joeygambino/MiniMax-H3-Multishot-Workflow

MiniMax-H3 Seamless Chain

A ComfyUI node pack and two workflows that render a multi-shot MiniMax-H3 scene as one continuous take: no visible cut at the shot boundaries, no colour shift between shots, and continuous audio across the whole piece.

MiniMax-H3 natively generates blocks of roughly 10-15 seconds. This pack chains those blocks into arbitrarily long scenes and hands back a single master video with a single master audio track.

Released as v2.0 of the ComfyUI-H3-Multishot pack.

What this repository contains

• ComfyUI-H3-Multishot/ - the ComfyUI custom-node pack (samplers, loaders, studio controls, LoRA stack, GGUF architecture patch).
• workflows/ - two ready-to-load ComfyUI graphs:
  • H3_Seamless_Chain_v2.json - the full workflow (42 nodes, 9 grouped lanes, 5 on-canvas notes).
  • H3_Seamless_Chain_CORE.json - the same chain with zero third-party dependencies.
• INSTALL.md, SETTINGS.md, PROMPTING.md - install steps, the full settings reference, and the boundary/prompt rules.

What this repository does not contain

No model weights. Nothing here is a checkpoint, a text encoder, a VAE or a LoRA. Download the weights separately:

Component	Where
MiniMax-H3 checkpoint (ref2va shipped, fl2va also chains), GGUF quants	joeygambino/MiniMax-H3-GGUF
Text encoder, video VAE, audio VAE	Comfy-Org/MiniMax-H3

GGUF sizing guide from the quant repo: Q8_0 for 32 GB cards, Q5_1 for 24-32 GB, Q4_0 for 16 GB. The curve variants are pruned-form requants.

The ref2va checkpoint is only needed for the reference/bank workflows. It is not required for seamless chaining.

Nodes added in 2.1

Node	What it does
RiftPromptSource	One dropdown over LPFF-style .txt briefs and passthrough .json scripts. Emits story_idea / character / count. Reads input/rift_prompts/, and still reads the older input/joyecho_prompts/ so existing folders keep working.
RiftScriptPicker	JSON script dropdown, and the speaker/voice stash RiftPromptSource feeds.
(not a node) unload_model_after	A switch added to the LLM prompt writer (JoyEcho_LLMEnhance). On, the writer frees its own model from Ollama once the script is written, so the video model gets the card. Uses the writer's existing base_url and model_name — nothing to keep in sync. Added in memory at startup by this pack, so the writer's own package is not modified; the switch simply appears on the node. Off by default. Ollama's OpenAI-compatible endpoint has no keep_alive field and its shim never sets one, so without this the model sits for the server default of five minutes — your whole first shot.

JoyEcho_PromptSource and JoyEcho_ScriptPicker still resolve as deprecated aliases, so graphs saved against 2.0 open unchanged. They were never published under those names — that was the 2.0 bug.

The prompt writer needs a model you actually have

The full workflow points at a local Ollama with model_name = qwen3:14b. Pull it before the first queue or the run stops with LLM API error 404: model 'qwen3:14b' not found:

ollama pull qwen3:14b

Any OpenAI-compatible endpoint works — its URL in base_url, its exact tag in model_name (ollama list prints the tags you have). A remote endpoint is often better: a local writer large enough to be good competes with H3 for the same card and on under 32 GB will evict the model mid-render. When you do run local, turn on unload_model_after on the writer — it frees the model as soon as the script is written.

No LLM at all? Set use_file_prompts to manual entry, delete the writer, and feed your own ----separated shot script into the sampler's script input. The CORE workflow already works this way.

Try a render with every switch off and the reserve at 0 before touching any of this. The activation reserve measures each shape and conditioning payload as it renders and sizes the pool itself, and it holds on 24 GB cards as well as 32 GB. A hand-set reserve overrides that measurement. These switches are for digging out of a spill the console has already named.

Requirements

Always required:

• ComfyUI v0.30.0 or newer (native MiniMax-H3 support).

The CORE workflow needs nothing else. It uses only this pack plus ComfyUI built-ins (LoadImage, LoadAudio, SaveVideo, SaveAudio, CreateVideo, VAELoader, PrimitiveFloat, Note).

The FULL workflow additionally uses:

Pack	Needed for
ComfyUI_JoyAI_Echo_GGUF_Nodes	the LLM prompt writer — ships in the release zip, modified with attribution; use that copy, not upstream
ComfyUI-H3-Motion-Context	continuity=context_pin, the shipped default
RES4LYF	the beta57 scheduler the full workflow ships with
ComfyUI-sol-attn + comfyui-minimax-h3-blockcache-T8	the VRAM/SPEED patch switches
ComfyUI-Custom-Scripts	the on-canvas script preview

ComfyUI validates every node class in a graph before it will queue, so a missing pack stops the whole workflow — even for switches that ship OFF. Each pack can be removed instead (one-widget change or node deletion, documented in INSTALL.md); e.g. no RES4LYF → scheduler=beta (measured cost: lip-sync 8/10 vs 10/10, all else equal), no Motion-Context → continuity=first_frame.

GGUF users also need ComfyUI-GGUF, then run apply_gguf_arch_patch.py once to teach it the minimax_h3 architecture.

Install

1. Copy ComfyUI-H3-Multishot/ into ComfyUI/custom_nodes/, or install the GitHub repo through ComfyUI-Manager (Install via Git URL: https://github.com/jlucasmcrell/ComfyUI-H3-Multishot).
2. Place the MiniMax-H3 checkpoint, text encoder and VAEs in the usual ComfyUI model folders.
3. Restart ComfyUI.
4. Load workflows/H3_Seamless_Chain_CORE.json (no extra packs) or workflows/H3_Seamless_Chain_v2.json (full).
5. GGUF only: install ComfyUI-GGUF and run apply_gguf_arch_patch.py once.

Full detail is in INSTALL.md.

Usage

1. Open a Seamless Chain workflow.
2. Set the MASTER CONTROLS panel (H3StudioControls): resolution, frames per shot, steps. One node drives both the sampler and the prompt writer's dialogue pacing, so the writer knows how much dialogue fits a shot.
3. Write your script into the sampler: one prompt per shot, --- between shots. JSON of the form {"prompts": [...]} is also accepted. In the FULL workflow you can instead let the LLM lane write the shots for you.
4. Leave continuity on the shipped default and hit Queue.
5. The master video lands in output/video/H3CHAIN/ with a paired audio file.

preview_first_shot is ON by default so you can judge shot 1 before paying for the whole chain.

Shipped defaults

Setting	Default
Resolution	1280x736
Frames per shot	362 (about 15.1 s at 24 fps)
Steps	14
Sampler / scheduler	euler / beta57 (full; RES4LYF) — CORE ships stock beta
Continuity	context_pin
Checkpoint	ref2va (shipped) — fl2va also chains, see below
Bank	OFF
VRAM/SPEED switches	all OFF
preview_first_shot	ON
Mux	24 fps
Output	output/video/H3CHAIN/ + paired audio

All switches OFF reproduces the verified recipe exactly.

How the chaining works

Two mechanisms ship in the pack. Both produce a single master; they differ in what crosses the boundary.

1. first_frame chain - H3MultishotSampler

Each shot's last frame is handed to the next shot as its first frame, through fl2va's trained continuation task. That is the model doing what it was trained to do, not a hand-rolled trick. The duplicated boundary frame is trimmed, and the seam audio gets a 40 ms equal-power weld so the join does not click.

No third-party dependency. This is what the CORE workflow uses.

2. context_pin - H3MultishotMemorySampler, continuity=context_pin

The previous shot's last 22 frames ride into the next shot as raw latents - bit-identical, with no VAE round trip - placed at interior keyframe coordinates, alongside a timeline-placed audio reference. The model regenerates that overlap as its own head; the regenerated 0.92 s is trimmed on decode, so what survives is the original tail followed by continuous new material.

Requires the third-party ComfyUI-H3-Motion-Context pack. As of v2.1.1 the two packs coexist cleanly — this pack's payload wrapper declares Motion-Context's compatibility marker, so load order does not matter. seamless and seamless_tail are legacy comparison modes: seamless is a soft latent-only pin that often reads as a cut, and seamless_tail conflicts with Motion-Context and stops up front, before sampling, when that pack is installed.

Why identity holds with no reference images

Two mechanisms stack:

1. The frame relay. Every shot after the first begins from an actual rendered picture of the character. Faces and wardrobe propagate as pixels, not as a re-imagining from text.
2. Byte-identical text. The prompt writer repeats each character's appearance block verbatim in every shot.

The frame pins the specific instance; the repeated text pins the category so the model cannot drift the description out from under the pixels. This was verified on a 40-second two-character scene with zero reference images supplied.

Settings reference

SETTINGS.md in the release carries the complete list. The dials that matter most:

Dial	Notes
shot_count	0 = one shot per prompt in the script.
seed_per_shot	Leave ON. Measured: per-shot seeds hold the face; a single seed shared across all shots drifted both face and voice.
continuity	first_frame or context_pin (see above).
chain_gain_control	Set to flatten on chains past roughly 5 shots. Texture ratchets about 1.3x per join otherwise, so late shots come out visibly over-sharpened.
color_level	off / mvgd / scene. Levels each shot's colour and exposure statistics to shot 1's settled tail (a fixed reference - matching neighbour-to-neighbour re-accumulates drift). Usually unnecessary; for long chains that drift warm or cool.
self_anchor_voice, voice_ref	Voice identity across the chain (the v1.5 headline feature).
output_scale	Lanczos resize of each shot AFTER decode - resolution, not detail. Works with every continuity mode including context_pin; applied per shot and after the bank takes its clip, so conditioning and VRAM are unchanged. Measured 1.78x faster than native for the same output size, and visibly softer.
upscale_model	Optional UPSCALE_MODEL link (ESRGAN and friends via ComfyUI's own loader). Synthesizes detail rather than resizing. Its invented texture never reaches the memory bank, so it cannot feed the sharpening ratchet.
reference_image_size	match or max.
preview_first_shot	Renders shot 1 alone first so you can abort early.

MASTER CONTROLS (H3StudioControls)

One node sets resolution, frames per shot and steps for both the sampler and the prompt writer, so the writer's dialogue budget always matches the shot length actually being rendered.

VRAM/SPEED panel (H3StudioSwitches + reserve control)

Three lazily gated model patches: memory-efficient attention, chunked feed-forward, and block cache. The gates are lazy, so a switch left OFF never executes its patch - and all OFF is exactly the verified recipe.

The activation-reserve heuristic decides how much VRAM to hold back. Its cache keys include a conditioning-payload signature (keyframes / audio refs / two-pass), so a bare shot 1 and a reference-laden shot 2 are measured separately instead of sharing one wrong number. Measured pools are no longer overridden by a fixed floor, the first run of a new payload variant estimates from a measured sibling, and a VRAM spill into system RAM is now detected and named in the console - previously it only showed up as an unexplained ~5x slowdown.

Prompt and boundary rules

These are render-verified. The FULL workflow's prompt writer applies them automatically through its join_style control, which appends them to the system prompt. Hand-written scripts must follow them manually.

• AIRLOCK. Every shot after the first opens holding the previous shot's exact closing arrangement, and gives about 2 quiet seconds before anyone speaks - a breath, a weight shift, real micro-motion, not a freeze.
• The first ~1 second of every chained shot is discarded replay. Dialogue starting at frame 0 loses its opening syllables.
• LAND SETTLED. End each shot back in a stable arrangement, dialogue finished, about 2 seconds spare.
• A spoken line never straddles two shots. Budget: dialogue plus 4 seconds of quiet must fit inside the shot. 243 frames fits one long line; 124 does not.
• Repeat verbatim. Each character's appearance description and the room/lighting description, word for word, in every shot.
• Keep fps at 24. Other rates audibly shift voice accents.

PROMPTING.md has worked examples.

New in v2.1.3

Correction, 2026-08-12 — read this before turning any drift dial on. A render with chain_gain_control=flatten, color_level=mvgd and the per-shot audio leveller all ON came out +142% texture and +18% brighter over three shots, with a visible brightness step at each join. The per-shot approach cannot work, for a reason the code already knew: under context_pin the drift is carried by the raw latent pin, and every one of those dials operates on decoded frames after the pin has been stored. They correct what you see and not what feeds forward. audio_tone_control has been removed. color_level=mvgd is deprecated — its own source comment records a 29% warmth step at every join. Use color_level=scene (one target for the whole piece, applied per frame at the end) and the new master_normalize=luma, both of which run outside the feedback loop and land every frame on the same number, so they cannot create a seam. Texture drift is not fixable after the fact: the only lever is blur, and blur removes real detail along with the invented kind.

• Picture darkening over chains: measured, and the existing dial verified. User-reported (−1.5 luma/shot, monotonic). Same autoregressive mechanism as the audio dulling; the raw-latent pin carries it directly. color_level=mvgd
  • shipped since 2.1, never verified - holds an 8-shot chain to −1.0 total luma where uncorrected loses −10.5 (both seeds). On long chains turn on all three drift dials:
• Audio dulling over long chains: measured, mechanism found, countered. Five-arm A/B on 8-shot chains: with bank_pinned=0 (pure recency conditioning) the voice band collapses - 84-92% of 4-10 kHz energy gone by shot 8; with the default pinned slot, 8-50% depending on seed. It is the audio twin of the seam sharpening ratchet, running the other way, and continuity mode is irrelevant - the bank decides. Two counters ship: a console warning when bank_pinned=0 on a chain past 4 shots (there is no true "bank off" - 0/0 leaves one recency slot, the worst configuration), and audio_tone_control=flatten - the audio twin of chain_gain_control=flatten, EQ-matching every shot's long-term spectral envelope to shot 1's before the weld. Constant per-shot gains, clamped +/-9 dB, half-strength in the top band so it cannot manufacture hiss. Paired A/B on the worst seed: HF loss halved (-49.5% -> -23.7%), rolloff drift cut to a third. It reduces the drift rather than eliminating it (the context_pin replay carries raw latents the EQ cannot reach), and it ships OFF until ears, not spectra, have judged it.
• The Audio Spine produced static with real-world audio files (ref2va, user-reported). The spine encoded guide_audio at whatever sample rate the file arrived in, while the audio VAE expects its own rate (32 kHz) - the native node resamples, the spine path did not. Nearly every real voice or music file is 44.1/48 kHz, so the encoded latent was garbage, and because the spine LOCKS the audio stream to that latent at every sampling step, the render came out as noise. The same file worked through the native MiniMaxH3ReferenceToVideo node, which is exactly what the reporter observed. The spine now resamples to the VAE's rate and upmixes mono to stereo, and the console says so. Measured: guide-to-output correlation went from 0.06 (unrelated noise) to 0.97 on a 48 kHz voice track - identical to a native-rate control. Also verified at 44.1 kHz mono.
• The spine's tooltip claimed latent_handoff only - wrong. It works with every continuity mode; the per-shot stride table has carried each mode's seam trim all along, and the fix above was render-verified on context_pin. This is the locked-audio music-video path, now documented as such.
• two_pass_upscale is removed. It spatially interpolated the raw latent between passes. H3's latent is not a spatially smooth representation, so the interpolated values landed off-manifold and pass 2, running at low sigma, had no room to pull them back. Every arm tested came back as colour-noise mush against clean single-pass controls - including at 14 steps / beta57, the recipe SETTINGS.md previously called render-verified, and including shot 1, which carries no pin at all. It was never a context_pin incompatibility; it did not work in any mode. The guard around it is gone with it.
• output_scale replaces it: a lanczos resize of each shot's finished frames, after decode, so it cannot leave the latent manifold and works with every continuity mode. It adds resolution, not detail - measured at 1.78x faster than rendering the same output size natively (45.5s vs 80.9s at 672x384) and visibly softer. Applied per shot, so a long chain never holds a full upscaled master in memory at once.
• upscale_model: optional UPSCALE_MODEL input for real detail synthesis (ESRGAN and friends via ComfyUI's own loader), per shot, at the model's own factor. Render-verified with RealESRGAN x2plus: 448x256 -> 896x512, and combined with output_scale it lands exactly on the requested size.
• video_latents / audio_latents / head_frames outputs on the memory sampler (issue #12). Every shot's latent exactly as sampled, batched along dim 0, untrimmed. Shots after the first open with head_frames of replayed material that is only removed at decode, so they do not line up with the master until you trim it - the outputs are deliberately raw rather than trimmed on your behalf, because the pin material cannot be recovered later. Verified: 124-frame shots return 37 latent rows (5*((F-5)//17)+2), and 124 + 124 - 22 is exactly the 226-frame master.

Both upscales are applied after the memory bank has taken its base-resolution reference clip, so conditioning and VRAM are unchanged from an un-upscaled run, and the returned latents stay base-resolution.

If you saved your own copy of a v2.1.2 graph, reload the shipped workflow: removing four widgets shifts the saved widget order on that node.

---

Fixed in v2.1.2

Writer-half only (ComfyUI_JoyAI_Echo_GGUF_Nodes). If you paste your own shot list instead of letting the LLM write it, nothing here changes for you.

• Every shot can now be saved as it renders. A chain only became a file at the very end, so anything that failed after the last shot destroyed the whole run — one report was three hours lost to an OOM at the mux, after every shot had rendered successfully. save_every_shot (both samplers) writes each shot to output/video/H3_SHOTS/ the moment it decodes. Written before the seam trim, so consecutive files overlap ~1s and the master is still the clean join. Requested in issue #13.
• Custom sigma schedules. The samplers built the schedule themselves from steps + scheduler with no way to supply your own, so a turbo LoRA that ships the curve it needs simply ran wrong rather than refusing. Both samplers now take an optional SIGMAS input; connect one and it replaces the schedule, steps rebinds to len(sigmas)-1 so the two-pass split rides your curve, and the console says the widgets are being ignored instead of silently overriding you. It is a link-only input, so saved graphs are unaffected. Issue #14.
• --- separators were ignored in passthrough mode. example_script.txt ships --- separated and every doc tells you to write scripts that way, but the writer's passthrough path returned the whole file as ONE shot — which the sampler then repeated to fill shot_count. Pasting a finished multi-shot script rendered the entire text as shot 1, four times. It now splits on the same rule the sampler uses. A single paragraph is still one shot, so .txt batches are unaffected.
• Reference images had no way in. The sampler's reference_images input has always existed, and SETTINGS.md documented it — as unwired, because nothing in the workflow was connected to it. There is now a REFERENCE lane in the anchors column (two image loaders → ImageBatch → a gate), shipped with the gate off so nothing changes until you turn it on. This is the one item here that is a new capability rather than a repair.
• A stale prompt-set filename blocked the whole queue. ComfyUI validates every combo value in a graph before it will run anything, so if RiftPromptSource's saved source_file no longer existed — a renamed folder, a workflow shared from another machine, or simply the prompt lane switched to manual — the run died with Value not in list and nothing executed, including the lanes that were fine. The node now declares VALIDATE_INPUTS, so the filename is only resolved if the node actually runs; switching to manual genuinely disables it. If it does run and the file is missing, the error names the file.
• Every story came out 15 shots. The system prompt ordered exactly 15 when the brief gave no count. It now counts the story's beats — measured 4–7 on ordinary briefs, 7 with no length signal (65 s at 243 frames, past the 1-minute mark). An explicit count is still honoured exactly.
• The mode dropdown did nothing. The shipped workflow's system_prompt box held a frozen copy of the long prompt, and a filled box overrides the per-mode file — so every mode ran the long prompt and pack prompt updates never reached anyone. It ships empty now. If you saved your own copy of the v2.0/v2.1 workflow, clear that box by hand.
• short_story is 1–3 shots, not always exactly 1.
• Messy LLM JSON no longer kills the render. A markdown fence sharing a line with the payload used to destroy it; truncated replies now have their complete shots salvaged; and parsing moved inside the retry loop, where it should always have been. Order: clean → parse → retry ×3 → salvage → fail.

New in v2.0

• Complete single-purpose workflow H3 Seamless Chain v2 (42 nodes, 9 grouped lanes, 5 on-canvas notes), plus a CORE variant with zero third-party dependencies.
• MASTER CONTROLS panel (H3StudioControls): one node drives resolution, frames per shot and steps for the sampler and the prompt writer's dialogue pacing.
• VRAM/SPEED panel (H3StudioSwitches + reserve control): three lazily gated model patches, all OFF by default.
• Energy-aware seam audio ("smart weld"). The boundary audio cut now lands in the quietest gap within the incoming shot's first 0.75 s instead of blindly at sample 0, so a word placed at a shot head is no longer clipped.
• Rewritten activation-reserve heuristic (payload-aware cache keys, measured pools no longer floored, sibling-based first-run estimates, VRAM spill detected and named).
• Prompt writer gained a join_style control that appends the render-verified boundary rules to the system prompt.
• flf_chain with no boundary plates now raises a clear error instead of silently rendering an unanchored chain.

Versions v1.0 through v1.5 shipped the same pack; v1.5's headline was voice identity (voice_ref + self_anchor_voice).

Verified and not verified

Verified

• A 3-shot context_pin chain and multi-shot first_frame chains were reviewed blind by two independent video-understanding models. One described the result as one continuous unedited take, colour consistent, with nothing broken.
• Verified on both static talking-head content and dynamic moving content.
• Blind-reviewed recipe (differs from the shipped defaults, which add ref2va reference rows for explicit voice/identity): fl2va checkpoint, euler sampler, beta57 scheduler, 14 steps, 362 frames per shot (about 15.1 s at 24 fps).
• Identity retention across a 40-second two-character scene with no reference images supplied.

Not yet verified

• Very long chains. Audio dulls slightly at each hop. Restart the chain on scene cuts rather than running one chain indefinitely.
• The ref2va + bank + context_pin combination.
• Hard-FFLF boundary-plate mode.

Nothing above is a benchmark. These are render observations from the recipe as shipped; results will vary with content, resolution and quantisation.

Release contents

MiniMax-H3_Seamless_Chain_v2.0.zip

ComfyUI-H3-Multishot/LICENSE
ComfyUI-H3-Multishot/README.md
ComfyUI-H3-Multishot/__init__.py               defensive loader
ComfyUI-H3-Multishot/apply_gguf_arch_patch.py  on-disk GGUF arch fallback
ComfyUI-H3-Multishot/h3_advanced.py            advanced sampling helpers
ComfyUI-H3-Multishot/h3_avbank_probe.py        AV bank diagnostics
ComfyUI-H3-Multishot/h3_cartridge.py           portable character cartridges
ComfyUI-H3-Multishot/h3_episode_tools.py       StudioControls, StudioSwitches, AnySwitch
ComfyUI-H3-Multishot/h3_gguf_arch.py           teaches ComfyUI-GGUF the minimax_h3 arch
ComfyUI-H3-Multishot/h3_interior_patch.py      interior anchors (stands down for Motion Context)
ComfyUI-H3-Multishot/h3_keyframes.py           keyframe anchor nodes
ComfyUI-H3-Multishot/h3_lora_stack.py          H3LoraStack
ComfyUI-H3-Multishot/h3_multishot_utils.py     samplers, loaders, gates
ComfyUI-H3-Multishot/h3_ref_folder.py          reference-folder picker
INSTALL.md
PROMPTING.md
SETTINGS.md
example_script.txt                             worked four-shot two-hander
workflows/H3_Keyframes.json                    single-clip keyframe anchoring
workflows/H3_Seamless_Chain_CORE.json          same job, zero third-party packs
workflows/H3_Seamless_Chain_v2.json            everything, optional lanes gated off

Three workflows, one reason each. v2 is everything with the optional lanes gated off. CORE does the same job with zero third-party packs — start there if you want a render before installing anything else. Keyframes is a different job: a hand-built sampling graph for anchoring a single clip at chosen frame positions with per-anchor condition strength, not multishot.

H3_Multishot_AIO and H3_Multishot_MEMORY from earlier versions are retired — every lane they had is in v2 (the AIO's episode source, plate chain and audio spine were folded in; MEMORY had nothing v2 lacks). Existing copies keep working.

Links

• Source: https://github.com/jlucasmcrell/ComfyUI-H3-Multishot
• Civitai listing: https://civitai.com/models/2833322
• GGUF quants: https://huggingface.co/joeygambino/MiniMax-H3-GGUF
• Encoder and VAEs: https://huggingface.co/Comfy-Org/MiniMax-H3

License

Apache-2.0 for the node pack and workflows in this repository. Model weights are covered by their own licenses at their respective repositories.