MiniMax-H3 Multishot Workflow

Chain multiple MiniMax-H3 shots into one continuous video with audio - in

one node. Each shot starts from the last frame of the previous one; the

duplicated seam frame and its 1/24s of audio are trimmed automatically.

The demo below was made by this workflow: 30 seconds, three chained shots

from one script, same presenter and same voice across both seams, rendered on

the Q5_1 GGUF.

Update your node pack before using H3_Multishot_MEMORY.json

H3MultishotMemorySampler called vae_decode_audio without importing it, so it raised a NameError at audio decode — after all sampling had finished, destroying the completed render. It was broken in every released version until now.

Fixed in ComfyUI-H3-Multishot (issue #1). Pull the latest node pack and restart ComfyUI fully — a browser refresh is not enough. The workflow JSON itself never needed changing.

Files

H3_Multishot_AIO.json - easy mode: loaders > script box > one sampler node > save. Write one prompt per shot with --- between them, pick a shot count (0 = one shot per prompt, 1-8 forces it), queue.
H3_Keyframes.json - keyframes at any position, not just first/last, in a single generation - so the audio is one continuous stream. Six anchor slots plus an unbounded images_batch, placed by percentage (0%, 50%, 100%), absolute frame index (0, 121, 242) or inclusive range (0-9, 352-361). A descending range like 30%-20% reverses that section of the batch.
H3_Multishot_MEMORY.json - long form (2-5 minutes). Keeps a persistent identity anchor so drift cannot compound across many shots.

Requirements

ComfyUI v0.30.0+ (native MiniMax H3 support)
The node pack: ComfyUI-H3-Multishot (Manager > Install via Git URL;

includes the one-line ComfyUI-GGUF architecture patch)
Models: MiniMax-H3 GGUF quants (Q5_1 for 24-32 GB cards, Q4_0

for 16 GB) or the originals; text encoder + VAEs from

Comfy-Org/MiniMax-H3

Notes

frames_per_shot sits on H3's 17k+5 frame grid (243 = ~10.1s at 24 fps;

362 = ~15.1s, the trained max).
End every shot on what the NEXT shot expects to see - the chain hands each

shot the previous final frame, and matching that bridge to the next shot's

framing is what makes seams invisible.
Malformed JSON scripts fail loudly instead of rendering the raw text.

v1.3 (2026-08-05)

Percentages and ranges in keyframe positions, contributed by @viralesveras. A bare 1 used to mean the last frame rather than frame 1, so addressing an early frame absolutely needed 1.0001 - which is how the ambiguity was found, in real use. Existing workflows keep working: a saved 0, 0.5, 1 is still read the old way and logs the percentage spelling to switch to, while an all-integer 0, 1 takes the new absolute meaning and warns rather than silently moving an anchor.
images_batch on the keyframe node, also from @viralesveras - for when six slots is not enough. It adds to the six slots rather than replacing them.
Loaders find GGUFs in subfolders - a model under diffusion_models/gguf/ was invisible in the dropdown because the scan was not recursive.
New: curve-form GGUFs, ~40% smaller at the same quant tier, with a Q8_0 that fits a 24 GB card: joeygambino/MiniMax-H3-curve-GGUF. They need ComfyUI 0.30.0+; the original-form files stay maintained.

v1.2 (2026-08-04)

Keyframes at any position. Stock ComfyUI pins H3 keyframes to the first and last frame and raises only first/last keyframe anchors are supported for anything else. That is a positional-maths limit, not a model limit - both stock cases are the same expression:
```
cond_t = text_len + FRAME_RESCALE * pixel_index
```
which is defined for every frame. Measured on an RTX 5090, 243 frames, one anchor at pixel frame 121: the rendered frame closest to the anchor image was frame 122 - off by one - reached by continuous motion with no cut, audio unbroken through it. A three-anchor run landed frames 121 and 242 exactly.

Applied in memory; it does not edit any ComfyUI file, self-tests against the stock formula, and rolls back if first/last positions do not reproduce.
Move vs cut. Anchor images with a plausible camera path between them make H3 interpolate. Images with no possible path (a kitchen and a diner) make it cut, then hold - stock first/last does the same with such a pair, so that is the model, not the node. The cut case is still useful: a timed shot change inside ONE generation, which keeps the audio continuous across it.
Two new controls: H3 Condition Strength (exposes minimax_visual_cond_noise_aug / minimax_audio_cond_noise_aug, read by ComfyUI core but written by no stock node) and H3 Reference Audio, a stereo guard - a mono reference clip crashes the sampler with an unhelpful shape mismatch because the layout reserves two channels.

v1.1 (2026-08-04)

Image-to-video. The Multishot Sampler now takes an optional start_image. Connect a LoadImage and shot 1 starts from that frame, then chains as usual. Leave it unconnected for unchanged text-to-video behaviour - v1.0 graphs keep working.
~4x faster on 32GB cards. The text encoder is evicted before sampling. The Qwen3-VL encoder (~~16.5GB even at Q4) and the H3 DiT (~~25GB) do not co-fit on 32GB, so the DiT was loading partially and streaming 19GB from system RAM every step. Measured on an RTX 5090: **60 min -> ~15 min** for the same render.
Text encoder GGUF + mmproj published: joeygambino/MiniMax-H3-encoder-GGUF. The mmproj sidecar is required for multi-shot, not just for reference images - shot chaining feeds the previous frame through the encoder's vision path.
Render at H3's native resolution and upscale afterwards: native 1920x1088 measured worse than 960x544 in blind review and cost ~4x the time.

Node pack: ComfyUI-H3-Multishot

On Civitai

Support

Everything here is free and stays free. If it saved you a night of debugging, tips keep the 5090 warm:

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for joeygambino/MiniMax-H3-Multishot-Workflow

Base model

MiniMaxAI/MiniMax-H3

Finetuned

(23)

this model