ComfyUI-H3-ContactSheet

Five coordinated views of one subject from one reference image, inside ComfyUI — Contact-Sheet diffusion for MiniMax-H3. Pair with a Turnaround LoRA from matlod/minimax-h3-turnaround.

Two nodes; everything else in the graph is stock:

UNETLoader -> LoraLoaderModelOnly (stock!) -> BasicGuider ---------\
CLIPLoader (type: minimax) ---\                                     v
VAELoader (video) ------------+--> H3ContactSheet ---> SamplerCustomAdvanced
                                     (cond + latent)                |
VAELoader (video) --> H3ContactSheetDecode <--- denoised latent ----/
                        -> 5 views (IMAGE batch) + 1 strip (IMAGE)

• H3 Contact Sheet (five views) — prompt + reference image + per-view size → conditioning (+ the five-slot AV latent). The reference rides as <Picture 1>; keep that tag in your prompt.
• H3 Contact Sheet Decode — decodes the five slots independently (a whole-video decode would smear them) and also emits a [v0..v4] strip.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/matlowai/ComfyUI-H3-ContactSheet
# restart ComfyUI

Models: the MiniMax-H3 ref2va DiT (pruned int8 works), the Qwen3-VL minimax text encoder, the H3 video VAE, and a Turnaround LoRA — the LoRA loads through the stock LoraLoaderModelOnly (key format verified). Sampler settings that match the LoRA's training: res_multistep / simple, 28 steps, denoise 1.0, LoRA strength 1.0 (0.7 favors scene fidelity over rotation).

Sizes 512 / 1024 / 2048 per view all work. Measured through this node on an RTX PRO 6000 (450 W cap, --gpu-only so everything stays in VRAM): 2048 sheet in 175 s (~20% faster than the reference sampler), peak ~54 GiB VRAM; 512 sheets in ~25 s warm. Output verified artifact-free at 2048 (flat-region crops).

Smaller GPUs: those numbers are NOT a requirement. Stock ComfyUI offloads/streams weights to system RAM on smaller cards, exactly as with any big model — if your setup runs MiniMax-H3 video generation at all, it runs this node (a sheet is computationally a ~17-frame clip). Expect slower steps and a hefty system-RAM footprint (~50 GB weights mirror), as usual for H3 on consumer cards.

example_api_workflow.py submits a complete graph over the ComfyUI API if you'd rather script it.

Decode fidelity note (resolved in-node, 2026-08-07)

Early versions decoded each slot as a raw single-latent-frame — a shape that is off the VAE's 17k+5 frame grid and decodes with heavy artifacts in ComfyUI (roundtrip error 31.4 tiled / 93.7 untiled, vs 3.9 on-grid; Comfy-Org/ComfyUI#15416). The node now duplicates each slot token to a legal 2-token clip and keeps pixel frame 0 — output is visually indistinguishable from the reference sampler. Nothing to configure; noted here for anyone building their own per-slot decode.

Honest notes

• Verified equivalent to the research sampler (matched checkpoint / ref / seed side-by-side).
• This rides ComfyUI's internal minimax_refs conditioning interface (comfy_extras/nodes_minimax_h3.py). Tested against master a464ac33 — if sheets degrade after a ComfyUI update, pin near that commit and file an issue here.
• Model behavior (rotation axis follows subject pose, 2048 rotation arrives late, don't stack the LoRA on normal video) is documented on the model card.