CoolFace
Modelpublic

rzgar/minimax_h3_fl2va_fp8_e4m3fn

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face
23likes2kdownloads
Model Card

MiniMax H3 MXFP8

MXFP8 quantization of the MiniMax H3 fl2va (first-last-frame to video+audio) diffusion checkpoint.

better audio at 5 (talking heads, less motions) & 8-steps (shift 8)

<video controls width="1280" height="720"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vafp8e4m3fn/resolve/main/demo/mxfp8vsfp8/mxfp8vsfp8.mp4" type="video/mp4"> Your browser does not support the video tag. </video>

MiniMax H3 FP16attn

Balanced quality and performance

fp16attn 8-steps | shift: 8.0 <video controls width="1280" height="720"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vafp8e4m3fn/resolve/main/demo/fp16attn/fp16-fp8_demo.mp4" type="video/mp4"> Your browser does not support the video tag. </video>

MiniMax H3 FP8 (E4M3FN) Quantized

Based on testing (this model and pruned_fp8), the minimal steps to get results similar to Lightx2v using 4-step LoRAs is 5step & 5steps. The sweet spot for quality and speed is 8step and 8steps. This quantization requires a minimum of 8step.

  • —Sampler: dpmpp_2m
  • —Scheduler: sgmuniform <img src="https://huggingface.co/rzgar/minimaxh3fl2vafp8_e4m3fn/resolve/main/demo/minimax-5step-ksampler.png" width="1196" height="709" />
5-Step pruned fp85-Step mxfp88-Step fp8
<video controls width="50%" height="50%"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vafp8e4m3fn/resolve/main/demo/MiniMaxH35stepprunedfp8.mp4" type="video/mp4"></video><video controls width="50%" height="50%"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vafp8e4m3fn/resolve/main/demo/mxfp8vsfp8/MiniMaxH3mxfp85step5steps_.mp4" type="video/mp4"></video><video controls width="50%" height="50%"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vafp8e4m3fn/resolve/main/demo/MiniMaxH38stepfp8.mp4" type="video/mp4"></video>
8-steps fp16attn
<video controls width="50%" height="50%"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vafp8e4m3fn/resolve/main/demo/fp16attn/MiniMaxH3fp16-fp88steps.mp4" type="video/mp4"></video>

FP8 quantization of the MiniMax H3 fl2va (first-last-frame to video+audio) diffusion checkpoint.

<video controls width="1280" height="720"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vafp8e4m3fn/resolve/main/demo/demo.mp4" type="video/mp4"> Your browser does not support the video tag. </video>

Original model: MiniMaxAI/MiniMax-H3

File sizes

VariantDownload
minimaxh3fl2va_mxfp8.safetensorsDownload 47.6GB
minimaxh3fl2vafp16attnfp8.safetensorsDownload 28.7GB
minimaxh3fl2vafp8e4m3fn.safetensorsDownload 47GB
minimaxh3fl2vaprunedfp8_scaled.safetensorsDownload 21GB

Simple Upscalers

Old but gold upscale models. Lower image resolution for faster inference and more steps.

VariantDownload
2xNomosUniesrganmultijpg.pthDownload
RealESRGAN_x2plus.pthDownload

736x416 | ~56 sec | 1472x832 60fps <video controls width="1280" height="720"> <source src="https://huggingface.co/rzgar/minimaxh3fl2vafp8e4m3fn/resolve/main/upscalemodels/demo/2xNomosUni60_fps.mp4" type="video/mp4"> Your browser does not support the video tag. </video>

Left at native precision (quality-critical)

TensorsDtypeNote
adaln_proj.linear.* (100 tensors)BF16Per-block modulation, largest single weight
token_refiner.* (17 tensors)BF16Text embedding refinement, 2 layers only
condition_proj.* (2 tensors)BF16Cross-modal conditioning bridge
All norm weights (norm1, norm2, q_norm, k_norm, final_norm)BF16104 tensors, tiny and precision-sensitive
All biasesBF1656 tensors
time_embedder.* (4 tensors)FP32Timestep signal, original precision preserved
video_patch_proj.*, audio_patch_proj.* (4 tensors)FP32Input/output projections
final_layer.*_out.* (4 tensors)FP32Decoder output heads
rope.inv_freq (1 tensor)FP32Position encoding

ComfyUI-MiniMaxH3-Text-Enhancer node

What this is: the result of looking into why male anatomy comes out misshaped in text-to-video scenes or when a nude male appears in the I2V scene. Along the way we learned some things worth knowing about how this model really works, read README.md and notes.md in case of you are interested.

Q: Is it a text-encoder problem? Would an abliterated Qwen3-VL solve it?

  • —A: No. The stock encoder doesn't refuse NSFW prompts, the tokenizer and embedding table handle them normally. Its only strong divergence from stock is on harmful-instruction content (~11% embedding shift); sexual content shifts only ~1–3%. Abliteration targets the wrong thing here.

Q: Is the base model censored? Would a domain-specific LoRA be ineffective?

  • —A: No, the base is uncensored. The misshape comes from weak text conditioning: only ~3–5% of the embedding actually carries the prompt, so the DiT's generic body prior wins over the anatomy request. Modifying internal values (condition_proj ×1.2 + 12,658 FFN columns ×1.2) improves the result slightly and a properly trained LoRA would refine anatomy further on top of it.

Experimental: 1.2 strength in both nodes improves NSFW rhythmic animations in some scenarios <img src="https://huggingface.co/rzgar/minimaxh3fl2vafp8e4m3fn/resolve/main/ComfyUI-MiniMaxH3-Text-Enhancer/txtenhancerexprmntl.png" width="1280" height="720" />