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 v2.7.0 of the ComfyUI-H3-Multishot pack.

v2.7.0 - per-subject voices, flf_chain fixed, chain leveller, ComfyUI 0.34

• Per-subject voice refs (voice_ref_2 / voice_ref_3, both samplers): each character keeps their own voice across a chained scene - verified blind across a full chain, no cross-speaker bleed.
• flf_chain fix: the memory bank stood in flf chains and pinned a clip of shot 1 (which opens on boundary plate 0) into every later shot as an unbound reference - PLATE0 bled back in from shot 2 onward. The bank now stands down automatically in flf_chain; plates alone carry the continuity there.
• H3ChainNormalize: post-chain texture + colour leveller - the 2.6.0 long-take ratchet limit is addressed.
• refresh_pin splice alignment, x0 clamp dial (x0_clamp_window, dose cap 0.30), in-loop latent upscale + refresh_renoise + pin_noise_ramp + auto_chunk_ffn.
• ComfyUI 0.34 supported: native interior keyframe anchors detected; the pack's layout patch stands down automatically on 0.34, unchanged on older cores.
• Engine-aware writer: separate H3 (<d>[English] ...</d>) and LTX (quoted dialogue) system prompts for every mode via an appended engine widget.

v2.6.0 adds the extend take: one prompt, one continuous speech, as long as you want

Set take_seconds on MASTER CONTROLS (or open H3_Extend_Take, now the main workflow, shipped at 1280x736 with a 30-second take) and give the writer one premise. The panel sizes a window for your card and the number of windows that fills the time; the writer's new extend take join style writes ONE continuous speech and cuts it across the windows at sentence boundaries; the chain continues it under context_pin in H3's own voice — no TTS, no airlock, no per-shot dialogue budgeting. Verified before shipping: seven writer-driven renders including a 65-second, 7-window take, every join continued the speech, reviewed blind as one uninterrupted take. Also in 2.6.0: reference photographs now win over the writer's prose (the writer points at the photos instead of describing a face - same-seed measured), auto refs reach the sampler on their own (a second gate downstream used to discard them silently), three reserve-planning fixes from the 24 GB test lab and a writer fix for silent shots echoing the voice anchor. Full notes in the changelog below.

Known limit (2.6.0): the chain's texture ratchet is not fully solved for long takes - measured about +13% fine texture per join at 736x1280 with the anti-drift set on. Under ~4 windows (~30-40 s) it is slight; at 7 windows it is visible sharpening. Keep extend takes to ~4 windows for now; a pin-side fix is in progress for 2.6.1.

v2.5 is the memory release

Four new memory systems, all measured, two of them fully automatic:

• The driver-headroom rule (automatic). High-resolution renders used to take anywhere from 27 minutes to 3 hours for identical work — a lottery caused by the Windows driver demoting GPU memory when the card fills past roughly 95%. The pack now detects that zone and deliberately streams a few GB of weights instead of riding the ceiling. The lottery render became 15 minutes, every time. Nothing to configure.
• low_ram_master. Long chains used to hold every finished shot in system RAM until the final join — tens of GB at the very last step. Switch it on and shots stream to lossless disk staging as they finish; peak RAM becomes about two shots regardless of chain length, with verified-identical output (42.8 dB against the RAM path — codec noise).
• Remote text encoder. The text encoder runs for seconds per shot and holds 15+ GB the rest of the time. Point the new node at any second PC running ComfyUI with this pack and that memory leaves your render card — identical results, verified across machines, with a local cache so repeated text never touches the network. One flag on the switches panel turns it on; it ships off.
• H3 TAE Decode. 2-second full-resolution draft previews from a 9 MB tiny decoder, versus about a minute per shot through the real VAE. For seed hunts and batch triage, never finals.

Plus a Speed Boosters panel (Spectrum, TeaCache, block cache, EasyCache — each measured and eye-tested, with honest notes about which distort people), and new defaults tuned for 16-24 GB cards. Details in the changelog below.

What this repository contains

• ComfyUI-H3-Multishot/ - the ComfyUI custom-node pack (samplers, loaders, studio controls, LoRA stack, GGUF architecture patch).
• workflows/ - three ready-to-load ComfyUI graphs:
  • H3_Seamless_Chain_v2.json - the full workflow: master controls, LLM prompt writer, speed boosters, remote encoder lane, anchors.
  • H3_Seamless_Chain_CORE.json - the same chain with zero third-party dependencies.
  • H3_Keyframes.json - a single clip with anchors at chosen frame positions.
• 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:

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

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:

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. Search H3 Multishot in ComfyUI-Manager and install it - the pack is on the Comfy Registry as comfyui-h3-multishot. You can also copy ComfyUI-H3-Multishot/ into ComfyUI/custom_nodes/ by hand, or use Manager's Install via Git URL with https://github.com/jlucasmcrell/ComfyUI-H3-Multishot. The Registry carries the node pack only; the LLM prompt writer the full workflow uses ships in this repo under nodes/.
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

These defaults target a 16-24 GB card at the best speed/quality mix. Every booster except block cache changes what people look like — leave them off unless you have compared for yourself.

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:

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 switches: memory-efficient attention, chunked feed-forward, and the remote text encoder (speed boosters have their own node). The gates are lazy, so a switch left OFF never executes its path - and all OFF is exactly the verified recipe. The panel keeps its 2.5.x eight output slots so saved graphs keep working; the removed ones always emit False.

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 v2.6.0 - the extend take

One prompt, one continuous speech, as long as you want

Type a length into MASTER CONTROLS take_seconds. The panel picks the largest window whose activation pool fits with most of the weights resident on your card (window = auto; wire the loader's MODEL into the panel for a real weight size) and the count that fills the time; the writer's new extend take join style writes ONE speech and cuts it across the windows only at sentence boundaries; the memory sampler chains them under context_pin. H3 continues the speech across every join in its own voice. Verified: seven writer-driven renders at 141/192/243-frame windows plus a 65 s / 7-window take — every join continued, zero repeats, zero clipped words. take_seconds = 0 is the old behaviour exactly.

Also new: audio_pin_frames on the memory sampler (audio reference window independent of the picture pin; 96 = the JoyEcho 4 s audio memory; neutral at n=1, ships 0), the standalone H3ExtendTake node, the H3_Extend_Take workflow.

Fixed: reserve planning (from the 24 GB test lab)

The bare-to-payload x1.6 no longer fires between the two samplers' differently-named payload signatures; a first run at a new shape borrows measurements only from the same quant family (GGUF figures do not transfer to w4a8/int8 — a cross-family borrow under-reserved a first run into driver paging); the streamed master's metadata is the API graph again.

Fixed: silent shots echoing the voice anchor

A silent shot whose framing boilerplate said "visible lip movement clearly readable", with shot 1's voice riding along as <Audio 1>, re-spoke shot 1's line word for word. The few-shot example no longer mentions lip movement, the rule cites the render, and the writer warns by shot number.

New in v2.5.5 - the memory release

New: the driver-headroom rule (the random-slowdown fix)

The single biggest fix this pack has shipped. High-resolution renders would randomly take anywhere from 27 minutes to 3 hours for identical work — same seed, same settings. The cause: when model weights plus working memory fill the card past roughly 95%, the Windows driver starts demoting GPU memory unpredictably, and whether your render crawled was luck.

The auto-reserve now detects that zone before sampling and deliberately streams a few GB of weights instead — streamed weights are nearly free on modern ComfyUI, the last few percent of VRAM are not. The lottery render became 15 minutes, every time, verified on both the clean-shot and payload-shot branches. Automatic; the console prints a "driver headroom" line when it engages.

New: low_ram_master streams the master to disk

Off (the default), every finished shot is held in system RAM until the final join — historically the point where long renders killed whole machines. On, each shot streams to lossless disk staging as it finishes and the master is assembled from disk through the exact same levelling math: verified at 42.8 dB against the in-RAM path, which is codec noise. Peak RAM stays near two shots regardless of chain length. The finished file's path now comes out of a new master_path output either way.

New: remote text encoder

The text encoder works for a few seconds per shot and occupies 15+ GB for the whole render. Install this pack on any second PC with ComfyUI, point the new H3 Remote Text Encoder node at it, turn remote_encoder ON in the VRAM / SPEED SWITCHES panel, and prompts are encoded over there while your render card keeps the memory. Results are identical — verified across two different machines to float precision — and repeated scene text is answered from a local cache with no network call at all. Ships wired into the full workflow; the panel flag defaults OFF, so single-PC setups see zero change.

New: H3 TAE Decode draft previews

A 9 MB tiny decoder (Kijai's taeh3, downloaded separately into models/vae_approx/) turns latents into full-resolution draft frames in about 2 seconds, versus roughly a minute per shot through the real VAE. Drafts smear fine texture but composition, framing and motion read clearly. For seed hunts and batch triage — never finals.

New: the Speed Boosters panel

Spectrum, TeaCache, block cache and ComfyUI's own EasyCache, each behind a switch on one node, each measured on the same seed and eye-tested on the same masters:

A booster whose pack is missing prints an install link and passes the model through unchanged instead of breaking the graph.

Changed: defaults tuned for 16-24 GB cards

736x1280 at 192 frames and 14 steps, euler/beta57, curve-Q5_1 checkpoint, the full anti-drift set on (context_pin, chain_gain_control=flatten, master_normalize=luma+contrast, pin_renorm, memory_frames=0). The resolution floor is deliberate: the base model distorts faces below roughly 1 MP.

Changed: the writer knows how long a shot really is

The prompt writer now receives each shot's actual speakable span — clip length minus the replay, airlock and settle a chained shot spends — and sizes dialogue against it. Overruns and silent scripts are flagged at generation time with the numbers. Measured on renders: lines over budget garble, lines under budget drag, matched budgets reviewed as natural and fully intelligible.

Changed: writer craft rules from failed renders

Silent shots with visible people now state what mouths are doing — stops the model inventing mumbling. Revealed objects are written as already present — stops mid-shot pop-ins. Shots that open a mouth without dialogue must name the sound it makes.

New: clickable title blocks in every workflow

Each workflow carries its version number and live links to the GitHub repo, this page and the Civitai page, plus a plain-language pass over every on-canvas note.

New in v2.2.5

Fixed: chained shots could stall before the first sampling step

The auto-reserve learns how much activation pool a given shape needs by measuring it. On a chained render the later shots often load partially

• the weights are already resident, so only part of them streams in. The old code discarded every partial measurement as untrustworthy, which meant a chained shot could never contribute what it learned. If the very first shot was also partial, the cache stayed empty forever, the fallback reserve was too small for the shape, and every subsequent attempt failed the same way. A render could sit there refusing to start with no error to point at.

Partial measurements are now kept when the pool is genuinely large, and clamped by a named floor (_auto_cache_floor) that only ever ratchets upward, never down. Verified live on a six-shot chain: offload dropped from 4308 MB to 638 MB once the cache started learning.

The warning that fires when a shot asks for far more pool than is available now distinguishes the two cases it was conflating. If this shape has been measured before, it says so and gives you the number. If it is the first run at this shape, it says that instead of implying something is broken.

Fixed: output_scale was documented backwards

With an upscale model selected, output_scale is not a multiplier on top of it - the model applies its own fixed factor (4x for the ESRGAN family) and output_scale is the final size you want, not an extra step. The tooltip said otherwise, in both samplers. Both are corrected, and upscale_model_name now explains the interaction rather than leaving you to discover it.

New: the upscale says what it will cost before it runs

Decoded frames accumulate in host RAM until the master is joined, so the bill is per-shot multiplied by shot count. A 1344x768 chain through a 4x model is 5376x3072 per frame, and an output_scale left at a value that looked harmless has taken machines down at the join - after all the sampling was already paid for.

Before sampling starts you now get the projected dimensions, MB per frame, GB per shot and GB total, measured against actual free RAM:

[H3Memory] upscale will produce 5376x3072 (4.0x): 198 MB per frame
[H3Memory] WARNING: that will not fit. Frames are held in host RAM
until the master is joined - lower output_scale to 1.60
or lower, or turn the upscaler off.

New: the folder walk prints which index is which file

Queueing a folder told you nothing about the mapping between EPISODE INDEX and the file it would pick, so a cancelled run left you doing off-by-one arithmetic against a single log line to work out where to restart. The ordered list now prints once per folder per session, and each job reports its own position and the index to set to resume there. It re-prints if the folder's file count changes, which is the one case where a remembered listing would mislead.

Fixed: garbled speech - the writer was never told how long a shot is

num_frames on the LLM Enhance node defaults to 0, and it shipped as 0 in the bundled workflow. At zero, the writer is told nothing about clip length, so it sizes every spoken line for the "~10 second clip" assumed in the system prompt regardless of what you actually render. Too long for a short shot is crammed, garbled speech; too short for a long one is dead air.

Three things now close that:

• num_frames = 0 prints a warning naming the widget and telling you to match it to the sampler's frames_per_shot. It used to pass in complete silence.

• The bundled workflow ships with num_frames set to 243, the same value its frames_per_shot already uses, so a fresh install is correct out of the box.

• Shots whose dialogue overruns the budget are reported at generation time with the numbers, instead of being discovered forty minutes later by ear. The check is symmetric: a script that comes back with no dialogue at all is also flagged, because that renders as a silent slideshow.

Fixed: revise asked for lines that could not fit

Chained styles spend about five seconds of every block on the replay, airlock and settle, so the speakable span is shorter than the clip. The main writer subtracted that. revise did not - it sized lines against the raw clip length, from its own separate copy of the arithmetic.

On a 243-frame chained shot that meant revise asked for 12-20 words where only 6-10 fit: double the budget, every line, which is exactly the overrun that renders as garble. It also never checked its own output, because it returns long before the check the main path runs.

Both paths now call one _speakable_budget() and both check their result, so they cannot drift apart again. The duplicate arithmetic is why they had.

Changed: the story writer stops prescribing creativity

The system prompts had accumulated a five-beat dramatic arc, a banned-cliche list, a "land near seven shots" target and a dialogue quota. None of those are properties of the model - they are one person's taste, hardcoded, and they were flattening every brief toward the same shape.

They are gone. What remains is what the renderer actually needs: valid JSON, identity restated verbatim per shot so the room does not drift, literal physical description over mood language because abstract adjectives have nothing to render, and the note that audio is half the model. The story, its length, its tone and whether any given shot speaks are the model's call now.

Changed: the node menu no longer says "Rift"

Four nodes sat in a category called Rift, which means nothing to anyone who is not me. They are in H3/script now, the labels drop the prefix, and ten console messages say [H3 Multishot] - the name you actually installed - instead of [Rift].

Saved workflows are unaffected. Categories are never written into a graph and display names resolve fresh on load. The underlying class names are deliberately unchanged, because those strings are written into every saved .json and renaming them would turn your nodes into red missing-node boxes.

New: installable from the ComfyUI Registry

The pack is published as comfyui-h3-multishot, so ComfyUI-Manager can find and update it without a manual clone. The zip on this page stays the complete bundle - both node packs, the workflows and the docs - and remains the right choice if you want everything in one drop.

The dependency list is deliberately empty. Every import this pack uses (torch, torchaudio, psutil, av, numpy, PIL) is already guaranteed by ComfyUI itself, and a custom node that declares torch can talk pip into replacing a working CUDA build with a CPU wheel.

Documentation: drift on long chains, and the dials that fight it

SETTINGS.md gains a measured section on why chained shots accrete detail - each conditions on the previous shot's own output, so invented texture compounds - and which controls actually counter it. Measured over ten shots at 960x544:

• memory_frames = 0 is the big one. The bank's RECENT slots feed accreted output forward on top of the pin.

• master_normalize = luma+contrast levels brightness and contrast of the finished chain.

• pin_renorm = on holds each pinned latent at shot 1's sigma.

• chain_gain_control = flatten on chains longer than about five shots - texture ratchets roughly 1.3x per join.

Residual with all of those on is about 1.02 per hop. Not zero. Long chains still drift, slowly. pin_noise is small, scene-dependent and gets worse above 0.10; it is not the fix it was described as in 2.1.5.

New in v2.2.4

Four things people asked for or tripped over, plus the documentation that should have prevented two of the questions.

Fixed: reference images of different people blended into one face

If you gave the sampler reference pictures of two characters, it rendered something that resembled neither. That was not you holding it wrong. Internally every reference picture was declared to the model as <Picture N> is a reference photograph of <Subject 1> - the same Subject 1, every time. With one character that is correct and it is why single-character references work well. With several you were telling the model that photographs of different people all showed one individual, and it produced the average.

There is a new reference_subjects field on the sampler. Leave it empty and nothing changes - every picture is one person, exactly as before. Fill in comma counts in picture order to group them:

reference_subjects: 3,3 pictures 1-3 are person A, 4-6 are person B
reference_subjects: 2,2,2 three people, two pictures each

Each group is then declared its own subject, with an explicit instruction that it is never blended with the others. Only Subject 1 is described as speaking, because H3's voice conditioning is single-speaker.

New: point the prompt source at a folder and queue the whole thing

start_index has always selected a block inside one file. That works for LPFF-style prompt batches, where one file holds many prompts. It does not work for H3 scripts, because in an H3 script --- separates shots within one scene - concatenating scenes into one file would fuse them into a single enormous take.

Turn on the new walk_folder switch and start_index selects which file instead, walking the chosen folder in sorted order and emitting the whole script. Wire it to a Primitive set to increment, queue thirty times, and thirty finished scenes render unattended.

New: hand H3 an existing clip as a video reference

There is now a V2V REFERENCE lane in the workflow - Load Video into Get Video Components into the new H3 Reference Video node, then into two new sampler inputs. It ships muted; un-mute the three nodes to use it.

Read what this is before you wire it. H3 is told a video reference is "a clip from an earlier moment of this same continuous scene" and asked to keep its framing, camera distance, room contents and colour temperature. It is scene and appearance conditioning. It is not motion control - there is no pose, depth or optical flow path in H3, so your subject will not copy the movement in the clip. It borrows the place and the look, not the action.

Keep the window short. References are subsampled to 2 fps and then ride through every sampling step, so a 25-second clip is roughly 50 reference frames of permanent per-step cost. The H3 Reference Video node trims to a window and prints what that window will cost before you commit to it. Requires ref2va; fl2va has no reference rows and ignores video references entirely.

Fixed: an empty prompt dropdown that blamed the wrong thing

The prompt source finds .txt files through the inspire_prompts folder path, which is registered by comfyui-inspire-pack. Without that pack installed the path does not exist, so the dropdown came back empty however many prompt files you had, and the error blamed the prompts folder. It now tells you which of the three fixes applies: install the pack, register the path yourself in extra_model_paths.yaml, or set manual_path and ignore the dropdown. The dependency is written into INSTALL.md as well.

Documentation: what has to change between shots

Several people have hit chains where shot 3 comes back as a near-copy of shot 2. The prompting guide is partly responsible: rule 5 says to repeat descriptions word-for-word, and never said what must differ. Rule 5 is only about appearance and the room. It is not an instruction to restate the action, and when it gets read that way most of shot 3 is byte-identical to shot 2 - and the model is separately instructed to preserve the subject, the room and the colour temperature, so "keep everything the same" wins.

There is now a sixth rule covering it. The short version: each shot's action must leave the world in a state the previous shot's world was not in, physical and irreversible, not a mood or a camera move. If you can swap two shots' action lines and the script still reads correctly, the model cannot tell them apart either.

Documentation: fl2va is not a smaller file

Both this guide and INSTALL.md described fl2va as "lighter and faster", which everyone reasonably read as a size claim while choosing a file for a 24 GB card. It is not. fl2va and ref2va are exactly the same size at every quant level. "Lighter" only ever meant fewer tokens per sampling step, because there are no reference rows riding along.

What actually separates them was never written down: fl2va lands on a supplied frame and ref2va only nudges toward one. Measured against the same frame, 26.35 dB on fl2va versus 16.15 dB on ref2va with a keyframe at 6 turbo steps - and 16.81 dB at 20 stock steps, which rules out the sampler and leaves the checkpoint. fl2va can also take a first and a last frame and plan a camera move between them, which ref2va cannot do at all. So: ref2va when identity or voice must persist, fl2va when a shot must start exactly where the last one ended.

Improved: the story writer, for thin briefs and long shot counts

Give the writer a couple of characters and a genre with no plot, ask for fifteen shots, and it tended to return atmospheric moments rather than a story. That was the system prompt's fault in four specific ways, all now fixed.

It was calibrated to about seven shots and argued against more

• "most scenes land between 4 and 10", "go higher only when the story clearly has that many real beats", plus a padding test. Asked for fifteen, the model was simultaneously told to produce exactly that count and that this many was probably padding, and it resolved the contradiction by rationing one thin premise across the shots. It is now told the opposite: a high count is an instruction to invent more story - another location, a second complication, a character who arrives partway through, a reversal that changes what the earlier shots meant. The no-padding rule is unchanged and is exactly the point; the way to satisfy a high count is to invent enough real plot that no shot is padding.

When the brief names only characters and a genre, the writer is now told plainly that it is the author: decide who wants what, the incident that starts it, the complication, the turn and the ending before writing any shot. Returning a series of moods involving the named characters is called out as the most common failure on a thin brief.

Dialogue density now has a floor. Speech was marked optional per shot and non-speaking shots were actively encouraged, with nothing on the other side, so long pieces came back nearly silent. It now aims for two thirds or more of shots to carry speech, with silent shots as punctuation.

And a new FIT THE SHOT section covering both dialogue and action. Overrunning the clip is what produces crammed, garbled speech and distorted motion, so the writer now budgets each shot - settle, action, line at an unhurried pace, a beat to land on - and cuts the action rather than speeding up the speech. One clear physical action per shot; if more than one is written, split the shot.

Note on saved workflows

Every new control in this release is appended at the end of the node's input list. Widget values are stored positionally, so inserting a control anywhere else silently shifts every value after it in workflows you have already saved. Your existing graphs load unchanged.

New in v2.2.3

Everything here is a fix. Nothing changed about how you use the workflow.

Fixed: long multi-shot renders died at the very last step

A long job would sample every shot, upscale every shot, finish master normalize - and then die while assembling the result, with nothing written. Every expensive stage had already succeeded.

RuntimeError: DefaultCPUAllocator: not enough memory:
you tried to allocate 33791016960 bytes.

That is host RAM, not VRAM, and it is one contiguous request. Upscaled frames were held on the host as fp32 for the whole run, and the final assembly then allocated a second complete timeline while the first was still alive. Peak was twice the timeline, and the timeline grows as shots × frames × height × width × 3 × 4 bytes with the upscale factor squared. Six shots of 243 frames upscaled is a 36.5 GB timeline and a ~70 GB peak, which no 64 GB machine can satisfy.

Three changes: the timeline is now assembled into one preallocated buffer, releasing each shot as it is copied (bit-exact - the same bytes as before); frames are parked as fp16, which is well clear of the 8-bit the encoder writes anyway; and master normalize upcasts to fp32 for its arithmetic and writes back, so the memory saving costs no precision.

Effect: six shots of 243 frames upscaled drops from ~70 GB peak to ~21 GB. Twelve shots of 192 frames drops from ~58 GB to ~31 GB. Both now fit in 64 GB with the per-shot upscale unchanged. Diagnosed by a second operator on a 3090 box; thank you.

Fixed: the auto-reserve inflated its own estimate until the render spilled

If your chains got slower as they went - shot 1 fine, shot 3 sluggish, shot 4 crawling - this is why. After each shot the pack measures how much activation memory that shape needed. It subtracted the weight bytes currently resident rather than the model's full size, so whenever the model was only partly loaded the measurement came out too large by exactly the amount that had been offloaded. That number then raised the next shot's reserve, which left less room for the weights, which offloaded more.

shot 1 recorded 6.4 GB true 6.4 (full load - correct)
shot 2 recorded 8.0 GB true 5.3 (2749 MB offloaded)
shot 3 recorded 8.9 GB true 4.3 (4784 MB offloaded)
shot 4 asked for 19.9 GB -> clamped -> 262 s/step

The overstatement equals the offload every time. The measurement now uses the model's full size, and a shot that offloaded weights records nothing at all

• it measured a spill, not an activation pool.

The cache also only ever grew, so a single bad shot poisoned a shape permanently. Entries written by earlier versions are dropped once on first load; you will see a line saying so, and those shapes re-measure on their next run.

This does not make an over-committed render fit. If a shot honestly needs more memory than the card has you still have to lower frames_per_shot or resolution - the difference is that the warning now fires on true numbers instead of the reserve quietly climbing.

Fixed: the sol_attn and chunk_ffn switches did nothing

Reported by sdktertiaire2. Those switches shipped ON, but the nodes they gate ship bypassed - they need third-party packs that cannot be bundled (ComfyUI-sol-attn, comfyui-minimax-h3-blockcache-T8). A toggle routing into a disabled node changes nothing and warns about nothing. That contradictory default was mine.

Both now ship OFF so the panel states what the canvas actually does, and the panel is labelled with the fix: install the pack, then select the node and press Ctrl+B to un-bypass it. The switch only routes; it cannot enable a bypassed node. If you do not want those packs, leave the switches off and lose nothing - they are speed and memory optimisations, not quality features. Everything renders identically without them, just slower.

The nodes stay bypassed on purpose. Enabling them by default would hard-fail every install that lacks the optimizer packs.

New in v2.2.2

One fix, reported by a user against 2.2.1. If you render with reference images, take this one.

Fixed: reference renders forced glasses onto the subject

Reported on Civitai: a reference image produced a character wearing thick black frames in every shot, and prompting to remove them changed nothing.

The pack injects a retention_analysis block alongside reference images so identity and voice stop drifting across a chain. That block was hardcoded to say the subject "retains the same face, skin, hair, glasses and wardrobe" - so every ref2va render was told to keep glasses whatever the prompt said. And it could not be argued with, because the sampler runs on a BasicGuider: cfg 1.0, no negative branch. There is nothing for a negative statement to subtract from, so "remove the glasses" only put the word into the conditioning a second time.

The block now says "the same face, skin and hair" and names no accessories. Eyewear, hats and jewellery belong to your prompt.

The rule this exposes, and it applies to every prompt you write for this graph: anything hardcoded into unconditional conditioning is permanent from your side. At cfg 1.0 you can add but never subtract, so phrase everything positively - "clear unobstructed eyes", never "no glasses". Negations do not work anywhere in this pack.

New in v2.2.1

A bug-fix release. Four defects, three reported from real renders on real machines rather than found in review. If you have been unable to finish a long chain, or a high-resolution one died partway with the whole server going down with it, this is the release for you.

Fixed: two crashes that only appear on long or high-resolution chains

Both were DefaultCPUAllocator: not enough memory - system RAM, not VRAM - and both killed the job after it had already done the sampling.

• master_normalize built the entire finished timeline four times over. A 12-shot 1088x1920 chain asked for a single 31.2 GB block and died on a 24 GB machine after 81 minutes of work. One of those four copies was pure waste: the code rebuilt a tensor byte-for-byte identical to one already in memory, purely to compute two numbers for a log line, then threw it away. Every statistic it needs is one-dimensional. It now measures per shot and releases each shot as it is consumed, so peak memory is one finished timeline plus one shot instead of four timelines. Output is bit-identical - verified against the old implementation across every colour mode and median width, on the pixels and on the log strings.

• The upscaler was handed every frame at once. 243 frames at 1472x2560 is an 11.0 GB float32 allocation on top of the input. It now runs in chunks into a preallocated buffer, so peak is one chunk rather than the whole batch. Bit-identical, verified at two chunk sizes. Set H3_UPSCALE_CHUNK if you want to trade memory for a little speed; default 16.

Fixed: auto-reserve could clamp itself into a hard crash

On a 12-shot 1280x736 run, shot 1 loaded completely and rendered fine. Shot 2 carries the context pin and the reference rows, so its activation-pool requirement roughly doubles - the node correctly costed it at 18.2 GB. It then clamped the pool to 9.4 GB "so the weights still load completely", and the weights loaded partially anyway, 401 MB offloaded. Neither constraint was met and the render aborted inside a CUDA kernel, taking the whole ComfyUI process with it.

Two causes, both fixed. The keepout was 384 MB, which did not cover ComfyUI's own buffer plus the difference between what the driver reports free and what ComfyUI treats as usable - about 400 MB short. It is now 1 GB. And the "this is tight" warning compared the clamped reserve against a previous shot's measurement rather than what this shot had just asked for, so a 9.4 GB clamp looked fine against shot 1's 9.1 GB while the payload needed 18.2. It now compares against the real request and says plainly that this usually dies inside a kernel, and that raising the reserve cannot help because the memory is not there.

Worth knowing regardless of this fix: shot 1 succeeding tells you nothing about shot 2. Shot 1 has no pin and no reference rows. If shot 2 will not fit, lower frames_per_shot or the resolution, or load a smaller quantisation of the DiT.

Fixed: start_image on the memory sampler silently did nothing

Reported by a user. start_image on H3MultishotMemorySampler is an identity reference row, not a first frame - and on an fl2va checkpoint, which has no reference rows, it is built and then ignored entirely. Meanwhile the sibling node H3MultishotSampler has an input with the same name that really is an I2V first frame, which is where the perfectly reasonable expectation comes from. The combination produced no warning at all.

It now prints one, naming both ways to actually open shot 1 on a picture: use H3MultishotSampler, whose start_image is a true first frame, or set continuity=flf_chain here and feed keyframe_images (N+1 stills for N shots). Note that continuity=first_frame is not about your image either - it hands over the previous shot's last frame, and does nothing on shot 1.

Verified

Three chained shots rendered end to end on ComfyUI 0.32.0 with the new auto-reserve: every shot full load: True, which is exactly where the crashing run went partial, and the new warning fired on shot 2 with real numbers and did not stop the render. master_normalize ran on that same 685-frame chain. The chunked upscaler is proven bit-identical by test rather than by render - that graph had upscaling off, so I will not claim end-to-end coverage it did not get.

New in v2.2.0

Everything since 2.1.2, which is where most people still are. Seven point releases in one: five separate defects that stopped the workflow running, a measured campaign against chain drift that changed the shipped defaults, and a whole ComfyUI version this pack could not previously run on.

(2.1.9 on GitHub and HuggingFace is this same content under a smaller number, tagged an hour earlier.)

Run-blocking, all user-reported or found by finally testing what we ship:

New controls: pin_frames, pin_noise, pin_renorm, and master_normalize's luma+contrast mode.

Changed defaults that change your output: memory_frames 2 → 0, and join_anchor_noise / handoff_release to 0 because they are inert under context_pin.

New runtime: ComfyUI 0.32.0, which needed real work — its ModelSamplingAV carries the audio half of the latent on a different scale. 0.30.0 is unchanged and still supported.

Every section below is the original release note for each of those versions, newest first.

---

New in v2.1.9

Fixed: H3_Seamless_Chain_CORE could not be queued at all

value_smaller_than_min: Value 0.0 smaller than min of 1.0 - output_scale

CORE's sampler still carried the widget array it was saved with before 2.1.3 removed the four two_pass_upscale dials. Nineteen saved values against the class's fourteen live widgets, so the frontend put false into output_scale (minimum 1.0) and 1.5 into save_every_shot, and the server rejected the whole prompt. The workflow advertised as the one with no third-party dependencies — the safest thing for a new user to open — has been un-runnable since 2.1.3.

It was never caught because CORE had never been rendered end to end. Our own release checklist said "one CORE render before posting" and that step had been carried, unticked, through six releases.

The array is now generated from a name→value map resolved against the server's schema, and the map is stored on the node so the pack's JS can re-apply it by name — the same treatment the full workflow's sampler got in 2.1.6. CORE has now rendered: three chained shots at 960x544, 370 frames, picture and audio, holding framing, wardrobe and lighting across both joins.

All three bundled workflows are now submit-tested against a live server as part of the release routine, not read.

Bypassed nodes now say what they need

Four nodes ship bypassed because their packs are not in the zip, and a missing class fails the entire prompt. That is the right default — but a bypassed node sitting on the canvas invites you to un-bypass it, and doing that without the pack installed breaks the workflow with no explanation. Joe's point, and he is right.

Now it says so in four places, in order of how hard they are to miss:

The two-step rule is the part people lose: un-bypassing a patch node changes nothing on its own, and flipping its switch while the node is still bypassed changes nothing either. Both, in that order, after installing the pack.

SCRIPT PREVIEW is the odd one out — it is a leaf, so leaving it bypassed costs you the on-canvas preview and nothing else.

Why not just bundle the packs? The writer pack is bundled because it is modified and pinned. These three are not: shipping a second copy of Custom-Scripts, sol-attn or blockcache-T8 inside the zip would shadow whatever the user already has installed and freeze it at the version we happened to vendor. Naming them and linking them is the honest version.

---

New in v2.1.8

Fixed: one node in the workflow came from a pack that is not in the zip

SCRIPT PREVIEW is ShowText from ComfyUI-Custom-Scripts, and it shipped active. A node whose class is missing serialises as class_type: null and the server rejects the whole prompt, so without that pack installed the workflow could not be queued — for the sake of an on-canvas text box. INSTALL.md said to delete the node, which only helps someone who reads it before pressing Queue.

It now ships bypassed, the same remedy the three accelerator nodes got in 2.1.4. A bypassed node is dropped from the prompt entirely. Install Custom-Scripts and Ctrl+B the node if you want the preview; the writer feeds the sampler either way.

Found by mapping every node type in all three bundled workflows to its owning Python module against a running server. That check is now part of the release routine instead of something I do after a user tells me. The other two workflows were already clean, and everything else in the full one is either core ComfyUI, this pack, the writer pack that is in the zip, or one of the three bypassed accelerators.

Verified end-to-end on ComfyUI 0.32.0

Not by reading the graph — by loading the shipped file in a browser on a 0.32.0 install and pressing Queue. Twice: once through the prompt-writer lane exactly as shipped, and once with a hand-written script.

• loads with no missing node types, and every sampler widget lands on its own name (memory_frames 0, master_normalize luma+contrast, pin_renorm on) — the shift that caused The value 1 for reference_image_size is not available cannot reproduce
• no null class_type: the bypassed preview and the three bypassed accelerators are all dropped cleanly
• three chained shots at 960x544, 124 frames each → 328 frames, exactly 372 minus the two 22-frame context_pin head trims
• a reviewer given the clip cold, with no idea how it was made, read it as one continuous static take, found no cut or jump anywhere, transcribed all three lines of dialogue, reported the lip-sync as matching, and found no shift in framing, colour, brightness or wardrobe and no hiss, dropout or click in the audio

0.30.0 remains the version everything else here was measured on; both are now tested before release.

---

New in v2.1.7

Three things that stopped the workflow running. All user-reported, all reproduced, all fixed and verified by render rather than by reading the code.

Everything below was verified on ComfyUI 0.32.0, not just 0.30.0. Two of these three only ever appear on 0.32, which is why they survived several releases.

Fixed: the Audio Spine produced static on ComfyUI 0.32.0

guide_audio came out as hiss while the same file through voice_ref or the native node was perfect. Reported against 2.1.2, still present through 2.1.6.

ComfyUI 0.32.0 introduced ModelSamplingAV, which carries the audio half of the packed AV latent scaled onto the video schedule — process_latent_in multiplies it by shift / audio_shift (12/3 = 4 for H3) and process_latent_out divides it back. Everything this pack injects into the sampler's latent is in the stream's native domain, so on 0.32 it landed 4x too small and decoded as broadband noise. On 0.30.0 there is no such scaling, so the same code was correct — which is why it never reproduced here until a 0.32 rig existed.

Three injection sites had it, not one:

All three now scale at the point of use, reading the factor off the live model_sampling object so a changed sigma shift stays correct. getattr's 1.0 default leaves 0.30.0 byte-identical.

Verified on 0.32.0 with a real 44.1 kHz stereo voice track: loudness-envelope correlation against the guide +0.965, speech-band energy 34.2% against the guide's 39.3%, and a blind listener transcribing the guide's words with "clean, no background hiss". Before the fix the same render was hiss.

Fixed: naming an mmproj file broke GGUF encoders

mat1 and mat2 shapes cannot be multiplied (3680x1152 and 3456x1152)

thrown at the handoff into shot 2, GGUF encoders only, unaffected by resolution or by turning every image input off.

Setting mmproj_name explicitly took a different code path than (auto) and skipped the vision key-renaming step entirely — so the vision tower loaded under raw llama.cpp names (v.blk.*, mm.*) that nothing downstream reads, the merger was never populated, and the first matmul touching vision features had the wrong width. Same file, two loaders, 19 key names in common out of 351.

The bitter part: mmproj_name is documented as the reliable escape hatch for when filename pairing fails, and it was the broken path.

It now runs the same post-processing as (auto), using ComfyUI-GGUF's own key map rather than a copy, so it follows their changes. Verified: an explicitly named file now yields a state dict identical to (auto) — 351 tensors, every shape matching. A new guard also fails by name if a chosen mmproj produces no visual.* tensors, instead of dying in a matmul twenty minutes later.

Fixed: The value 1 for reference_image_size is not available

The shipped workflow's saved widget values were written for a layout that did not present sampler_override and scheduler_override as widgets. The current schema does, so everything from index 27 read two slots early and output_scale's 1.0 landed in reference_image_size, a combo of match/max. This affected 2.1.3, 2.1.4 and 2.1.5.

The array is now generated from a name→value map resolved against /object_info, and the same map is stored in the node's properties so the pack's own JS can re-apply by name if a future schema change shifts anything.

Also: memory_frames now defaults to 0

The bank's recent slots hand each shot's accreted output forward as reference images on top of the latent pin, so invented detail compounds. Measured over ten shots at 960x544, moving 2 → 0:

The last row matters most: at 2 the drift accelerates, which is what a runaway loop looks like. At 0 it holds flat.

The obvious worry was motion continuity, since the recency slots exist to carry it. Tested on a scene with continuous large-amplitude movement: anchor-only retained motion slightly better (−5.9% vs −6.5% over four shots) with better framing (0.983 vs 0.971). The cost does not exist. If a busy scene ever does lose continuity between shots, raise it to 1.

join_anchor_noise and handoff_release now ship at 0 — both are inert under context_pin (one noises keyframes the mode never creates, the other belongs to latent_handoff) and non-zero values read as tuned settings while doing nothing.

---

New in v2.1.6

Chained shots stop getting brighter-edged every hop. Not by the route 2.1.5 claimed - see the correction below.

master_normalize matched every frame's MEAN to one global target, which is why a chain shows no brightness step. It was also masking a second drift it never touched. Measured on a 3-shot chain at 960x544: mean held flat, 27.31 -> 27.43, while the DISTRIBUTION stretched - p25 fell 6 -> 2 and p95 rose 85 -> 96. Contrast climbing every hop, re-centred each time, and handed to the next shot as a higher-contrast starting point.

New: master_normalize=luma+contrast (now the default) matches the spread as well as the mean. Rescaling amplitude about each frame's own mean is an affine remap: it moves no edges, so it is not the blur this pack has always ruled out for texture drift.

Texture growth per hop, from a log fit across all shots. 1.000 is no accretion:

The baseline ratchet scales with canvas (1.130 at 640x352, 1.199 at 960x544); after normalising it stops caring (1.055 vs 1.064). There is nothing in it to tune per resolution - it works on decoded frames with per-frame statistics against one global target.

The target anchors to shot 1, not the timeline median. Contrast only ratchets upward, so shot 1 is the one frame-set with nothing accreted onto it; a median target pulls shot 1 UP to meet the drift (+11.7% texture, for no benefit) where anchoring to shot 1 leaves it untouched (+0.1%) and only ever pulls later shots down.

1:1 crops of the last shot show no loss of real detail - lamp vent slots, hinge rivets, hair strands and knit weave all survive. What leaves is the invented crispness.

What is left, honestly

About 1.05 per hop. That residual is spatial accretion, and this pass cannot reach it: master_normalize runs on the finished master, outside the feedback loop, so it cleans what you see while the next shot is still handed the inflated pin. Four shots is slight. Ten shots is roughly +50%. If you are chaining long, expect it.

Correction to v2.1.5

v2.1.5 said pin_noise=0.05 fixed this. It does not. That was measured on two seeds of a single scene at 640x352 - a scene whose background was nearly black and which barely ratcheted to begin with - and the control that would have caught it, pin_noise=0.00 at the reporting user's own resolution, had never been run. With it run:

Both on a detail-heavy scene. The dial is small and scene-dependent, it cannot touch the dominant drift in a busy frame, and above 0.10 it gets worse (0.20 measured 1.228 against a 1.211 control). Its range now stops at 0.10 and its tooltip says all of this. It stays in the pack because it costs nothing and does help where the ratchet is already small; it is not the fix.

Not tested

Portrait canvases. Everything above is landscape - 640x352 and 960x544. The mechanism is resolution-independent by construction and the two landscape sizes agree, but 768x1344 and 736x1280 have not been measured and are not claimed.

Measuring this yourself

Texture comparisons are only meaningful while the framing holds. If the model cuts to a different setup, texture reflects content and the number is meaningless - one portrait run here scored a flattering 0.878 per hop purely because shot 3 cut to a close-up of a film reel. Correlate each shot's mean frame against shot 1's before trusting any of it; a held framing sits above 0.95.

That cut is worth a writing rule of its own: do not name a nearby object in a shot's closing beat. "She glances down at the reel" reads as a request for a shot of the reel. Keep closing beats on the speaker's own body.

---

Fixed in v2.1.4

• The main workflow would not queue without two third-party packs. H3_Seamless_Chain_v2.json ships three optional accelerator nodes - sol_attn and chunk_ffn (ComfyUI-sol-attn) and block_cache (comfyui-minimax-h3-blockcache-T8) - and they were saved active, so a clean install hit missing_node_type: Node 'attention patch (gated)' has no class_type and nothing ran, even though the shipped recipe keeps all three gates off. They now ship bypassed: dropped from the prompt entirely, with the model passing straight through. To use one, install its pack, select the node, Ctrl+B to un-bypass, then turn its gate on.
• Value 4 bigger than max of 3: memory_frames on a workflow you never edited. v1.2 inserted seed_per_shot into the middle of the sampler's input list. Widget values are positional, so every dial after it shifted by one in workflows saved on v1.0/v1.1 - your old anchor_frames arrived as memory_frames. Pre-v1.2 workflows are now repaired on load (save to make it permanent), the sampler also stores values by name so no future change can shift anything, and the validation error explains the shift instead of naming the wrong dial.

---

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

• Guides: the 5-minute guide and every setting explained
• 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.