WAS/Minimax-h3_Singularity_v2_pruned

MiniMax H3 Singularity ref2va v2, pruned

Pruned and quantized versions of the ref2va v2 release of WarmBloodAban/Minimax-h3_Singularity, in the same layout as ComfyUI's own pruned MiniMax H3 releases. Load them in ComfyUI with Load Diffusion Model, with weight_dtype left at default.

File	Size	Built from	Transformer blocks	Token refiner
Minimax-h3_Singularity_ref2va_v2_pruned_int8_convrot_r2.safetensors	21.0 GB	bf16 release	int8 convrot	bf16
Minimax-h3_Singularity_ref2va_v2_pruned_w6a8_r2.safetensors	16.0 GB	bf16 release	w6a8, group 32	int8 convrot
Minimax-h3_Singularity_ref2va_v2_pruned_w4a8.safetensors	11.8 GB	int8 release	w4a8, group 16	int8 convrot

Build r2

The int8 and w6a8 files are built from Minimax-h3_Singularity_ref2va_v2_bf16.safetensors. The first build started from the int8 release. Measured against the bf16 release:

	r2	r1
int8 weight error, sampled layers	0.87% to 1.10%	1.30% to 1.48%
Timestep modulation error	0.02%	0.3% to 0.6%
int8 render, mean PSNR over four seeds	26.1 dB	24.6 dB
w6a8 render, mean PSNR over four seeds	24.4 dB	23.9 dB

The w4a8 file is unchanged. Built from bf16 it rendered further from the bf16 release on every seed tested (20.2 dB against 21.5 dB), so the build from the int8 release stays.

The r1 int8 and w6a8 files remain in the repository history at commit ad7eb66.

What was pruned

The timestep embedder and the full-width adaLN projection in each of the 50 transformer blocks, 13 billion weights in all, are replaced by a 1025 x 8 adaln_t_table and 8-wide adaLN weights: the curve layout of ComfyUI's own minimax_h3_ref2va_pruned_int8_convrot. ComfyUI loads it natively. In the r2 files the table comes from the bf16 release's own timestep embedder, and the modulation it produces is within 0.03% of the bf16 release's at every timestep tested.

The r2 int8 file quantizes the transformer block linears from bf16 with each row's scale set for the lowest error. The w6a8 file quantizes the block linears and final_layer.video_out, with the token refiner in int8 convrot, as ComfyUI's pruned w6a8 release does.

Comparison

1344 x 768, 124 frames at 24 fps, euler, simple, 6 steps, DMAD 4-step LoRA at 0.7, PyTorch attention, minimax_h3_video_vae_int8_convrot and the Qwen3-VL 32B MiniMax H3 text encoder. Mean PSNR of every frame against the bf16 release rendered with the same seed:

File	Seed 42	Seed 43	Seed 44	Seed 45	Mean	Render, seed 42
bf16 release						reference_bf16.mp4
pruned int8 convrot r2	23.3	24.9	28.2	28.0	26.1 dB	pruned_int8_convrot_r2.mp4
pruned w6a8 r2	23.5	24.2	24.3	25.8	24.4 dB	pruned_w6a8_r2.mp4
pruned w4a8	21.9	21.1	21.7	21.4	21.5 dB	pruned_w4a8.mp4

The bf16 release rendered twice with the same settings is identical frame for frame, so every difference in the table comes from the weights. A single seed can rank two close files either way; the mean over four is the comparison.

LoRAs without adaLN or timestep-embedder keys apply unchanged, including the source card's recommended minimax_h3_ref2v_turbo_4step_v0.1.

Modified files

These files are modified versions of WarmBloodAban's Singularity ref2va v2 release (MiniMax H3 Community License Agreement, Section III.2): the int8 and w6a8 files of Minimax-h3_Singularity_ref2va_v2_bf16.safetensors, the w4a8 file of Minimax-h3_Singularity_ref2va_v2_int8.safetensors. The timestep path is replaced as described above and the transformer weights are quantized. Each file's safetensors metadata states the same.

Credits

Fine-tune: WarmBloodAban. Base model: MiniMax H3 by MiniMax.

License

MiniMax H3 Community License Agreement. See LICENSE and NOTICE.