MiniMax H3 Fused Turbo: One 21GB File, 4-Step Video and Audio

The 4-step single-pass reference run: 1152x640, 243 frames (10 s with audio), 76 seconds on one RTX PRO 6000.

What is inside the file

The bake starts from the pruned fl2va transformer with a rank-1024 SVD of the (ref2va - fl2va) weight delta fused in, so one partition serves both first/last-frame conditioning and reference conditioning. On top of that:

A single INT8 ConvRot quantization pass runs after the merge: the four heavy Linear weights in each of the 50 blocks (qkv_proj, out_proj, fc1, fc2) are INT8 with ConvRot (group 256, per-channel) in ComfyUI's native comfy_quant layout, everything else stays BF16/F32. That order matters: quantizing the merged weights avoids quantized-base-with-float-deltas drift, and same-seed tests measured the bake identical to the live-LoRA path.

Why merge instead of stacking LoRAs

The two LoRA files total about 540 MB, so adapters are not the problem. A LoRA is a low-rank delta that ComfyUI adds at load time; to stay removable it keeps a pristine copy of every weight it touches. Patch all 200 core layers of an INT8 model and that backup is a second full model in memory. Measured twice at 8 steps on the same seed:

Same take, same wall time, audio meters identical within noise. The bake simply removes two loaders, one quantization round-trip and the ~21 GB patch backup. If you need adjustable LoRA strengths, the README documents the equivalent live path with public files.

Steps: it says 8, run it at 4

The merged turbo is lightx2v's 8-step LoRA, but it is a 4-or-8 NFE model. The measured ladder on one RTX PRO 6000 (96 GB), 1152x640, 243 frames, SLA sparse attention:

The same take at 25 steps: a little sharper drawing, 292 seconds instead of 76, soundtrack unchanged.

The de-rope is the reason to run 4 steps: pass 2 only spends steps on the frames the jerk oracle flagged, so 4 + 4 together cost about the same as one 25-step pass while the held regions come out cleaner. Two recipe findings matter: res_multistep beats euler for audio quality at 4 steps (lightx2v's example graphs say euler), and SLA block-sparse attention at 0.90 sparsity is audible, restoring top-end soundtrack detail dense attention flattens while running about 40% faster.

Workflows

The repo ships five ComfyUI graphs (UI and API format), each with one PROMPT box feeding every stage. Requires ComfyUI-MAINodes and an SLA sparse-attention node pack:

Getting started

1. Download minimax_h3_fused_refdelta_r1024_turbo8_mystic07_int8_convrot.safetensors (21 GB) into ComfyUI/models/diffusion_models/ and load it with a stock UNETLoader, weight_dtype default. No custom quantization loader needed.
2. Download the companion files from Comfy-Org/MiniMax-H3: minimax_h3_video_vae_int8_convrot.safetensors, minimax_h3_audio_vae_fp32.safetensors (VAEs) and qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (text encoder).
3. Install ComfyUI-MAINodes and an SLA sparse-attention node pack, then drag in a workflow above.

Smaller cards work too: ComfyUI's model management loads and offloads the transformer, text encoder and VAEs as needed, so the same graphs run, just slower. On AMD/ROCm, the recipe was developed on an AMD box with patientx's ComfyUI ROCm fork, which also carries a ROCm build of the SLA attention node.

Links

• Hugging Face: MATLOWAI/minimax-h3-fused-turbo-int8-convrot
• GitHub: matlowai/ComfyUI-MAINodes