Safe Sharing: When to Strip Midjourney Metadata, How to Keep Internal Provenance](https://www.numonic.ai/blog/midjourney-privacy-sharing-guide): Downloaded MJ PNGs embed your full prompt in metadata. Share an image, share your creative process. Here is a redact-vs-preserve decision framework.
- [The Midjourney Delivery Package: Hand Off AI Images Without Losing Context — or Oversharing It](https://www.numonic.ai/blog/midjourney-client-delivery-package): Clients need high-res finals and AI disclosure. They should not see your prompts, iterations, or MJ username. Here is a standardised delivery checklist.
- [Building a Searchable Style Reference Library Beyond SrefHunt](https://www.numonic.ai/blog/midjourney-style-reference-library): SrefHunt and Midlibrary are great for discovering styles. But discovery is not governance. Here is how to build a style library with approval status, usage tracking, and export.
- [Midjourney Notion vs DAM: When Your Prompt Database Isn't Enough](https://www.numonic.ai/blog/eagle-vs-notion-vs-numonic-midjourney): Notion is great for prompt catalogues. Eagle is great for visual browse. But neither is a system of record for assets, provenance, and compliance. Here is a decision framework.
- [Midjourney vs ComfyUI: Beyond Aesthetics to Provenance, Reproducibility, and Governance](https://www.numonic.ai/blog/midjourney-comfyui-workflow-bridge): MJ embeds a text Description. ComfyUI embeds a full workflow JSON graph. Here is how the metadata model of each tool shapes governance, compliance, and cross-tool workflows.
- [Midjourney Library Health Metrics: A Practical Scorecard for Teams](https://www.numonic.ai/blog/midjourney-library-health-metrics): "Organised" is not a metric. Here are 7 measurable KPIs for your MJ library — from metadata coverage to compliance readiness — with a red/amber/green scorecard.
- [Bulk Rename vs Metadata: Why Filename Schemes Fail for Midjourney at Scale](https://www.numonic.ai/blog/midjourney-filename-vs-metadata): Renaming files feels like organisation. But filenames encode 3-5 attributes at best. Metadata encodes dozens. Here is why the UUID in your MJ filename is more valuable than you think.

(93 more posts at https://www.numonic.ai/blog)

## ComfyUI node reference

- [KSampler](https://www.numonic.ai/integrations/comfyui/ksampler): The primary sampling node in ComfyUI. Runs the diffusion process that transforms random noise into a coherent image using a model, positive/negative conditioning, and configurable parameters like seed, steps, CFG scale, and sampler algorithm.
- [KSampler (Advanced)](https://www.numonic.ai/integrations/comfyui/ksampler-advanced): An extended version of KSampler with fine-grained control over the noise addition and start/end steps. Used for advanced workflows like multi-pass rendering, img2img refinement, and ControlNet-guided generation.
- [Load Checkpoint](https://www.numonic.ai/integrations/comfyui/checkpoint-loader-simple): Loads a Stable Diffusion checkpoint model file (.safetensors or .ckpt) and outputs the MODEL, CLIP, and VAE components. The primary entry point for model loading in most ComfyUI workflows.
- [Load LoRA](https://www.numonic.ai/integrations/comfyui/lora-loader): Applies a LoRA (Low-Rank Adaptation) weight file to an existing model and CLIP encoder. LoRAs modify the base model behavior—adding styles, concepts, or characters—without replacing the full checkpoint.
- [Load VAE](https://www.numonic.ai/integrations/comfyui/vae-loader): Loads a standalone VAE (Variational Autoencoder) model file to override the VAE bundled with a checkpoint. Used when a higher-quality or specialized VAE produces better image decoding results.
- [DualCLIPLoader](https://www.numonic.ai/integrations/comfyui/dual-clip-loader): Loads two CLIP text encoder models simultaneously, required for architectures like SDXL and Flux that use dual text encoders (CLIP-L and CLIP-G or T5) for richer prompt understanding.
- [CLIP Text Encode (Prompt)](https://www.numonic.ai/integrations/comfyui/clip-text-encode): Converts a text prompt into CONDITIONING tokens that guide the diffusion sampling process. This is where the user prompt enters the generation pipeline, making it the primary source of creative intent metadata.
- [Load ControlNet Model](https://www.numonic.ai/integrations/comfyui/control-net-loader): Loads a ControlNet model that enables image-guided generation. ControlNets add spatial control (edges, depth, pose, segmentation) to the diffusion process, allowing precise composition from reference images.
- [Apply ControlNet](https://www.numonic.ai/integrations/comfyui/control-net-apply): Applies a loaded ControlNet model to conditioning with a reference image and strength parameter. This node is where spatial guidance from a reference image merges into the generation pipeline.
- [VAE Decode](https://www.numonic.ai/integrations/comfyui/vae-decode): Decodes a latent representation into a visible pixel image using a VAE (Variational Autoencoder). This is the final conversion step that transforms the sampler output into a viewable image.
- [VAE Encode](https://www.numonic.ai/integrations/comfyui/vae-encode): Encodes a pixel image into a latent representation using a VAE. Used in img2img workflows where an existing image serves as the starting point for generation instead of random noise.
- [Upscale Image](https://www.numonic.ai/integrations/comfyui/image-scale): Resizes an image to specified dimensions using configurable interpolation methods (nearest, bilinear, bicubic, lanczos). Used for resolution changes, upscaling, and preparing images for further processing.
- [Upscale Latent](https://www.numonic.ai/integrations/comfyui/latent-upscale): Resizes a latent representation before decoding, enabling hi-res fix workflows that sample at a higher resolution in a second pass. Operates in latent space for better quality than pixel-space upscaling.
- [Save Image](https://www.numonic.ai/integrations/comfyui/save-image): The primary output node in ComfyUI. Writes generated images to disk as PNG files and embeds the full workflow JSON in PNG metadata chunks, preserving the complete generation graph for reproducibility.
- [Load Image](https://www.numonic.ai/integrations/comfyui/load-image): Loads an image from the ComfyUI input directory for use in img2img, ControlNet, or other image-based workflows. Records only the filename, creating a fragile provenance link to the source material.
- [Empty Latent Image](https://www.numonic.ai/integrations/comfyui/empty-latent-image): Creates a blank latent tensor of specified dimensions, serving as the canvas for txt2img generation. The width and height set here determine the output image resolution (in latent space, 8× for pixels).
- [Video Combine (VHS)](https://www.numonic.ai/integrations/comfyui/vhs-video-combine): Combines a batch of frames into a video file — the standard video-output node for ComfyUI animation workflows. Choose video/h264-mp4 with yuv420p for the broadest playback compatibility; WebM, AV1, ProRes, FFV1, GIF, and WebP cover web, archival, and preview needs.
- [Load Video (VHS)](https://www.numonic.ai/integrations/comfyui/vhs-load-video): Loads a video file and extracts individual frames as an image batch for processing in ComfyUI. Part of the VideoHelperSuite extension, enabling video-to-video workflows.
- [AnimateDiff Loader](https://www.numonic.ai/integrations/comfyui/animate-diff-loader): Loads an AnimateDiff motion model and injects it into the base diffusion model, enabling frame-coherent animation generation. AnimateDiff adds temporal awareness to standard image generation models.
- [AnimateDiff Combine](https://www.numonic.ai/integrations/comfyui/animate-diff-combine): Combines AnimateDiff-generated frames into a video or GIF output. The animation-specific alternative to VHS_VideoCombine, with settings tailored for AnimateDiff frame sequences.