MiniMax H3 Plan Studio (Experimental)

Write your whole MiniMax H3 movie before H3 ever loads a weight

By ethanfel·Created 17 days ago·Updated about 19 hours ago· 315

plan
source_timeline
source_audio
tagged_references
reference_schedule
chain_policy

plan
preflight
ready
status
report_json
plan_summary
clip_count
width
height
video_blend_frames

◄start_clip1►

◄scene_range►

◄verify_resume_historytrue►

◄plan_json{ "shots": [ { "id": "intro", "prompt": "Describe the opening shot." }, { "id": "continuation", "prompt": "Continue the same take." } ] }►

◄run_nameh3_chain►

◄generation_fingerprint►

◄width960►

◄height544►

◄context_length22►

◄encode_modevideo►

◄anchor_modehead►

◄cropdisabled►

◄audio_modegenerated_audio►

◄audio_context_length22►

◄default_duration_seconds15.00►

◄default_steps20►

◄base_seed0►

◄segment_crf18►

◄video_blend_frames0►

◄continuation_modeguide►

The Plan Studio is where a Contex Loop production stops being a graph and starts being a movie. Everything else in this pack - the Preflight, the sampling body, Review Gate, the assembly - is built around the Plan object this node produces. If you're here, you already know the pain: MiniMax H3 tops out around 15 seconds a clip, and stitching six of those into one coherent scene with continuity is where most multi-shot H3 projects quietly die. This node is the "sit down and write the scenes" step.

It's marked experimental, and it shows in the best way: instead of one more wall of widgets, you get a timeline-oriented authoring studio with scene navigation, per-scene prompt editing, saved-segment status, a source-audio waveform, and synchronized preview playback. Think of it as the script supervisor for a node graph.

How it works

The Studio runs in two modes. Connect an existing H3 Chain Plan into the plan socket and it edits and mirrors that Plan live - change a scene in the Studio, the Plan node reflects it. Leave plan unconnected and the Studio builds and validates its own complete Plan from the Plan settings tab, then hands you that Plan on the plan output to feed into Loop Start. Either way the thing you get out is a validated H3_CHAIN_PLAN, and you get a free preflight report alongside it.

Under the hood the Plan is just structured JSON, and the Studio has a plan_json field with a Raw JSON mode for import, export, or edits the visual editor can't reach. Scene prompts support {wide shot|close-up} alternatives - Plan resolves each group from that scene's Prompt alternatives control, which can derive a stable seed, keep an exact fixed seed, or randomize on every queue. The exact resolved choice gets saved with the checkpoint, which matters for resume.

The fields a beginner actually touches

run_name - the identity of one production. It maps to a folder under ComfyUI/output/h3_chains/. Keep it unchanged to resume or regenerate; new name, new production. Reusing a name on purpose re-exposes that run's checkpoints to Review Gate.
width / height - generation canvas for every scene. Connect the Plan's width/height outputs to the stock Ref2VA/I2V node so its latent always matches the Plan.
base_seed - the root seed that derives a stable, different seed per scene. Review Gate's reroll writes an explicit override per scene rather than touching this, so the rest of your production stays reproducible.
chain_policy - the one-wire hook for MiniMax H3 Chain Policy, which sets transitions and audio in a single connection. Use it; the legacy continuation_mode/context_length/audio_mode fallbacks are hidden in the normal 0.5 interface for a reason.
segment_crf - H.264 quality of each saved scene MP4. 18 is visually great; lower is bigger files.

There are also start_clip, scene_range, and verify_resume_history for resuming an interrupted production, plus generation_fingerprint - a compatibility tag you bump whenever the model, VAE, LoRAs, CFG, sampler, or scheduler changes so resume doesn't silently mix runs.

Outputs: plan, preflight (a structured report), ready (boolean), status, report_json, plan_summary, clip_count, width, height, and video_blend_frames.

Install

This is one node from the ComfyUI-MiniMaxH3-Contex-Loop pack. Easiest is ComfyUI Manager → search "ComfyUI-MiniMaxH3-Contex-Loop" → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop.git

Restart ComfyUI. There's no requirements.txt - the pack runs on ComfyUI's own built-ins. What it does need is a current ComfyUI build with the native Add Guide for MiniMax H3 node from ComfyUI PR #15439, and ffmpeg on your PATH is preferred (PyAV is the fallback). Models are not bundled: you supply the H3 weights yourself, which is where the elephant sits - the MiniMax H3 Community License excludes the US, EU, UK and South Korea, so if you're in those regions the local weights aren't licensed for you at all.

Common gotchas

The Studio validates the Plan as you go, so most mistakes (a scene whose context length doesn't line up with the AV clock, a reference timestamp off the canvas) show up in the preflight report before you spend GPU on them. Two things trip people regardless: forgetting to update generation_fingerprint when you swap models or LoRAs mid-production (resume will reject the run), and editing an earlier scene's prompt after later scenes are already saved - that invalidates dependent resume. The good news is both failures are loud, not silent. Start from a maintained example workflow and you'll rarely hit either.

Categoryconditioning/minimax/contex_loop

Inputs (26)

Name	Type	Default	Description
planopt	H3_CHAIN_PLAN		Optional H3 Chain Plan. When connected, Plan Studio edits and mirrors that Plan. When omitted, the Studio builds and outputs its own complete validated Plan from the Plan settings tab.
source_timelineopt	H3_SOURCE_TIMELINE		Optional source media for the shared model-free preflight.
source_audioopt	AUDIO		Legacy source route; do not connect with Source Timeline.
start_clipopt	INT	11–128	Resume scene to preflight.
scene_rangeopt	STRING		Optional contiguous scene selection to preflight.
verify_resume_historyopt	BOOLEAN	true	When resuming after scene 1, compare every saved predecessor dependency with the active Plan. Disable only for an intentional advanced recovery from known artifacts.
tagged_referencesopt	H3_TAGGED_REFERENCES		Optional active prompt-driven reference registry.
reference_scheduleopt	H3_REFERENCE_SCHEDULE		Optional legacy scheduled reference registry.
plan_jsonopt	STRING	{ "shots": [ { "id": "intro", "prompt": "Describe the opening shot." }, { "id": "continuation", "prompt": "Continue the same take." } ] }	The editable production plan behind the large Scene Plan interface: shared prompt, ordered scene prompts, optional lengths, sampler steps, per-scene context, continuation, and seed overrides. Use the visual editor for normal work and Raw JSON only for import, export, or advanced editing. Reference media is connected elsewhere; this JSON only mentions native @tags, Tagged Picture semantic anchors such as #hero[2.50s], or native <Picture/Video/Audio N> labels. Scene prompts may use {first option\|second option}; Plan resolves each group from that scene's Prompt alternatives control.
run_nameopt	STRING	h3_chain	Identity of one render history and its folder under ComfyUI output/h3_chains. Keep it unchanged to resume or regenerate scenes from that same production. Use a new name for a separate render; reusing a name intentionally exposes that run's existing checkpoints to Review Gate and resume.
generation_fingerprintopt	STRING		Checkpoint compatibility tag for generation inputs not stored in plan_json. Connect Scheduled Ref2VA's schedule_fingerprint when using scheduled references. Reference-node outputs retain append lineage, so a newly added ref that is inactive in a completed scene does not invalidate that scene. Otherwise enter/change a stable tag whenever the model, VAE, LoRA, global references, CFG, sampler, or scheduler changes. Resume rejects a mismatched fingerprint instead of mixing runs.
widthopt	INT	96032–4096	Generation width for every scene. Connect the Plan width output to the stock Ref2VA/I2V node so its latent always matches the plan.
heightopt	INT	54432–4096	Generation height for every scene. Connect the Plan height output to the stock Ref2VA/I2V node so its latent always matches the plan.
context_lengthopt	COMBO	22	Legacy 0.4 fallback, hidden in the normal 0.5 interface. Connect Chain Policy for new workflows or the Legacy 0.4 Policy Adapter when rebuilding an old control surface. Default previous-scene video frames used to continue motion. Use 22 for guide mode and 39 for masked_av, tapered_av, feathered_av, or audio_feathered_av/Drift AV so the AV clocks meet exactly. A scene's Advanced selector can override this; blank inherits it and 0 starts a visually new scene. Audio context is controlled separately. With head anchors, those frames are regenerated at the start and Loop Trim removes them, so later scenes deliver raw scene frames minus context_length. Larger values strengthen motion continuity but produce fewer new frames per scene. This does not control reference-audio duration.
encode_modeopt	COMBO	video	How the carried visual overlap is encoded. Use video (recommended) to preserve the previous frames as one motion-bearing latent clip. frames creates separate still-image anchors, costs more conditioning space, and is mainly for diagnosing or experimenting with anchor behavior.
anchor_modeopt	COMBO	head	Where previous frames sit on the next scene's timeline. head is the tested default: it repeats the overlap at the beginning, and Loop Trim must remove exactly trim_frames. before places context at negative time and returns no repeated head; use it only for workflows deliberately built around that experimental timing.
cropopt	COMBO	disabled	How saved context frames are fitted when their shape differs from the Plan canvas. disabled resizes directly to width x height and may change aspect ratio. center preserves aspect ratio, then center-crops overflow. It does not crop Ref2VA picture/video reference inputs.
audio_modeopt	COMBO	generated_audio	Legacy 0.4 fallback, hidden in the normal 0.5 interface and used only when no H3 Audio Policy is connected. It controls timeline continuity and final audio; it does NOT enable or disable @voice/<Audio N> references. For a finished prerecorded voice, dialogue, or song that must remain exact, choose source_track: wire the full track to Loop Start and Assemble, and feed Current Shot's exact slice to Ref2VA/Scheduled Audio. For a short @voice identity/timbre reference while H3 generates new speech and sound, choose generated_audio: no full source track is required, connect the audio VAE to Loop Context, and save trimmed generated audio. source_plus_timeline provides both an exact source slice and previous generated-audio context; it is experimental and usually not the first choice.
audio_context_lengthopt	INT	220–240	Legacy 0.4 fallback amount of previous generated sound carried into the next scene, measured in 24-fps video frames. Normal Chain Policy derives 0/22/39 from Cut, Guide, or AV automatically. This field is active only when the Audio Policy has Generated continuity=on. Legacy generated_audio and source_plus_timeline enable that axis; source_track disables it. AV mask modes use zero versus positive as an audio-carry switch: positive preserves audio for the exact video-prefix duration required by the shared AV clock, while zero leaves audio fully denoisable. A scene's Advanced Audio context can override this default; there, explicit 0 means no audio carry.
default_duration_secondsopt	FLOAT	15.000.1–149.6666666666667	Fallback duration only when the scene and JSON defaults both omit a duration/length. H3 cannot generate every frame count, so seconds round UP to the next valid 17k+5 raw length. In head mode, continuation scenes then lose their effective context repeated frames from their delivered duration.
default_stepsopt	INT	201–10000	Fallback sampler steps only when a scene and the JSON defaults both omit steps. A value set under a scene's Show advanced section overrides this.
base_seedopt	INT	00–18446744073709550000	Base used to derive a stable different seed for each scene that has no explicit seed. Review Gate's Reroll seed does NOT change base_seed; it writes an explicit override into that scene's always-visible Scene seed field, leaving every other scene reproducible and checkpoint-compatible.
segment_crfopt	INT	180–51	H.264 quality for each saved scene MP4 (and normalized imported prelude): lower means higher quality and larger files. 18 is visually high quality; 0 is lossless and 51 is lowest quality. This does not change model sampling or the saved safetensors continuation checkpoint.
video_blend_framesopt	INT	00–243	Default visual blend entering each scene, in frames. A scene's Blend entering scene field can override it; blank inherits this value, capped to that scene's context, and 0 makes a hard cut. A positive value requires head anchors. In a 0.5 workflow, connect Current Shot's state output to Loop Trim's state input; Loop Trim resolves this default plus any scene override automatically. Connect Loop Trim's images_with_overlap output to Segment Save. Final and partial videos are re-encoded with a linear cumulative blend; audio timing and the delivered duration remain unchanged.
continuation_modeopt	COMBO	guide	Legacy 0.4 fallback, hidden in the normal 0.5 interface. Connect Chain Policy for new workflows. Inherited default for scenes without a per-scene continuation override. guide keeps the established Motion Context path: previous AV is supplied as fixed guide rows while the overlap is regenerated. tone_carry_guide uses the RGB guide path and applies the predecessor's detected direct tone curve before VAE encoding. masked_av (experimental) VAE-encodes the previous video tail into the current target latent and protects it with a denoise mask. tapered_av (Detail AV) makes a disposable Gaussian-noised copy of only that carried video prefix before applying the same hard mask; audio is never noised. With Generated continuity on it also copies and protects the matching sampled-audio tail; with that policy off, target audio remains fully denoisable for source/reference guidance. feathered_av uses the same prefix but progressively denoises its final temporal steps for a softer handoff. audio_feathered_av keeps the picture prefix hard while releasing only the last eight audio latent ticks with a half-cosine ramp. drift_control_av keeps the clean predecessor checkpoint unchanged but applies a per-model-evaluation, next-sigma video mask with an 8+4 clean-seam taper; it requires the H3 MODEL to pass through Chain Context. All AV modes require video/head, context >= 5, the Chain Context latent output wired to the sampler, and native or compatible H3 AV-mask support.
chain_policyopt	H3_CHAIN_POLICY		Connect MiniMax H3 Chain Policy directly, or pass it through Advanced Policy or the Legacy 0.4 Policy Adapter. It combines transition, final soundtrack, source-audio reference, generated audio continuity, and automatic audio context in one connection.

Outputs (10)

Name	Type	Description
plan	H3_CHAIN_PLAN	The connected Plan unchanged, or the complete validated standalone Plan authored by Studio.
preflight	H3_PREFLIGHT_REPORT	Structured model-free preflight report.
ready	BOOLEAN	True when no blocking preflight error remains.
status	STRING	Concise error/warning count.
report_json	STRING	JSON serialization of the structured report.
plan_summary	STRING	The same normalized timing and compatibility summary as H3 Chain Plan.
clip_count	INT	Number of scenes in the authored Plan.
width	INT	Validated generation width; connect to stock H3 conditioning.
height	INT	Validated generation height; connect to stock H3 conditioning.
video_blend_frames	INT	Legacy Plan-wide blend default; Current Shot state remains preferred.