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.
Type a story. Get one continuous video, with sound. Multi-shot scenes render as a single take - no last-frame chaining, no quality loss from shot to shot. That's the whole pitch.
Both workflows now read left to right: numbered lanes, and you only ever touch lanes 2-4. Everything below the main row is optional.
What you need
ComfyUI + this node pack (Manager: MiniMax-H3 Multishot, or the zip on this version).
A MiniMax-H3 checkpoint (links on this page). 24 GB card? Take a GGUF.
SEAMLESS CHAIN - multi-shot scenes as one take
Lane by lane:
README - the quick start lives on the canvas itself.
1 - MODELS - pick your H3 checkpoint and text encoder. VAEs are preset; LoRA slots are empty until you fill one.
2 - ANCHORS (optional) - a photo to open shot 1 on (enable its gate), and a short voice clip to lock the speaker's voice.
3 - YOUR PROMPTS - type your idea in the box, or point the switch at a prompt file. The writer expands it into shot prompts. Writing your own? Set the writer to
passthrough (raw JSON, skip LLM)and paste shots separated by---lines.4 - CONTROLS - size, frames per shot, steps, and
take_seconds(total length; 30 is a good first run). The switches stay off unless you installed the pack a switch names.5 - ENGINE - nothing to change. The remote encoder lives here if you want the text encoder on a second PC: enter its address, flip the encoder switch, free ~15 GB.
6 - OUTPUT - your video and its audio save here.
Optional panels below the main row: reference images (your character, ref2va checkpoints - folder per character + AUTO REFS on), V2V reference (a clip whose look guides the render), FFLF plates (flf_chain mode only), audio spine (a soundtrack the take follows).
EXTEND TAKE - one person talking, as long as you want
Same lanes, different job: one premise becomes ONE continuous speech cut across windows.
1 - MODELS - same as above.
2 - ANCHORS - a photo of your speaker (shot 1 opens on them) and a voice clip. More useful here than anywhere: one person carries the whole take.
3 - YOUR PROMPT - ONE premise, one speaker. The writer writes the whole speech. num_shots 0 = it decides. Passthrough works here too.
4 - CONTROLS -
take_secondsis the star: 30 ships, 60 clears TikTok's minute.windowstays on auto - it sizes itself to your card.5 - ENGINE / 6 - OUTPUT - same as above.
Keep takes to about 4 windows for now - very long takes slowly sharpen.
Rules of thumb (both workflows)
Spoken lines: 8-12 words per shot. Short lines sync; long lines garble.
Say the sounds you want ("rain on the roof, a fridge hum") or it invents its own.
Keep your character's face in frame - faces carry identity between shots.
If something breaks
Red node? Update the pack in Manager, restart, reload the workflow from disk.
Render crawls at low wattage? Lower resolution or frames per shot, or use the remote encoder.
Only one of the two workflows shows in your sidebar? Fixed in 2.6.5 - re-download both.
Still stuck: comment with your console log. I answer.
Deep dives: the two articles linked on this page. Every lane also has a short note on the canvas.
Detailed guide for people that can read good:
Every setting explained: the Seamless Chain deep manual | Civitai
Description
The big one. Everything since v1.4, consolidated - headlined by something no other H3 workflow does: voice identity that survives the chain.
A character's voice now holds across every chained shot
Chained multishot re-performs the voice from text each shot, and it drifts - that has been the known limitation since v1.1. Two new controls on the multishot sampler end it:
voice_ref (AUDIO input): feed a voice clip and it anchors EVERY shot as
<Audio 1>. Stereo, sample rate, and length are guarded internally - a mono clip no longer crashes the packed layout.self_anchor_voice (toggle): no file at all. Shot 1's own rendered voice automatically becomes the reference for every later shot. Write shot 1 so the character speaks a clean solo line, and the scene anchors on itself.
Verified A/B - same seed, same start frame, same continuation line: the plain chain drifted off the character's voice; the anchored render held it and matched the source loudness within 1 dB. Requirements: run the chained shots on a ref2va checkpoint (fl2va was not trained with reference rows) and bind the ref in each shot's prompt: Her voice is the voice in <Audio 1>. The AIO workflow ships with a muted VOICE ANCHOR node ready to unmute.
VRAM management is now automatic
Leave activation_reserve_gb at 0 and the pack sizes the reserve from the actual render shape, measures the true allocator peak of every run, and tightens itself per GPU, model, and shape. The trap where lowering the resolution made renders 5-10x slower (the reserve shrank below the allocator pool and the driver thrashed) is gone. Manual values still pin, and multi-GPU rigs no longer evict a dedicated-GPU text encoder every shot - thanks @VladiCz for measuring that one at a third of his render.
LoRA stack, in every workflow
A four-slot MODEL-only H3 LoRA Stack sits inert between loader and sampler in every bundled workflow. Empty slots are a true pass-through; pick a LoRA in slot 1 (the turbo distill, a style LoRA) instead of rewiring.
New nodes
H3 Ref Folder / H3 Auto Refs / H3 Ref Batch - folder-driven reference casting, up to 9 images, with dialogue stripped before name-matching so a merely mentioned character does not cast.
H3 Cartridge Loader - portable
.riftcastcharacter cartridges (anchor frames + references + voice), drop-in auto-install.H3 Episode Split / Last Frame / Concat A/V - two-stage episode assembly. The concat matches stage B's sharpness and tone to stage A at the seam, and passes stage A through untouched for single-shot renders.
H3 Studio Controls (+ Sampler by Name / Sigmas by Name) - one node whose width/height/frames/steps/sampler/scheduler drive an entire multi-stage graph. No more stage A rendered at one size and stage B at another.
H3 Keyframe Inject (experimental) + the merge patch that lets references and keyframes COMPOSE in one conditioning instead of the refs silently discarding the keyframe.
Worth knowing
With the 4-step turbo distill at 10-12 steps, sample with euler + beta. Multistep solvers extrapolate across step history, and H3's audio runs on a different sigma clock than the video - on a distilled model that math over-steps the AUDIO stream and it comes out distorted.
GGUF loading stays fully automatic (the v1.4.1 fix), and the pack in this zip is the complete current node set - nothing requires a separate download.
FAQ
Comments (34)
Hi, how are you mate? Look I'm just curious on how to load lora stuck works? Thank you.
I am genuinely not sure what you're asking. Can you re-word your question?
My bad I had to update your node pack
Can this be used to make a music video?
I've already tried but its altering the uploaded song unfortunately.
No, not yet anyway. The audio reference in these workflows are only just reference, it won't pass your audio exactly to use in the render. I will look into adding the functionality though.
@joeygambino 大佬要是有空的话，有没有兴趣做一款支持输出ASMR级别音质的视频lora
@joeygambino Thanks.
I found if I type the exact lyrics in my prompt lip-sync works 90% of the time.
Very nice work! But could you add an output to your Multishot node for the LTX 2.3 Spatial Upscaler? See this WF: https://civitai.red/models/2842324/minimax-h3-base-gen-ltx-23-spatial-upscale?modelVersionId=3208714
You can add this yourself... You're pointing them to the thing but aren't doing it yourself? Just add the node. It's like 2 seconds of work and you're asking them to do it for you?
@DaddyWolfgang You are right. I thought it depended on a missing latent sampling output, but I was wrong. It works. Thanks.
@ingriduterus513 New workflow up with built in upscaler and a bunch of other new features.
This is friggin awesome :)
Thanks!
If anyone finds out how to prevent the clip to clip degradation, let me know :)
I've been working on it all day! Update soon ;)
@joeygambino started using H3 Motion Context. It's unreal
@Unisol58930 I did not even know that existed, I've been working on integrating LTX architecture into H3 myself for this same reason. I will scour their Git. Thanks!
I disabled SageAttention and the blurriness disappeared. I made a second batch file without it. No need to thank me))
@bionovafood863 Ha, if that ends up being the whole problem I will hang my head in shame. I will test it though!
@bionovafood863 And there is no texture and lighting changes between segments? I've tried with and without SageAttention, made no difference for that
@Unisol58930 I just found the H3 motion repo today do you have any reccomended settings when you chain videos. Was wondering if it was good glad to know it works.
@Hzoid I use SamplerCustomAdvanced and from that I send the denoised_output to H3 Motion Context Save Latent. Then on the next segment H3 Motion Context Load Latent -> H3 Motion Context as the context_latent.
H3 Motion Context conditioning come from MiniMax H3 Reference to Video in my case. Then just send the H3 Motion Context conditioning to your Basic guider.
Keep in mind that this will keep the context_length X frames in your generation. For example if set to 22, 22 first frames of the new segment will be from the old latent. Just cut it off based on frames or time from the video before combining the clips. 22 frames is 0.917 seconds in 24fps
@Unisol58930 Thanks for the info exciting times
@Unisol58930 @bionovafood863 @Hzoid
New workflow up that should take care of most of these problems and add some new features too. The demo video is three 10 second shots.
What i discoverd is: lets say walking pace, first video > slow walking, second video > fast walking, third video > slow walking again. every new block do not know how to extend the movement from the previous block. thats the same problem that Wan I2V has when you want to extend it, with only one picture, the model know nothing about the movement. Whats about to use a frame sequence for a motion anchor like vace-tools or wan-vace-prep do to extend videos?
What I would suggest to add to all 3 workflows:
- Possibility to edit single wrong rendered prompt segments.
- Add Music (Music Video mode)
- Multiple image reference inputs with multiple voice reference inputs (Multi person mode)
I will see what I can do, I have a running list!
how do you toggle "self_anchor_voice"? I don't see it anywhere in the workflow.
It's in the AIO and MEMORY workflow, but I just noticed I hadn't actually updated the keyframes workflow with it. I will update the workflow, sorry!
I'm actually using the AIO workflow, I can't find it in there. I see the node for uploading an audio clip to use as a reference, but I can't find the toggle you refer to.
Nevermind! I found it. Don't know how I missed it before lol.
Do you have a good way to prompt for music? When I try to do it with "non_diagetic_music:", it just repeats the same snippet of the song during each segment, like identically to the first segment.