pottokao/MiniMax-H3-FastH3-NVFP4-rotated

MiniMax-H3 · FastH3-NVFP4-rotated (single-file, 16 GB)

A rotated-NVFP4 quantization of FastH3 — FastVideo's 4-step, VSA-distilled MiniMax-H3 (video + synchronized audio) — as one self-contained ~12.8 GB file that runs fully resident on a single 16 GB Blackwell card, zero CPU offload, on the native nunchaku W4A4 kernel. This is the fast line (4-step, ~5× fewer steps). For the higher-quality 8-step line see MiniMax-H3-NVFP4-rotated.

Loads with our H3RotNVFP4Loader node — get it at H3-RotNVFP4-ComfyUI-Loader (why a custom node, full method, and dependency versions are there).

Measured (1× RTX 5070 Ti 16 GB, warm, DiT-only)

• Genuine 16 GB residency, zero offload — ~1.2 GB CPU offload (base-class, NOT the 26 GB an un-shrunk AdaLN needs), peak VRAM 13.7 GB.
• 480p (832×480): ~5.7 s/step × 4 → full 124-frame clip in ~25–30 s of denoise.
• Max clip length on a single 16 GB card (dynamic mode): 480p ~12.9 s (310 frames), 768p ~5.17 s (124 frames) — the ceiling is resolution-driven (same for the base line). VSA runs much faster per step but is shorter (480p ~2.8 s / 68 frames; 768p won't fit).
• Only 4 steps. FastH3 is 4-step distilled — vs the 10 a stock-nvfp4 recipe typically needs. Per step it's on par with community quants on the same card (rotation overhead offset by nunchaku's efficient W4A4 kernel), so end-to-end it's **1.9× faster**.
• Audio preserved (clean 32 kHz stereo). File 12.8 GB (T1, all-NVFP4 blocks + rank-32 curve-basis AdaLN + bf16 islands).

Sampler — use Euler (important)

Run with the euler sampler + simple schedule (sigma shift 12), CFG 1.0, 4 steps — the settings FastH3 was distilled for. This is not our quantization: it's a standard sampler choice, and it matters at 4 steps.

• euler → coherent, temporally stable video (recommended default; the included workflow.json uses it).
• res_multistep (a 2nd-order solver) looks a touch sharper but strobes on high-contrast neon / rain / water — the composition churns frame-to-frame. Avoid it here.
• Want more sharpness/motion on the stable Euler base? Raise steps (6–8) — trades some speed. For maximum sharpness, the 8-step base line is crisper still.

Pick sampler/steps per scene and taste — the model is the same either way.

Why "rotated"?

The rotated in the name is a runtime block-256 Hadamard applied before the NVFP4 W4A4 GEMM (QuaRot-style — it spreads outliers so the 4-bit grid isn't dominated by a few large values). A clean same-weights check — identical weights and settings, only rotation toggled — confirms it earns its place: the rotated render keeps crisp edges and sharp wet-pavement reflections, while the un-rotated version hazes over, softening mid-ground detail and neon glyphs (measurably sharper on every no-reference IQA metric, ~+70% on a Laplacian-variance sharpness proxy). It isn't free — the runtime Hadamard costs about 15% per step and gives a different draw for a given seed. Un-rotated NVFP4 isn't broken (fine on easy scenes at enough steps), but where detail counts rotation is the sharper half — so we ship rotated. Full method + measurements: H3-RotNVFP4-ComfyUI-Loader.

Samples

close-up face — eyelashes, iris, skin	forest waterfall — moss, spray
sailboat / sea — rigging, sun-glitter	sci-fi rooftop — flying cars, neon

Full MP4 + synchronized audio in samples/. Rotated-NVFP4 vs fp8 vs bf16 quality comparison (frames + IQA table) in quality/ — same perceptual sharpness as fp8/bf16, but NVFP4 runs the native fp4 kernel (fp8 falls back to bf16 GEMM). (GIFs are video-only previews; use the euler sampler per above for coherent motion.)

Setup

1. Install the loader node + nunchaku — see H3-RotNVFP4-ComfyUI-Loader (verified torch 2.12.1+cu130 + a nunchaku cu13/torch2.12 wheel; Linux + Blackwell RTX 50-series / GB10).
2. Download h3_fasth3_T1.safetensors into ComfyUI/models/diffusion_models/.
3. Load workflow.json: H3RotNVFP4Loader (ckpt_name = the file), text encoder on CPU (CLIPLoader device=cpu), 4 steps. Log marker: [rot] single-file wrapped 200 linears (200 rotation-NVFP4 + 0 bf16-protected).

Credits & license

MiniMax-H3 (MiniMaxAI) · FastVideo (FastH3 4-step + VSA distillation) · nunchaku / SVDQuant (NVFP4 W4A4 kernel) · ComfyUI. Derivative under the MiniMax-H3 Community License — commercial use with attribution; US/UK/EU/South Korea users apply to MiniMax; >$20M/yr needs separate authorization. Verify for your use.