Node guide
This is a user-facing map of the nodes you normally see in a Context Loop workflow. It focuses on what to connect and what comes out; exact expert fields remain documented by the node tooltips and the linked specialist guides.
Read a node at a glance
The tables use this direction:
input socket or widget ──> NODE ──> output socket
- Socket means connect another node.
- Widget means choose or type a value on the node.
- Inputs marked optional can stay disconnected for the normal path.
- Outputs described as status or summary are for inspection; they do not need to be wired for generation.
The flow overview in the README shows where the core nodes sit.
How disabled nodes are shown
Disabled nodes remain part of the saved workflow, so their sockets and purpose still matter. These docs keep their node-shaped representation and add an explicit state label:
| Representation | ComfyUI mode | Meaning |
|---|---|---|
| Solid card — ACTIVE | 0 |
Executes when its output is requested. |
| Grey dashed card — MUTED | 2 |
Present, but never executes. No output is produced. |
| Dashed amber card — BYPASSED | 4 |
Skips its operation and forwards a compatible input. |
This follows the useful behavior of disabled-pack schema previews: input sockets, widgets, defaults, and outputs remain visible even though the node is not live. It prevents a muted recovery branch from looking like missing or broken documentation.
In the shipped generation examples, Load Manifest → Assemble — recovery is muted. Unmute that pair only to assemble saved scenes without rendering. In some deferred-upscale examples, an attention override is bypassed intentionally for compatibility.
Main loop
| Node | Inputs you normally use | Outputs you normally use | What it does |
|---|---|---|---|
| MiniMax H3 Generation Profile | Widgets: Scene continuity, Audio profile. Optional: lip_sync_options. |
chain_policy, status |
Turns two user choices into the complete policy expected by Plan. |
| MiniMax H3 Plan (Modern) | Required chain_policy; organized Project, Canvas, Generation defaults, and Delivery settings; optional project_assets. |
plan, summary, scene count, width, height, blend frames |
Recommended for new graphs. Keeps the familiar scene-column editor, removes all legacy fallback controls, and compiles the same Plan contract as the original node. |
| MiniMax H3 Context Loop Plan | Widgets: Scene Plan, run_name, size, duration, seed. Optional sockets: chain_policy, project_assets. |
plan, summary, scene count, width, height, blend frames |
Original compatibility-preserving Plan. Existing workflows keep it unchanged. Connect Generation Profile, then use Upgrade to Modern Plan… from its context menu when desired. |
| MiniMax H3 Chain Preflight | plan |
checked plan, preflight, ready, status, report JSON |
Validates the full plan before model-heavy nodes execute. Wire its Plan output to Loop Start. |
| MiniMax H3 Context Loop Start | plan. Optional: scene range, source timeline, source audio, external context. |
flow, state, status |
Starts scene 1, a bounded range, or a compatible resume. |
| MiniMax H3 Context Loop Current Shot | state. Optional source-audio fallback. |
state, scene index/count, prompt, seed, raw length, steps, size, source-audio slice, blend frames | Exposes the current scene as ordinary ComfyUI values. |
| MiniMax H3 Current Tagged Ref2VA Scene | state, CLIP, video VAE, audio VAE, tagged references. Optional: Tagged Scene Options. |
state, positive conditioning, latent, typed scene data | Compact prompt-driven Ref2VA route. It replaces the visible Current Shot + Tagged Ref2VA pair in new reference graphs. |
| MiniMax H3 Context Loop Context | state, conditioning, video VAE, latent. Optional: audio VAE, model, drift sigmas, anchors, lip-sync voice. | conditioning, trim frames, continuation flag, latent, model | Adds the selected visual/audio continuation. Scene 1 is effectively a pass-through. |
| MiniMax H3 Context Loop Trim | decoded images, trim count. Optional: audio, FPS, overlap controls, state. | delivered images/audio, overlap-ready images, overlap count | Removes the repeated head context and retains any frames needed for assembly blending. |
| MiniMax H3 Context Loop Segment + Checkpoint | state, delivered images, sampled latent. Optional: audio, overlap-ready images, denoised latent. | segment, status |
Saves the scene movie, latent continuation, dependency record, and immutable revision. |
| MiniMax H3 Context Loop Review Gate | state, segment. Widgets: enabled, timeout, candidate count, memory cleanup, partial assembly. | reviewed segment, status |
Pauses for approve, retry, reroll, candidate selection, or clean stop. |
| MiniMax H3 Context Loop End | flow, state, images, sampled latent, reviewed segment. | manifest, manifest JSON, final context frames, final context latent | Recurses into the next scene or finishes the selected range. |
| MiniMax H3 Context Loop Assemble | manifest. Widgets: audio source, filename, bitrate. Optional: output copy, blend/tone controls, recovery media. | video_path |
Verifies selected segments, joins them, and muxes the chosen audio into the final MP4. |
The H3 model loader, scheduler, sampler, and VAE decoders are ordinary ComfyUI or MiniMax H3 nodes. Context Loop provides the plan, continuity, checkpoint, and recovery layer around them.
Authoring and policy nodes
| Node | In → out | Use it when… |
|---|---|---|
| Scene Prompt Editor | Plan → Plan | You want the stable per-scene prompt editor and prompt history. |
| Rich Scene Prompt Editor (Experimental) | Plan → Plan | You want the experimental rich editor and optional optimizer UI. |
| Plan Studio (Experimental) | Plan and optional Source Timeline → Plan | You want editorial placement, gaps, chapter views, previews, timeline controls, or picture-only alternate final-cut takes. |
| Project Asset Carousel | Project name/catalog; optional existing tagged refs or upscale model → project assets, references, Source Timeline | You want one project library instead of many loader/tag nodes. See the Carousel guide. |
| Scene LoRA Scheduler | Model plus lazily revealed LoRA inputs; state → scene model | Different scenes use different connected LoRA routes. |
| Lip-Sync Options | Timing/denoise widgets; optional vocal stem → lip-sync options and voice | The Generation Profile uses Lip-sync to source audio. |
| Advanced Policy Override | chain policy + advanced transition → chain policy | You need Tone, Latent, Detail, Drift-Control, or Color-Stable Drift behavior. |
| Manual Chain Policy (Legacy) | Low-level transition/audio widgets → chain policy | An existing workflow already uses it. Prefer Generation Profile for new graphs. |
| Legacy 0.4 Policy Adapter | Old 0.4 fields; optional current policy → chain policy | You are rebuilding or migrating an old control surface. |
See Scene authoring, Audio and continuity, and the complete Plan format.
References and source media
| Node group | Main inputs → outputs | Use it for |
|---|---|---|
| Tagged Picture / Video / Motion / Audio Ref | Media + tag/options → tagged references | Build a prompt-selected @tag reference collection. |
| Tagged Ref2VA | Tagged refs + CLIP/VAEs + current scene values → conditioning, latent, fingerprint | Compile only the references used by the current prompt. |
| Tagged Scene Options | Reference/backend widgets → options | Keep non-default reference settings off the compact Current Tagged Scene node. |
| Scene Data Extract | typed scene data + selected field → typed value | Recover a secondary Current Tagged Scene value such as source-audio slice or RefMod sources. |
| Scheduled Picture / Video / Audio Ref | Media + scene selector/options → schedule | Assign media by scene number or range instead of prompt tags. |
| Scheduled Ref2VA | Schedule + current scene + CLIP/VAEs → conditioning, latent, fingerprint | Compile the active scheduled references for one scene. |
| Source Timeline | Source media/path and timing options → Source Timeline | Register the reusable video/audio timeline once. |
| Audio Tracks | Optional full mix, vocals, instrumental AUDIO → Source Timeline | Use vocals for source-locked lip-sync while preserving the full mix for delivery; mix stems only when a full mix is absent. See grouped audio. |
| Reference Video Prep | Video + fit/timing controls → prepared reference | Prepare motion/reference video for H3 timing. |
| Frame Gate | Opening image; optional last frame + scene index → first/last-frame outputs | Apply the opening image only to scene 1. Do not bypass it in I2V workflows. |
| Frame Index Switch | Scene index + image inputs → selected image | Alternate last-frame targets across scenes. |
| Existing Video Context | Plan plus source video or frames; optional audio → external context | Prepend an existing clip and continue from its tail. |
Lazy motion loaders, source-timeline previews, semantic anchors, and Patch Priority are advanced reference tools. Their behavior is covered by Scheduled references and Advanced workflows.
Runs, recovery, and output
| Node | Main inputs → outputs | Use it for |
|---|---|---|
| Run Manager | Plan, archive controls, optional loader assets/Source Timeline/tagged refs → Plan, Source Timeline | Browse runs, restore a saved Plan and loader bindings, and archive recovery assets. |
| Checkpoint Manager | Saved UI selection; optional Plan → selected manifest | Pin a workflow-local lineage or choose Selected chapter only for chapter export/upscaling. Chapter output preserves original scene numbers and timing, and permits different resolutions in earlier chapters. Local output leaves the project and connected Plan unchanged. |
| Load Manifest | Plan; optional source media/context → manifest, manifest JSON, status | Verify and load saved segments without executing the generation loop. Usually muted in generation examples. |
| Chapter Delivery | Manifest, enabled, chapter number → delivery manifest, snapshot path | Seal a chosen chapter's generated scenes for independent export. 0 selects the chapter containing the last generated scene; a positive number selects that chapter explicitly. |
| Chapter Recovery Load | Run name, chapter number, optional snapshot ID → chapter manifest | Recover a sealed chapter's exact saved lineage and editorial state without loading or extending the rest of the run. |
| Export PNG Sequence + Audio | Manifest + optional VAEs, or file-backed VIDEO + upscale state; output folder, 8/16-bit PNGs, workers, verification, reuse → directory, PNG count, status, WAV path, VIDEO | Legacy latent PNG/WAV export remains available. VIDEO mode decodes and publishes one scene at a time with bounded memory, then passes the original VIDEO downstream. Verified unchanged frames can be reused; changed contents are never overwritten. |
| Full-Chain Latent Video Adapter | Generated manifest + video VAE → cached continuous video | Build a low-RAM, disk-backed whole-run input for SeedVR2. |
| Assemble | Manifest + audio/output choices → MP4 path | Join a complete or partial source/upscale manifest. |
See Runs and recovery for review branches, resume rules, folder layout, deferred upscaling, and assembly options.
Masking nodes
| Node | Main inputs → outputs | Use it for |
|---|---|---|
| Masking · Loop Source AV Target | State + source video/audio + H3 target → scene target latent | Slice the exact source interval for the current loop scene. |
| Masking · Loop Mask Slice | State + mask batch → current scene masks | Keep a tracked mask aligned with recursive scene timing. |
| Masking · Trim Source AV | Source frames; optional audio → valid-length frames/audio, length, status | Trim the tail to H3's largest valid 17k+5 frame count without padding or resizing. |
| Masking · Grid Preview | Mask + H3 grid controls → snapped mask, preview, status | Inspect the actual 32×32 H3 cells that will change. |
| Masking · Apply Target Mask | H3 target latent + mask → masked target | Combine the spatial mask with any existing AV protection mask. |
| Masking · Master Audio + Video Prefix | H3 target + full source track + timing/prefix options → masked target, audio slice | Protect exact source audio while controlling the incoming video prefix. |
| Masking · Two-Clip AV Bridge | Start/end clips + H3 target → bridge target | Protect both ends and generate only the middle gap. |
Use Masked editing for the required wiring and mask/audio semantics.
Upscale and research nodes
Deferred-upscale workflows include a checkpoint adapter, current-scene node, reference-conditioning recovery, pass-2 AV preparation, segment save, and an upscale Loop End. The de-rope example adds Guard, Freeze Mask, Continuity, and Recovered AV nodes. These nodes are designed to be used as a supplied group, not assembled one at a time from this short reference.
Start from a deferred-upscale example and use Runs and recovery for the detailed contracts.
Nodes whose names end in Experimental, Research, Internal, or Legacy are not the normal starting point. Keep their supplied wiring unless you are following the matching specialist guide.