Support
Everything here is free and stays free — the format spec, the nodes, the workflows, the cartridges, the LoRAs. If it saved you a night of debugging (it contains several hundred of mine), tips keep the 5090 warm:
🔁 Liberapay (recurring)
⚡ Or right here: the Civitai tip button on this page sends Buzz directly.
2.6.0 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 - and give the writer one premise. The panel sizes a window for your card and the count that fills the time; the writer writes ONE speech and cuts it across the windows at sentence boundaries; the chain continues it in H3's own voice, no TTS, no per-shot budgeting. Verified: seven renders including a 65-second seven-window take, every join continued the speech. Details in this version's notes.
Known limit in 2.6.0: the chain's texture ratchet is not fully solved for long takes - we measure about +13% fine texture per join at 736x1280 with every anti-drift setting on. Under about 4 windows (30-40 seconds) it is slight; at 7 windows it is visible sharpening. Keep extend takes to about 4 windows for now; a fix is in progress.
2.5 is the memory release
Four new memory systems, all measured, two of them fully automatic:
The random-slowdown fix (automatic). High-resolution renders used to take anywhere from 27 minutes to 3 hours for identical work - a lottery caused by the Windows driver quietly demoting GPU memory when the card fills past roughly 95%. The pack now detects that zone and 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, after all the sampling was already paid for. Switch it on and shots stream to lossless disk staging as they finish; peak RAM stays near two shots no matter how long the chain, with verified-identical output.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 and ComfyUI's own EasyCache, each measured (-11% to -29%) and eye-tested. All ship OFF: the three real ones are faster but can visibly distort people, and block cache turned out to do nothing at 14 steps (0 hits measured on every run) - honest notes on each. Defaults across the workflows are retuned for 16-24 GB cards. Full details in this version's notes.
Guides
Make talking videos with MiniMax-H3: the 5-minute guide - what to open, what to type, what each setting costs your card.
Every setting explained: the deep manual - every workflow, every dial, what it does and what breaks if you move it.
What this is
A ComfyUI node pack and three 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. 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.
How identity holds with no reference images
This is the part most people do not expect. A 40-second two-character scene held both faces across every shot with zero reference images supplied. Two mechanisms do it, and neither works alone.
The frame relay carries the instance. Every shot begins from an actual rendered picture of the character - the previous shot's last frame - so the specific face, the wardrobe weathering, the exact hair propagate as pixels rather than being re-imagined from text. Drift can only accumulate per join, not per frame.
Verbatim text carries the category. Repeating each character's appearance description word-for-word in every shot re-asserts everything a picture cannot pin down, and authors anyone who walks into frame partway through. Once they are in a handed-over frame, they join the relay too.
Reference images and the identity bank are still there for long chains, characters who leave frame for a long stretch, or when you need a specific pre-existing face - they are just no longer the price of entry.
Voice identity across the chain
Chained shots used to re-perform the voice from text each time, and it drifted. Two controls end that, and the shipped default turns one of them on:
voice_ref- feed a clean solo speech clip and it anchors every shot as<Audio 1>. Mono clips, sample rate and length are all guarded internally.self_anchor_voice- no file at all. Shot 1's own rendered voice becomes the reference for every later shot. Write shot 1 so the character speaks a clean solo line.
Both ride the checkpoint's reference rows, so they need ref2va. On fl2va they do nothing except cost tokens - see the checkpoint note below.
Which workflow
Three, each with one clear reason to exist.
H3_Seamless_Chain_v2 - everything. Master controls, the LLM writer, the speed boosters, the remote encoder lane, identity and voice anchors, an episode/batch prompt source, FFLF boundary plates and an audio spine. The optional lanes ship gated off, so the default path stays simple.
H3_Seamless_Chain_CORE - the same job with zero third-party packs. Type your shots into the script box and queue. Start here if you want a render before installing anything else.
H3_Keyframes - a different job: a hand-built sampling graph for anchoring a single clip at chosen frame positions with per-anchor condition strength. Not multishot, does not chain.
Master controls, and the dials that matter
One panel drives resolution, frames per shot and steps for the sampler, and also feeds the writer so it paces dialogue to the real shot length. Shipped defaults: H3_Extend_Take 1280x736 landscape, take_seconds 30; H3_Seamless_Chain_v2 736x1280, 192 frames x 4 shots; 14 steps, euler/beta57, the curve-Q5_1 checkpoint. The resolution floor is deliberate - the base model distorts faces below roughly 1 megapixel, so shrinking the canvas to save VRAM costs you faces first. The dials worth knowing:
continuity-context_pin(raw-latent join, needs the Motion Context pack),first_frame(the model's own trained hand-off, no extra pack),cutfor episodic work, orflf_chainfor boundary plates.seed_per_shot- leave it on. Measured: varying the seed per shot holds the face; one seed for every shot drifted both face and voice. Identity lives in the conditioning, not the seed.chain_gain_control- ships atflatten. Each shot's tail anchors the next and the model returns roughly 1.3x the anchor's texture energy, so sharpness ratchets across a chain without it.Speed boosters - all ship OFF. Spectrum, TeaCache and EasyCache are real speed but can visibly distort people; block cache does nothing at the shipped 14 steps (0 hits measured on every run - only worth it at 30+ steps). The three real ones - fine for empty rooms and landscapes, compare for yourself before using them on faces. Never stack Spectrum with TeaCache: no extra speed, badly damaged output.
low_ram_master- turn ON for long chains or under 32 GB of system RAM. Shots stream to lossless disk staging instead of accumulating in RAM; identical output, verified.remote_encoder(VRAM / SPEED SWITCHES panel) - OFF by default. With a second ComfyUI PC on your network, turn it ON and the text encoder's 15+ GB moves over there. The how-to note sits next to the encoder node.output_scale/upscale_model- upscaling happens after decode, per shot, so it works with every continuity mode includingcontext_pin.output_scaleis a lanczos resize (measured 1.78x faster than rendering the same output size natively, and visibly softer - resolution, not detail);upscale_modeltakes ComfyUI's own Load Upscale Model (ESRGAN and friends) for synthesized detail.master_normalize- leave it atluma+contrast. It levels both the brightness AND the contrast of the finished chain against one global target taken from shot 1. Measured: texture growth per hop 1.126 -> 1.047 at 960x544, with contrast spread across shots falling from ~11% to under 0.5%.memory_frames- leave it at 0. The bank's RECENT slots hand each shot's accreted output forward on top of the pin, so invented detail compounds. Measured over ten shots: at the old default the drift accelerated; at 0 it holds flat, and motion continuity did not suffer.pin_frames- leave it at22. Longer pins also cut the sharpening, but the head trim scales with them, so speech planned under the replay gets clipped.Audio drifts the other way: chained audio gets duller per hop where chained video used to get sharper. There is no per-shot dial for it. For cut-grammar content it can be avoided outright:
continuity=cutwithbank_pinned=1, memory_frames=0conditions every shot's audio on shot 1 alone - measured flat.preview_first_shot- writes shot 1 the moment it decodes, so a bad take can be cancelled before the whole chain renders.
Which checkpoint
ref2va ships as the default, as the curve-Q5_1 GGUF (~14 GB): it carries the reference rows that voice anchoring and the identity bank need. fl2va has no reference rows and chains just as well - the voice rides the frame relay instead of being pinned - and it lands on a supplied frame where ref2va only nudges toward one. The two are the same file size at every quant level.
Writing a script
One prompt per shot, --- on its own line between them. The mechanics of a join are handled for you; whether the join reads as invisible is decided by the writing. These rules were each written after a specific failure:
The airlock. Every shot after the first opens holding the previous shot's exact closing arrangement, with about two quiet seconds before anyone speaks. The head of a chained shot is discarded on decode - anything you put there is thrown away.
Give the hold something to do. A held framing with nothing happening renders as a literal freeze. Write in a breath, a weight shift, an eyeline change.
Land settled. End each shot back in a stable arrangement, dialogue finished, about two seconds spare.
A line never straddles two shots. The LLM writer budgets this automatically now. Hand-writing? A chained shot at 192 or 243 frames fits roughly 6-10 spoken words; 362 frames fits 12-20. Over budget renders as crammed, garbled speech; far under budget renders slow and robotic.
Repeat the descriptions word-for-word. Each character's appearance and the room and lighting, identical in every shot. Rewording between shots is the single most common cause of a face changing mid-scene.
Silent shots with visible people: say what mouths are doing ("lips pressed closed") or the model invents mumbling. A thing revealed mid-shot is written as already present, or it pops in.
Camera cuts inside a shot are free and share one audio bed - so put cuts inside shots, and continuity between them. A worked four-shot two-hander ships as example_script.txt.
Quick fixes
A word clips at a join - the script put dialogue too close to a boundary. Move the whole line, do not split it.
Sharpening increases every shot - the texture ratchet.
chain_gain_control=flattenships on; check it was not turned off.A render randomly crawls - as of 2.5.0 the pack detects and fixes the main cause automatically (watch for the "driver headroom" line). If it still happens: drop frames or resolution, or load a smaller quant of the DiT.
The machine dies at the final join on a long chain - turn
low_ram_masterON.People look subtly wrong - if you enabled Spectrum, TeaCache or EasyCache, that is the trade. Turn them off; keep block cache.
LLM API error 404, model not found - the writer's
model_nameis not pulled on the endpoint it is pointed at. See the prompt-writer section below.Red or missing nodes - an optional pack is not installed. The optional nodes ship bypassed, so this only bites if you un-bypassed one: re-bypass it with
Ctrl+B, delete it, or use the CORE workflow.GGUF architecture error - the pack teaches ComfyUI-GGUF the
minimax_h3architecture at startup, so installing it is normally the whole fix. If the error persists, runpython apply_gguf_arch_patch.pyfrom the pack folder once and restart.Audio dulls on a very long chain - expected; restart the chain on a scene cut, where a fresh start costs nothing.
The prompt writer needs a model you actually have
The full workflow ships pointed at a local Ollama with model_name = qwen3:14b. Pull it before the first queue or the run stops immediately with LLM API error 404: model 'qwen3:14b' not found:
ollama pull qwen3:14b
Any OpenAI-compatible endpoint works - put its URL in base_url and the exact tag in model_name; ollama list prints the tags you have and it must match character for character. A remote or hosted endpoint is often the better choice, because a local writer large enough to be good competes with H3 for the same card.
Running the writer locally? Turn on unload_model_after on the writer node. It frees that model from Ollama the moment the script is written, so the video model gets the card - otherwise it sits for the server default of five minutes, which is your whole first shot.
Want no LLM at all? Set the master panel's use_file_prompts to manual entry, delete the writer, and feed your own shot script straight into the sampler's script input - one prompt per shot, separated by --- lines. The CORE workflow already works this way.
Requirements
Always: ComfyUI v0.30.0 through v0.33.1 - tested at three points and supported across the range (native MiniMax-H3 support), this node pack, a MiniMax-H3 checkpoint, the H3 text encoder, and the video and audio VAEs. The CORE workflow needs nothing else - it is built from this pack plus stock ComfyUI nodes.
For the FULL workflow. ComfyUI validates every node class in a graph before it will queue, so a missing pack stops the whole workflow - not just its own feature. Each can be removed instead; INSTALL.md gives the one-widget change or node deletion for every entry.
ComfyUI_JoyAI_Echo_GGUF_Nodes - the LLM prompt writer (
JoyEcho_LLMEnhance). It is inside the release zip - copy it intocustom_nodes/with the main pack. It is RealRebelAI's pack, modified with attribution (the NOTICE inside lists every change); use that copy, not upstream - the workflow drives inputs the upstream release does not have, and with upstream those values are dropped silently. Hand-written scripts can delete the writer instead.ComfyUI-H3-Motion-Context -
continuity = context_pin, the shipped default, plus interior keyframe anchors. Removal:continuity = first_frame.RES4LYF - the
beta57scheduler the full workflow ships with. Removal:scheduler = beta(measured cost: lip-sync 8/10 vs 10/10, everything else equal; CORE shipsbetaalready).Booster packs (Spectrum, TeaCache, block cache) - each switch names its pack and prints an install link if it is missing; the graph queues either way. EasyCache is built into ComfyUI, nothing to install.
ComfyUI-Custom-Scripts - the in-canvas script preview, a leaf node; removal: delete it.
ComfyUI-GGUF - quantised checkpoints and encoders.
Install in one click: search H3 Multishot in ComfyUI-Manager - the pack is on the Comfy Registry as comfyui-h3-multishot. The zip on this page stays the complete bundle - both node packs, the workflows and the docs.
Models
GGUF quants of the H3 checkpoint are at huggingface.co/joeygambino/MiniMax-H3-GGUF - Q8_0 for 32 GB, Q5_1 for 24-32 GB, Q4_0 for 16 GB. The shipped default is curve-Q5_1 (~14 GB), which the new automatic headroom rule keeps comfortable on 16 GB cards as well. The text encoder and both VAEs are at huggingface.co/Comfy-Org/MiniMax-H3. GGUF encoder quants are at huggingface.co/joeygambino/MiniMax-H3-encoder-GGUF; keep the -mmproj vision sidecar, and load it with this pack's H3 CLIP Loader (safetensors + GGUF) rather than the stock GGUF loader - the vision path is what carries frames between shots.
What is verified, and what is not
Stated plainly, because a release note is a claim.
Verified: chained renders reviewed blind by two independent models and reported as one continuous take with consistent colour and unbroken audio; a 40-second two-character scene holding both faces with no reference images; the streaming master at 42.8 dB against the in-RAM path; the remote encoder identical across two machines; the driver-headroom rule on both its branches; every booster measured on the same seed and eye-tested on finished video.
Not verified: very long chains still dull the audio slightly per hop.
flf_chainhas not been rendered against a fully colour-matched plate set.
Credits
The context_pin join and interior keyframe anchoring build on ComfyUI-H3-Motion-Context by NikoDemon80. The script preview uses ComfyUI-Custom-Scripts by pythongosssss. GGUF loading builds on ComfyUI-GGUF by city96. Speed boosters integrate ComfyUI-Spectrum-MiniMax-H3 by xmarre, ComfyUI-MiniMaxH3-TeaCache by Icyoung and comfyui-minimax-h3-blockcache-T8 by T8mars. Thanks to @viralesveras for contributions to earlier versions.
Free, and it stays free. If it saved you time: Ko-fi · GitHub Sponsors · Liberapay.
Description
Everything here is free and stays free — the format spec, the nodes, the workflows, the cartridges, the LoRAs. If it saved you a night of debugging (it contains several hundred of mine), tips keep the 5090 warm:
🔁 Liberapay (recurring)
⚡ Or right here: the Civitai tip button on this page sends Buzz directly
The joins are gone. Two independent video-understanding models were shown a four-shot chain from this version, blind and with no context. Both described it as one continuous unedited take - no cuts, colour consistent, audio unbroken.
That is the whole point of 2.0. Everything below is either what made it possible or what got cleaned up on the way.
A second way to join shots
Chaining used to mean handing the next shot the previous shot's last frame. That still works, is still shipped, and is still what the CORE workflow uses. 2.0 adds continuity = context_pin, which is a different mechanism entirely.
The previous shot's last 22 frames ride into the next shot as raw latents - bit-identical, never decoded to pixels and re-encoded - placed at interior keyframe coordinates, with a timeline-placed audio reference alongside them. The regenerated head is trimmed on decode. Colour, motion and voice cross the boundary as data instead of as a description, which is why the join stops being a place where things can drift.
This mode needs the ComfyUI-H3-Motion-Context pack. Without it, set continuity to first_frame - the model's own trained hand-off, no extra install, and what the CORE workflow ships with.
Seam audio that stops eating words
The boundary audio cut used to land blindly at the start of the incoming shot, which shaved the attack off any word the model happened to place there - audible as a tiny blip at a join. It now lands in the quietest gap within the shot's first 0.75 seconds, then welds with the same 40 ms equal-power crossfade. A line that begins early survives intact.
Voice and identity anchors, on the chaining sampler
The memory-bank sampler now takes the same anchor inputs the single-shot sampler always had: reference_images, voice_ref, self_anchor_voice, reference_image_size, preview_first_shot, two_pass_upscale and its three dials, plus sampler_override / scheduler_override so one master panel can drive the sampler and scheduler by name.
The bank already carried voice - each slot is a video reference with the audio under it - but only from shot 2 onward, because shot 1 renders against an empty bank. self_anchor_voice closes that gap: shot 1's own rendered voice becomes <Audio 1> for every later shot. Operator-supplied references are placed ahead of the bank slots, so <Picture n> and <Audio n> numbering stays fixed as the bank fills and a prompt's bindings cannot drift mid-chain.
Two-pass upscale cannot be combined with continuity = context_pin or latent_handoff, or with an audio spine. Those carry the previous shot's raw latents - or one locked denoise trajectory - across the join, and a two-pass render preserves neither. The node stops with an error naming the conflict rather than quietly producing a weaker join. Two-pass is available on cut, seamless, seamless_tail, first_frame and flf_chain.
A VRAM and speed panel
One panel, three lazily gated model patches - memory-efficient attention, chunked feed-forward, and a block cache - plus an activation-reserve control. Every switch off reproduces the shipped recipe exactly, and a gate that is off means the patch node never executes at all, so there is no cost to leaving them alone.
And a reserve that understands payload
Memory measurements are now keyed by shape and conditioning payload, so a bare first shot and a reference-laden later shot no longer share one estimate. Measured pools are no longer overridden by a fixed floor, an unseen payload variant estimates from a measured sibling, and a spill into system RAM is now detected and named in the console rather than presenting as an unexplained five-times slowdown with nothing in the log.
The writer knows the join rules now
The LLM prompt writer gained a join_style control. Set it to a chained style and the render-verified boundary rules are appended to its system prompt automatically - open each shot holding the previous shot's closing arrangement, land settled, never straddle a line across a boundary, repeat every description verbatim. Generated scripts obey them without you memorising anything. Hand-written scripts still need to follow them; they are documented in PROMPTING.md with a worked four-shot example.
Three workflows instead of five
H3_Seamless_Chain_v2 - everything, with the optional lanes gated off by default: master controls, LLM writer, VRAM panel, identity and voice anchors, an episode/batch prompt source, FFLF boundary plates, and an audio spine.
H3_Seamless_Chain_CORE - the same job with zero third-party packs. Type shots into the script box and queue. Start here.
H3_Keyframes - unchanged, and still its own thing: single clip, anchors at chosen frame positions, per-anchor condition strength.
H3_Multishot_AIO and H3_Multishot_MEMORY retire. Every lane they had is in v2 - the AIO's episode source, plate chain and audio spine were folded in, and MEMORY had nothing v2 lacks. Your existing copies keep working; there is just no longer a reason to open them.
Smaller things that were worth fixing
flf_chainselected with no boundary plates wired now stops with a clear error instead of quietly rendering an unanchored chain.Interior keyframe anchoring and the Motion Context pack no longer fight over the same patch site. If that pack is installed it owns the site and this one stands down with a line in the log; if it is not, this one fills the gap. First and last anchors work either way.
GGUF vision sidecars can be named explicitly. ComfyUI-GGUF pairs the
-mmprojfile to a GGUF text encoder by filename, in the encoder's own folder only - rename either, or split them up, and it loads the encoder without its vision tower, which presents as the model ignoring your reference image. This pack's CLIP loader now raises instead of continuing blind, uses the only mmproj beside the encoder when there is exactly one, and takes anmmproj_namewidget so you can point at the file directly - so the filename rule documented on the encoder listing stops being something you have to obey.The pack now imports defensively - one module failing no longer takes every node in the pack down with it.
Shipped defaults
ref2va checkpoint, continuity = context_pin, voice self-anchor and the identity bank on, euler + beta57, 14 steps, 362 frames per shot (~15.1s), 24 fps, every VRAM switch off.
Blind review ran on the lighter configuration - fl2va, no voice anchor - which is the same chaining mechanism with fewer reference rows. ref2va ships as the default because it makes voice and character identity explicit rather than emergent; switching to the reviewed path is two changes, documented in SETTINGS.md.
Honest limits
Audio dulls slightly per hop on very long chains - restart the chain on a scene cut, where a fresh start costs nothing.
Resolution cannot change mid-chain.
flf_chainhas not been rendered against a fully colour-matched plate set.Two-pass upscale is off by default and excluded from the raw-latent continuity modes (above), so the seamless chain has no built-in upscale path yet. Upscale after the fact, outside the graph.
Keep the mux at 24 fps. Other rates audibly shift voice accents; that is the model's audio lane, not the muxing.
FAQ
Comments (39)
the AIO script seems to be closer to what I want, the ref2a flexibility is really good and I trust it to preserve what I want at the seams more, the only thing lacking is sometimes tricky to get it to seamless transition instead of jump into slightly different positions.
Try latest v2.1 - seamless transitions built in by default.
Thank you for all the work you are putting into these workflows, it's appreciated.
Unfortunately I'm have difficulty with v2.0 as I can't find JoyEcho_LLMEnhance or JoyEcho_PromptSource anywhere. Of course, they just so happen to support the feature I'm eager to try. Any pointers?
Update1: I've installed JoyEcho_LLMEnhance from RealRebelAI's ComfyUI_JoyAI_Echo_GGUF_Nodes pack. Still looking for JoyEcho_PromptSource
Update2: I had to drop joyecho_prompt_source.py from HF joeygambino/joyai-echo-multishot-workflow into custom_nodes\Comfyui_custom_scripts folder. I don't know if it's vital but I removed the second underscore in the file name as that was the name thrown by error in Comfyui.
I think I'm ready to go, I'll leave this here in case it helps anyone else.
v2.1 going up shortly to fix some bugs I didn't catch locally and will take care of this. Sorry!
Both v2.0 and 2.1 do not seem to work with continuity=context_pin even with ComfyUI-H3-Motion-Context installed.
[WARNING] h3_motion_context: another pack has already patched MiniMaxH3.extra_conds (it now comes from '/home/vlady/apps/ComfyUI/custom_nodes/ComfyUI-H3-Multishot.h3_avbank_probe'). Both packs are solving the same keyframe/ref collision and they cannot both own it, so this one is refusing. Disable one of them and restart.
[ERROR] !!! Exception during processing !!! h3_motion_context: the payload patch could not be applied. Without it the audio ref would overwrite the pinned video latents and the motion context would be lost. The reason was logged just above this error.
A different issues found:
continuity=seamless behaves as a cut, not seamless at all.
continuity=seamless_tail errors out with T2V mode AFTER sampling, not before:
[ERROR] !!! Exception during processing !!! only first/last keyframe anchors are supported
I'm in search of seamless T2V (and I2V) clip chaining and could not find working setting in the current version. Both identity anchor gate [OFF] and FFLF PLATES gate [OFF - flf_chain only] are set to T2V
@vladulidlo All three confirmed, and thank you - this is an excellent report. 2.1.1 is up with the fixes.
context_pin + Motion-Context: my pack was grabbing the same patch site Motion-Context needs, before their pack could. Their code publishes a compatibility marker for exactly this situation; mine now honours it, so the two coexist and load order no longer matters. Verified with a live context_pin render.
seamless_tail: real conflict - it needs interior keyframe anchors, which collide with Motion-Context's ownership of that patch math. It now stops before sampling with a clear message instead of dying after your first shot. With Motion-Context installed, use context_pin - it's the stronger mechanism and what that pack is for.
seamless: you're right, and the tooltip now says so - it's a legacy latent-only soft pin kept for comparison, and it often reads as a cut. For seamless T2V chaining use context_pin (or first_frame on an fl2va checkpoint). Both are the measured, working paths.
The bug never showed on my machine because of an install-layout difference that disabled the conflict detection - also fixed, and my release testing now runs on a packaged clean install so this class doesn't slip through again.
@joeygambino
Thank you! For both the fixing and fixing it so quickly!
Seems like the bugs started eating into the bugs, chill, don't rush, take your time, customers can wait.
Ha, thanks. I do tend to rush when I have a new feature to show off. A lot of bugs don't pop up until someone reports them, because the workflows are functioning perfectly for me, but then I realize the things other people just don't have installed.
@joeygambino yeah people are too excited for new model and what they can do. also samples looks good, compared to previous degrading over time was visible, now it looks stable through all 30 secs. will be trying lastest WF later on, good job!
I would like to point to a fork https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop
ethanfel's fork is 51 comits ahead of original https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context, but also not (yet) compatible with ComfyUI-H3-Multishot.
So if you are seeing:
RuntimeError: continuity=context_pin needs the ComfyUI-H3-Motion-Context pack installed (github.com/NikoDemon80/ComfyUI-H3-Motion-Context)
You could have downloaded the fork instead of https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context
Thank you - you're right, and I've verified it against the fork's source. It registers 18 node ids (MiniMaxH3LoopTrim, the MiniMaxH3Chain* family, the Scheduled* reference nodes) and deliberately does not re-register MiniMaxH3MotionContext. From its own init.py: "The original Motion Context, Save Latent, and Load Latent ids remain exclusively owned by Niko's upstream pack."
So, the fork is a complement rather than a replacement - install both. They're built to coexist, and this pack works with one's runtime patches because all three honor the same patch-ownership markers.
My error message was unhelpful about that, so 2.1.1 now detects the fork and says exactly this instead of just naming a repo you thought you'd installed. It's also documented in INSTALL.md.
That fork looks well worth a look on its own merits, by the way, a disk-backed chain/loop system with review gates and checkpoint resume is solving a different problem than this pack and solving it further than I have.
@joeygambino Good to know! I'm glad you find the forked repo interesting as I do.
Just installed 1.5 last night, after wondering to myself "This new H3 stuff is amazing, just wish I could load more than start/stop keyframes" and looking at civitai for a lark. Blown away by this stuff - once 2.0 or its successors settle out from the bugs, I'll update, but this is already incredible, bordering on revolutionary for me to mess around with. Well done!
Thank you! 2.1.1 should mostly be bug free now - but if you find anything, I try to be quick about fixing things. Sometimes stuff that works perfectly on my machines, don't necessarily work well on others due to different Comfy versions, hardware, node packs installed, etc. I don't know something is broken until someone tells me.
I have noticed that this larryvrh/MiniMax-H3-Turbo-Lora is not burning the output like the lightning one does, anyone else been using it with this workflow?
The only one I've used so far is minimax_h3_turbo_4step_ckpt500.safetensors
And I can't even say I've tested it enough to know it works. When I did test it, it didn't seem to work very well with my workflows unless it was at 12 steps + only Euler - which at the time it was in Alpha so I didn't bother testing further and figure people will use their own anyway. I didn't realize he'd put a hundred more options up though, so I may have to give them a shot.
I've tried to optimize things enough so you don't need a lightning/turbo lora, and 4 steps on a video render seems nuts to me, but I suppose I should give it a shot.
@joeygambino with latest V4 he even says 4 steps is not enough need to be minimum 6-8, at 8 no improvement would be gained. he also uses custom sampler and custom lora loader for those who has less Vram.
Thanks for all the work on this pack, the chaining is genuinely great.
Found a bug though: guide_audio (Audio Spine) with a real voice track outputs static/hiss instead of the audio, on ref2va. Same audio file works perfectly through the native MiniMaxH3ReferenceToVideo node, so it seems isolated to the Audio Spine injection path
On it, next update, coming tonight.
@joeygambino Thanks for the reply! Quick update: I've now tested up through 2.1.6, same result.
The resample fix from 2.1.3 is confirmed working on my end (console shows the resample happening correctly), but the audio is still coming out as static/garbled, no change from before the fix.
I've ruled out continuity mode (context_pin/latent_handoff), checkpoint format (safetensors and GGUF), scheduler (beta/beta57), and LoRA, same result every time. voice_ref works perfectly on the exact same file, so it's specifically the guide_audio path that's still broken for me.
How are you testing this on your end? Trying to figure out what's different about my setup (RTX 5090, ComfyUI 0.32.0).
Got it — and the answer to "how are you testing this on your end" is the bug. I am on ComfyUI 0.30.0. You are on 0.32.0. It works here and cannot work there, and that is entirely on me for not testing across versions.
What changed. 0.32.0 introduced ModelSamplingAV, and ComfyUI now carries the audio half of the audio+video pack scaled onto the video schedule:
process_latent_in audio slice x (shift / audio_shift)
process_latent_out audio slice x 1 / (shift / audio_shift)
For H3 those shifts are 12 and 3, so the audio latent inside the sampler lives in a 4x-scaled domain. The Audio Spine locks its encoded audio into that pack during sampling — and it was writing raw, unscaled values. Every locked column lands 4x too small, which decodes as exactly the static you are hearing. On 0.30.0 there is no such scaling, so raw was correct.
That accounts for everything you found, and your process is what made it findable:
- voice_ref works on the same file — it goes through the text conditioning and never touches the sampler's latent, so the scaling never applies to it.
- The 2.1.3 resample fix fires and changes nothing — you were right, it works. Encoding was never the problem; the problem is one step later.
- Continuity mode, checkpoint format, scheduler, LoRA all irrelevant — none of them touch this path. Ruling them out is what pointed at the sampler.
Also worth knowing: *audio_lock has the identical bug** on 0.32.0, same code path, same cause. It is fixed by the same change.
The fix reads the scale off the live model_sampling object rather than hardcoding 4, so it stays correct if you change the shifts with MiniMaxH3SigmaShift, and it leaves 0.30.0 behaviour byte-identical. It is written and deployed on my side but I have not render-verified it on 0.32.0 yet - I am setting up a 0.32.0 instance to reproduce your exact failure and confirm the cure rather than ship it on code reading alone. It will be in the next release, which is close.
Until then, honestly, there is no clean workaround. voice_ref will hold one voice across shots and is the nearest thing, but it is not the spine — it does not lock every shot to one continuous performance. If you need the spine specifically, 0.30.0 is the only place it currently works, and I would not recommend downgrading a whole install for one feature when the fix is coming.
Thank you for staying with this through 2.1.6 and for testing so carefully. Four ruled-out variables plus "voice_ref works on the same file" is what turned this from a shrug into a one-line fix.
Oh, and, sorry about "coming tonight" - I got sidetracked trying to add too many features at once, which I tend to do. I am going to say it again though... fix is coming tonight (I hope).
Thanks for the deep dive, really appreciate it! No worries about the delay, I'll wait for the fix and try guide_audio again once it's out
attention (gated) and chunk (gated) nodes dont work for me
Can you paste the errors from the terminal?
@joeygambino [WARNING] invalid prompt: {'type': 'missing_node_type', 'message': "Node 'attention patch (gated)' has no class_type. The workflow may be corrupted or a custom node is missing.", 'details': "Node ID '#9'", 'extra_info': {'node_id': '9', 'class_type': None, 'node_title': 'attention patch (gated)'}}
@joeygambino also in multishot wflow now have error
Prompt outputs failed validation: H3MultishotMemorySampler: - Value 4 bigger than max of 3: memory_frames
Это может быть связано со следующим скриптом:
/extensions/comfyui-easy-use/assets/extensions-WrZZZUnM.js
@egin1992654 Sorry for the late response, I missed you replied.
## 1. The gated nodes: two packs to install
The full workflow uses three nodes from two packs I am not allowed to bundle.
I shipped them switched off, assuming that was enough - it is not.
ComfyUI checks that every node class exists before it will queue anything, even a node that is switched off, so a missing pack stops the whole workflow instead of just that one feature.
Install these two and the workflow runs exactly as shipped:
ComfyUI-sol-attn (provides two of the three)
https://github.com/Saganaki22/ComfyUI-sol-attn
comfyui-minimax-h3-blockcache-T8 (provides the third)
https://github.com/T8mars/comfyui-minimax-h3-blockcache-T8
Via ComfyUI Manager (easiest): Manager > Install via Git URL, paste each URL in turn, then restart ComfyUI.
Or by hand:
cd ComfyUI/custom_nodes
git clone https://github.com/Saganaki22/ComfyUI-sol-attn
git clone https://github.com/T8mars/comfyui-minimax-h3-blockcache-T8
then restart ComfyUI. Check the console on startup - if either pack fails to import it will say so there, and that message is the thing to send me.
Reload the workflow afterwards. The three nodes will resolve, and the error goes away. They are speed and memory optimisations, so you will also get a faster render out of it.
One thing worth knowing: those three ship bypassed on the canvas. Installing the packs stops the error. If you then want the speed as well, select each node and press Ctrl+B to un-bypass it, and turn on the matching switch on the FEATURE SWITCHES panel. Leaving them bypassed is fine too - everything renders identically, just slower.
## 2. The memory_frames error
Value 4 bigger than max of 3 - that dial only accepts 0 to 3.
Open the H3MultishotMemorySampler node and set memory_frames to 0 (that is the shipped default), then queue again.
If other dials on that node also look wrong, the workflow file you loaded was saved by an older release. The sampler gained widgets over several versions, and when a saved file has a different number of values than the node has dials,
ComfyUI fills them in order and everything after the mismatch lands on the wrong dial. In that case load H3_Seamless_Chain_v2.json fresh out of the current zip rather than reusing your saved copy - then re-enter any settings you had changed.
You're definitely getting there, man. This is good stuff.
Your own video above has a cut and isn't seamless though. And the image degradation is pretty severe by the end, like WAN 2.2. Seems like you did fix color shift and audio, and there are no overbright frames at the seams, so this is insane progress for only a few days.
Keep up the great work.
Yeah, I am working on the degradation, expecting to have some progress by morning. The cut.. I don't even know what happened there, it's been pretty steadily working for me.
Hey, first as almost everyone else has already said. Thanks for your amazing work! And also for being so engaged on follow ups!
I managed to get your Riftcast Studio up and running the other day. This flow was looking like a seamless startup for me after I got rid of a Fantasy Talking GGUF node conflict. Then unfortunately the flow ran to about 85% before throwing a math error. I believe I had everything in place as the models auto-populated when I loaded the H3_Seamless_Chain_v2 flow. This was just with the stock images and prompts.
H3 Multishot Sampler + Memory (long form)
Error log
# ComfyUI Error Report ## Error Details - Node ID: 30 - Node Type: H3MultishotMemorySampler - Exception Type: RuntimeError - Exception Message: RuntimeError: mat1 and mat2 shapes cannot be multiplied (3680x1152 and 3456x1152)
I ran the update patch and it came back success. Apologies if this is one of those long since asked and answered. Feel like I had a pretty solid look around for someone with the same issue and cam up empty.
Thanks
@joeygambino Hi, just to add a little more to this as I've been having the same error. My settings in the master control are 640x960 (2:3) I'm only using one image, or at least it's the only one activated and that's in the Identity Anchor Image Node and the resolution is 1024x1536 (2:3) so from a resolution perspective they should be compatible?
The things I have noticed:
1. It only errors if I'm using a GGUF text encoder, safetensors work fine although it doesn't bring the image in as the first image for the shot.
2. After the error if I look at the parameters for the node, which I assume reflect the state of play when the fatal error occurred, width and height are reported as 764x1344 which isn't 2:3. However, if I do get a successful run (using a safetensor text encoder) the finished video is as specified in the master control i.e. 640x960.
I don't know what any of this means, it's mostly all well above my brain cell count, but I thought I'd let you know incase it helps.
Edit: Just to mention after reading wallmonster's latest comments, my error occurs pretty much as soon as it hits the sampler.
@joeygambino Thanks for the quick response!
Should have clarified but as with Lemming below that was with the gguf models. That first run was with only the place holder 768x768 images in there slots. All use image toggles were turned off which should have defaulted to whatever T2I resolution the workflow was saved at and the example text in place. I do not believe I flipped a single toggle on that run. When I was first starting with comfy I threw a lot of math errors with text encoder mismatches but as you said those where always when the first merge happened. Here it goes pretty much all the way until it is getting ready to move off the ksampler. Rough previews were generating that looked to match your example script. Maybe trying to merge the shots?.
Anyway I'll keep digging through my settings. I do have a pretty robust local machine so I'll give the safetensors version a try. My main interest in gguf version is iteration speed as I am learning, as we see here it is sometimes better to fail fast. Local storage space is another big plus for gguf. Like most of your users I am a bit of a hoarder and hesitant to delete anything. Either I have happy memories of one good run with that file or it is on my mental list to go back and figure out how to optimize later.
Edit here: It was not actually at 85%. The ksampler goes to 100% of the first shot and the error throws on the handoff to the second shot. I made sure the image input toggles were all turned off and for extra security bypassed all of their loaders as well as the audio anchor loader. I tried 1152x1152 hoping for a direct match to the model but something, somewhere is adding a little to the image width no matter what I put in the master. I changed the reference image size toggle from match to max also with no joy.
Edit 2: I thought I kept everything 100% unchanged when I first loaded the workflow but I may have been a little too proactive. When trying to figure out where the extra width is coming from I changed the text encoder sidecar setting from referencing the actual file to auto and at least T2V it was able to join 2 shots and run to completion. It's certainly possible that I populated that field myself on the initial run. Testing I2V now with the sidecar on auto.
Edit 3: It runs to completion I2V with the sidecar set to auto.
Early days after only one run but the initial run off of the same reference image and resolution did not seem to generate the same quality as your Riftcast/JoyEcho workflows. Now that I have completed a run I'll move up to the Q8 model and see how that goes.
Thanks again!
Found it - and it is my bug, not your setup. Ignore my earlier answer about resolutions; that was wrong, sorry for the detour.
It's the mmproj_name widget on the H3 CLIP Loader. Naming a file there went down a different code path than (auto) and skipped the key-renaming step, so the vision tower loaded under names nothing reads. That is why it always died at the shot-2 handoff, why it was GGUF-only, and why nothing you changed about resolution or image toggles helped.
Manual fix, in order of least effort:
1. Set mmproj_name back to (auto). If it loads, you are done.
2. If (auto) then says "No vision sidecar resolved", the pairing is by filename - the mmproj must sit in the same folder as the encoder and contain the encoder's name minus its quant suffix:
So either rename the mmproj to match, or make it the only file with "mmproj" in the name in that folder - the loader falls back to "if there is exactly one, use it."
3. If you would rather not touch your model folder, one line in custom_nodes/ComfyUI-H3-Multishot/h3_multishot_utils.py. Find:
if mmproj_name and mmproj_name != "(auto)":
and change it to:
if False and mmproj_name and mmproj_name != "(auto)":
That makes the widget inert and forces the working path. Restart ComfyUI. It is a workaround, not the fix - the real one keeps the widget working for people with split folders.
Or just wait. It is already fixed and verified on my side, and the next release is close — it also carries two other things that stop the workflow running for anyone who installed from here: the accelerator nodes shipped switched on (so a clean install could not queue at all), and a widget mismatch that threw "The value 1 for reference_image_size is not available". If you are not blocked today, the update will be the cleaner path.
Thanks again - @wallmonster151, your Edit 2 is what found this. It would have stayed hidden for a long time otherwise.
@LemmingWolf01 - the 768x1344 you saw on the node after the error is a display quirk, not the cause: width and height are driven by links from MASTER CONTROLS, so the widget keeps showing its own stored default. Your render really was 640x960. Separately, tell me which safetensors encoder you used when the image did not come in as the first frame and I will chase that one too.
Found it - and it is my bug, not your setup. Ignore my earlier answer about resolutions; that was wrong and I am sorry for the detour.
It is the mmproj_name widget on the H3 CLIP Loader. Naming a file there went down a different code path than (auto) and skipped the key-renaming step, so the vision tower loaded under names nothing reads. That is why it always died at the shot-2 handoff, why it was GGUF-only, and why nothing you changed about resolution or image toggles helped.
Manual fix, in order of least effort:
1. Set mmproj_name back to (auto). If it loads, you are done.
2. If (auto) then says "No vision sidecar resolved", the pairing is by filename - the mmproj must sit in the same folder as the encoder and contain the encoder's name minus its quant suffix:
MiniMax-H3-encoder-Q5_K_M.gguf + MiniMax-H3-encoder-mmproj-F16.gguf pairs
MiniMax-H3-encoder-Q5_K_M.gguf + mmproj-F16.gguf does not
So either rename the mmproj to match, or make it the only file with "mmproj" in the name in that folder - the loader falls back to "if there is exactly one, use it."
3. If you would rather not touch your model folder, one line in custom_nodes/ComfyUI-H3-Multishot/h3_multishot_utils.py. Find:
if mmproj_name and mmproj_name != "(auto)":
and change it to:
if False and mmproj_name and mmproj_name != "(auto)":
That makes the widget inert and forces the working path. Restart ComfyUI. It is a workaround, not the fix - the real one keeps the widget working for people with split folders.
Or just wait. It is already fixed and verified on my side, and the next release is close — it also carries two other things that stop the workflow running for anyone who installed from here: the accelerator nodes shipped switched on (so a clean install could not queue at all), and a widget mismatch that threw "The value 1 for reference_image_size is not available". If you are not blocked today, the update will be the cleaner path.
Thanks again - @wallmonster151, your Edit 2 is what found this. It would have stayed hidden for a long time otherwise.
@LemmingWolf01 - the 768x1344 you saw on the node after the error is a display quirk, not the cause: width and height are driven by links from MASTER CONTROLS, so the widget keeps showing its own stored default. Your render really was 640x960. Separately, tell me which safetensors encoder you used when the image did not come in as the first frame and I will chase that one too.
@joeygambino Thanks man! Coincidentally I saw you were updating that file on git when I was digging around. I almost threw in the new utils file to test. I will do that on the next run. In truth I spent a fair amount of time hacking around in that file last night with no joy so I did a full revert to confirm the issue before reaching out.
@joeygambino Hi, thanks for the updates and I can confirm that GGUF text encoders work now. With regard to the first frame issue, the safetensors encoder I was using was just the stock "qwen3vl_32b_minimax_h3_nvfp4_awq" I've since tried with a GGUF + mmproj and still no first frame from my Identify anchor image. Perhaps I'm not fully understanding the process and it's something I'm doing wrong. As it stands I have the:
Reference Gate OFF,
FFLF Plate Gate OFF,
Identity Anchor gate ON (with image),
Continuity = first_frame,
and using a fl2va model (I have tried a ref2va model as well, though in this instance it shouldn't be required, should it?)
As I understand things, that should produce a shot with the first frame as per my Identity Anchor image.
As I said, maybe it's something I'm doing or not doing.
Thanks for all your fantastic work and help.