Feature traceability
This page answers two questions: where did a feature come from? and where is it implemented here? It is a navigation aid, not a replacement for the license notices or Git history.
Origin labels have precise meanings:
- Original — designed and implemented in this repository.
- Adapted — implementation work is based on an identified upstream project or algorithm.
- Inspired — an upstream idea or interaction informed a new local implementation.
- Integrated — this pack targets or interoperates with an upstream API; it does not claim that upstream implementation.
- Compatibility — retained only so older workflows or compatible packs continue to work.
For copied or adapted work, Third-party notices remain the authoritative attribution and license record.
Production loop and authoring
| Feature | What users get | Origin | Implementation and evidence |
|---|---|---|---|
| Recursive scene loop | One reusable H3 sampling body advances through a scene plan. | Original, built on the adapted Motion Context foundation described below. | chain_nodes.py, initial loop commit |
| Plan and scene editor | Multiline prompts, exact H3 timing, per-scene seeds, settings, and bounded scene ranges. | Original | chain_nodes.py, web/h3_chain_plan_editor.js, visual editor commit |
| Per-scene LoRA model routing | Each scene selects the Base model or one of the scheduler's connected LoRA-patched MODEL branches. Route inputs grow progressively from A through Z, Plan and Studio expose only connected routes, and unused branches remain lazy. Existing A-D workflows retain their saved route IDs and connections. |
Original; interoperates with ComfyUI's standard MODEL connections and lazy-input contract without loading or applying LoRAs itself. |
chain_nodes.py, web/h3_chain_lora_scheduler.js, web/h3_lora_scheduler_core.mjs, web/h3_chain_plan_editor.js, web/h3_chain_plan_studio.js, progressive routing commit |
| Strict H3 prompt authoring | Both prompt editors provide mode-aware section validation/repair, alignment helpers, completions, diagnostics, counts, and Plain/Rich presentation. | Backported from the original standalone H3 Prompt IDE, which itself was adapted from this repository's editor. | web/h3_prompt_schema_core.mjs, web/h3_prompt_schema_ui.mjs, web/h3_prompt_completion_core.mjs |
| Review, retry, and reroll | Every rendered scene can be approved, edited and retried, rerolled, or used to stop with a valid partial run. | Original | chain_nodes.py, web/h3_chain_review_final.js, Review Gate commit |
| Checkpoints, resume, and assembly | Accepted scenes are saved atomically and assembled from disk without a cumulative image tensor. | Original | chain_nodes.py, run_manager.py, resumable loop commit |
| Run Manager and prompt history | Saved plans, branching prompt revisions, archived loader assets, and run restoration. | Original | run_manager.py, prompt_history.py, Run Manager, prompt history |
| Plan Studio and Rich Scene Prompt Editor | Optional timeline authoring, source comparison, richer prompt presentation, and optimizer connections. | Original, experimental UI; selected quick-reference and dialogue interactions are inspired by ComfyUI-MiniMaxH3-Easy. |
web/h3_chain_plan_studio.js, web/h3_chain_rich_scene_prompt_editor.js, notice |
| Preflight and structured dependencies | Model-free checks explain source length, references, compatibility, and exactly which change invalidates a saved scene. | Original | chain_nodes.py, contracts_v05.py, preflight, scene dependencies |
| PNG/WAV re-decode | Saved scene latents can be decoded later into lossless PNG sequences, a synchronized PCM WAV, or both. Picture and audio VAEs are independently optional; cached integrity checks, bounded parallel atomic PNG saving, progress, timings, trim/order alignment, and partial recovery remain available. | Original; chunked CPU conversion and threaded image saving are inspired by ethanfel/ComfyUI-JSON-Dynamic's Fast Absolute Saver. |
chain_nodes.py, PNG export commit |
Continuity, audio, and editing
| Feature | What users get | Origin | Implementation and evidence |
|---|---|---|---|
| H3 Motion Context foundation | Previous frames condition the next scene; last-frame targets, 56-frame context, and seam diagnostics remain available. | Adapted from NikoDemon80/ComfyUI-H3-Motion-Context. |
nodes.py, probe_node.py, patch_layout.py, notice |
| Guide and Cut transitions | Clean RGB/VAE context or no carried picture, selected by an incoming-scene policy. Compatible Guide scenes transparently use the installed upstream Motion Context provider; the retained fallback supports exact arbitrary lengths, interior placement, and exact-grid source audio. | Local policy and compatibility adapter are Original; low-level Guide construction and the exact 40 Hz audio-grid helper are Integrated/adapted from seitanism/ComfyUI-H3-Motion-Context-MultiRef, with native placement from ComfyUI PR #15439. |
chain_nodes.py, motion_context_upstream.py, nodes.py, h3_audio_grid.py, contracts_v05.py, notice |
| Tone Carry Guide | Detects a coherent boundary tone difference and applies a saved correction to the next disposable RGB guide. | Original, experimental | chain_nodes.py, implementation commit |
| Latent Guide | Reuses the predecessor's compatible sampled video-latent tail and falls back to RGB/VAE Guide when needed. | Original, opt-in | chain_nodes.py, implementation commit |
| Detail Guide | Applies a deterministic, tapered chroma-noise treatment to a disposable Guide copy. | Adapted, experimental from MacroSony/minimax-h3-chained-character-swap. |
chain_nodes.py, local integration, notice |
| Hard AV and Soft AV | Carries an exact protected picture prefix; optionally carries exact or audio-feathered generated sound. | Adapted from seitanism/ComfyUI-H3-Motion-Context-MultiRef. Per-token masking is Integrated with ComfyUI PR #15375. |
masked_context.py, h3_mask_compat.py, masked AV commit, notices |
| Detail AV | Adds matched-variance latent noise to a disposable Hard AV picture prefix, tapering to a clean boundary while leaving audio exact. | Adapted, experimental from beijinren/ComfyUI-H3-Context-Noise. |
masked_context.py, local integration, notice |
| Drift-Control AV | Applies the sampler's existing noise field to a clean carried video prefix at the next scheduler sigma, with an 8+4 clean-seam mask shared by sampler blending and H3 row timesteps. | Original, experimental integration built on ComfyUI's dynamic denoise-mask and H3 per-token mask APIs. | drift_control.py, chain_nodes.py |
| Color-Stable Drift AV | Adds a tapered, low-frequency E(graded D(z)) - E(D(z)) scene-one color delta to Drift-Control's disposable copied video prefix; audio and saved latents remain immutable. |
Original, experimental integration. | latent_color_carry.py, masked_context.py, chain_nodes.py |
| Reference Video Fade | Keeps complete native Ref2VA video blocks at full early influence, then applies a full-schedule half-cosine value-row fade while preserving still refs, Qwen presentation, audio, guides, targets, and an existing attention backend. | Original, experimental. | reference_video_fade.py, design and wiring |
| Audio policies and exact timing | Final soundtrack, source reference, and generated continuity are independent; cumulative frame boundaries prevent long-run A/V rounding drift. | Policy contract is Original. Cumulative audio budgeting is Inspired by MultiRef; exact AV timing paths are Adapted from later MultiRef updates. | av_timing.py, chain_nodes.py, audio policy, timing integration, notice |
| General masked editing | Loop-aware source AV targets, tracked-mask slicing, 32px cell preview, audio preservation, inpaint/outpaint, and mask composition. | Adapted from ethanfel/ComfyUI-MiniMaxH3-PerRowMasking; mask execution uses the PR #15375 integration instead of the earlier model wrapper. |
masking_nodes.py, source_av_target.py, masking_ops.py, local integration, notice |
| Exact causal mask conversion | Thin or moving masks are max-reduced through H3's causal frame groups and 2x2 latent-token cells instead of being weakened by temporal interpolation. | Adapted from drozbay/MaskVidExperiments. |
masking_ops.py, masking_nodes.py, notice |
| Master-audio continuation and two-ended AV bridge | Protects exact soundtrack windows, optionally carries a video prefix, or preserves both clip endpoints while generating the gap. | Adapted from MultiRef Updates 4–6. | master_audio_context.py, masked_bridge.py, notice |
| Boundary blending and partial recovery | Scene-specific overlap blends also survive interrupted-run assembly. | Original | chain_nodes.py, frame-exact blending, recovery schedules |
References, media, and interoperability
| Feature | What users get | Origin | Implementation and evidence |
|---|---|---|---|
| Prompt-driven tagged references | Stable @tags activate only the pictures, videos, motion, and audio used by the current scene. |
Original | chain_nodes.py, web/h3_reference_autoconnect.js, v0.4 tagged workflow commit |
| Qwen-only semantic pictures and timed anchors | Bare #picture adds an untimed Qwen visual; #picture[time] adds explicit approximate placement or a picture-storyboard cue. Neither creates a native VAE reference. |
Original, experimental | chain_nodes.py, syntax, conditioning, storyboard mode |
| Source Timeline and lazy motion media | Register source media once, decode only the active scene window, and retain exact FPS/PTS and recovery provenance. | Original | chain_nodes.py, disk-backed loader, typed timeline |
| Existing-video continuation | Scene 1 can continue an imported clip and optionally prepend the normalized original during assembly. | Original | chain_nodes.py, workflow commit |
| Legacy scheduled references | Numeric scene selectors remain available for old graphs and explicit schedules. | Original, compatibility route | chain_nodes.py, initial schedule commit |
| H3-Multishot and SolAttn coexistence | Known external payload and attention wrappers remain active instead of being overwritten. | Integrated compatibility; SolAttn defensive detection includes a contribution by tsolful in repository PR #16. |
patch_layout.py, patch_payload.py, compatibility guide |
| Legacy widget-width repair | Older serialized widgets remain usable in current ComfyUI frontends. | Adapted with permission from pekkAi-dev/ComfyUI-LegacyWidgetWidthFix. |
web/h3_legacy_widget_width_fix.js, notice |
Example and media provenance
Example-specific prompts, images, footage, and licensing are recorded next to the workflow that uses them:
When adding a feature from another project, update this page and
THIRD_PARTY_NOTICES.md in the same change. Record the upstream repository,
author, revision, license, local files, and whether the relationship is
adapted, inspired, integrated, or compatibility-only.