MiniMax H3 ref2va — NVFP4 (ComfyUI-native)
🚀 Use this model hosted (API + playground): modelslab.com/models/minimax/minimax-hailuo03-reference-to-video
NVFP4 quantization of the MiniMax H3
reference-to-video DiT (ref2va), in ComfyUI's native quant layout — loads with
the stock UNETLoader on any Blackwell GPU (sm_120: RTX 5090 / RTX PRO 6000, and
newer). Derived from
Comfy-Org/MiniMax-H3's
minimax_h3_ref2va_bf16.safetensors.
Files
| File | Size | GEMM path |
|---|---|---|
diffusion_models/minimax_h3_ref2va_nvfp4.safetensors |
38.6 GB | native FP4 tensor-core |
diffusion_models/minimax_h3_ref2va_nvfp4_fpmm.safetensors |
38.6 GB | dequant → bf16 (quality-safe, same path the official NVFP4 text encoder uses) |
Quantization policy mirrors the official int8_convrot release: the 50 main
blocks' attn.qkv_proj, attn.out_proj, mlp.fc1, mlp.fc2 (200 layers) go
to NVFP4 (E2M1, per-16 FP8-E4M3 block scales + global FP32 scale); everything
quality-critical — adaln/modulation, norms, patch/condition projections, token
refiner, final layers — stays bf16. Weight-only PTQ (no activation calibration),
produced with the included convert_nvfp4.py via ComfyUI's comfy_kitchen
quantize_nvfp4 kernels.
Usage (ComfyUI)
Drop into ComfyUI/models/diffusion_models/ and select in UNETLoader
(weight_dtype default) inside the official
MiniMax H3 r2v template.
Pair with the official qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors text
encoder and both H3 VAEs from
Comfy-Org/MiniMax-H3.
Benchmarks (RTX PRO 6000 Blackwell Max-Q, 96 GB, SageAttention, 1344×768×124f, 20 steps)
All on the same card, same seed, SageAttention. torch cu130 build strongly recommended — it enables comfy_kitchen's optimized CUDA kernels (on cu128 the FP4 path runs emulated and is ~2.3× slower).
| DiT variant | Size | s/step (cu130) | s/step (cu128) | Notes |
|---|---|---|---|---|
| nvfp4 (this repo) | 38.6 GB | 8.47 | 19.8 (emulated) | full model, 1.5× faster than bf16 |
| convrot W4A4 (measured, not shipped) | 37.4 GB | 9.69 | — | visible text-rendering artifacts |
| bf16 (reference) | 66.3 GB | 12.76 | 14.65 | baseline quality |
| nvfp4_fpmm (this repo) | 38.6 GB | — | 14.99 | ties bf16 on cu128, 42 % less VRAM |
| pruned int8_convrot (official) | 21.0 GB | — | 22.7 | pruned arch, W8A16 dequant path |
Stacking ComfyUI's EasyCache (threshold 0.2, quality-gated) cuts wall time a
further ~1.4×.
End-to-end speed (full optimized stack)
600W RTX PRO 6000 Blackwell, nvfp4 + cu130 + SageAttention +
fp16_accumulation + fused adaln/gate kernels + EasyCache 0.2, 20 steps,
1344×768 @ 24 fps with native stereo audio:
| Video length | Wall time | Notes |
|---|---|---|
| 15 s (362 frames — the model's one-shot ceiling) | 8 min 07 s | vs |
| 5 s (124 frames) | ~2 min (hot server) | measured through a production API |
| 15 s draft (16 steps @ 960×544) | ~3.5–4 min | preview tier |
Important: H3's trained range tops out at 362 frames (~15 s) — longer one-shot generations collapse regardless of settings (verified empirically at 736 frames with multiple schedules/shifts).
Same-seed visual quality of nvfp4 vs bf16: no quantization artifacts observed
(trajectory divergence only — diffusion is chaotic under any weight
perturbation; see sample). The W4A4 experiment degraded fine text rendering,
which is why it is not shipped.
A same-seed sample generated with this checkpoint is in
assets/sample_r2v_5s.mp4.
License
MiniMax H3 Community License. This is a derivative work of MiniMaxAI/MiniMax-H3; see the license for terms.
Model tree for ModelsLab/MiniMax-H3-ref2va-NVFP4
Base model
MiniMaxAI/MiniMax-H3