Spectrallic/MiniMax-H3-DMAD-4step-int8-convrot

MiniMax-H3 FL2VA and Ref2VA (pruned) 4-step DMAD: INT8 ConvRot

These are the pruned MiniMax-H3 FL2VA and Ref2VA transformers from Comfy-Org/MiniMax-H3 with the full-rank DMAD full_critic 4-step LoRA merged in, quantized to INT8 ConvRot. Both were made the same way and differ only in the base model:

File	Base model	Same layout as (Comfy-Org diffusion_models/)
minimax_h3_fl2va_pruned_dmad_4step_int8_convrot.safetensors	FL2VA	minimax_h3_fl2va_pruned_int8_convrot.safetensors
minimax_h3_ref2va_pruned_dmad_4step_int8_convrot.safetensors	Ref2VA	minimax_h3_ref2va_pruned_int8_convrot.safetensors

The tensor names and order, dtypes, shapes, data offsets and .comfy_quant descriptors all match the Comfy-Org file. The difference is that the distillation is baked in, so you don't need a LoRA loader.

Modified files: the weights differ from the MiniMax-H3 release, as described under How they were made.

Layout: every block linear (blocks.N.attn.{qkv_proj,out_proj}, blocks.N.mlp.{fc1,fc2}) is INT8 and has its own .comfy_quant tensor ({"format": "int8_tensorwise", "convrot": true, "convrot_groupsize": 256}) and a per-row F32 weight_scale. Everything else keeps the Comfy-Org dtype.

Not trained on Ref2VA: DMAD was trained on the text-to-audio-video transformer, and the Ref2VA weights differ from it. The DMAD authors state that the LoRA loads on the Ref2VA partition but don't publish results for it.

Examples

FL2VA

Steps	CFG	Size	Frames	Seed	Sampler	Shift (video / audio)
4	1.0	1344×768	124 at 24 fps	42	DMAD	12 / 2

A cinematic view of an ancient floating city suspended above a sea of cumulus clouds. The camera begins in a wide establishing shot and slowly dollies in toward the city, gliding forward through the soft pastel cloudscape as its intricate, hand-drawn architecture comes into view. A waterfall cascades from the city's edge into the clouds, while shafts of golden sunlight break through the cumulus clouds and illuminate the ultra-detailed digital painting.

This render used SageAttention with sparse attention at density 0.15 (first step dense) as described here.

Ref2VA

Steps	CFG	Size	Frames	Seed	Sampler	Shift (video / audio)
4	1.0	1344×768	124 at 24 fps	42	DMAD	12 / 2

A close-up of a small wooden fishing boat in a stormy green sea at night, rain lashing the deck, while the floating city from <Picture 1> hangs in the sky far above and lightning flashes behind it. Handheld camera. Loud thunder and crashing waves.

<Picture 1> is the first frame of the FL2VA example above, scaled to match the output size. This render used the same setup as the FL2VA example: SageAttention with sparse attention at density 0.15 (first step dense).

Sampling

Use the DMAD student's own operating point: 4 steps, CFG 1 (no classifier-free guidance), video time shift 12 and audio shift 2. Workflows are in the DMAD repository.

Both examples were rendered in Nox with the DMAD sampler. For Ref2VA, refer to references in the prompt as <Picture i>, <Video k> and <Audio j>.

How they were made

<variant> below is fl2va or ref2va.

1. Source:
   • Base: diffusion_models/minimax_h3_<variant>_pruned_bf16.safetensors from Comfy-Org/MiniMax-H3 (Ref2VA: revision 912b6e38).
   • LoRA: minimax_h3/dmad_minimax_h3_4step_full_critic_comfyui.safetensors from ZhengmingYu/DMAD (revision b08c2bf3). This is the authors' exact conversion of full_critic, with no rank reduction: rank 128, with q/k/v fused into attn.qkv_proj adapters of rank 384.
   • Layout reference: diffusion_models/minimax_h3_<variant>_pruned_int8_convrot.safetensors from Comfy-Org/MiniMax-H3 (FL2VA: revision d6cfb4e5; Ref2VA: revision 912b6e38).
2. Merge: the LoRA is merged at strength 1.0 into its 208 targets: attn.{qkv_proj,out_proj} and mlp.{fc1,fc2} in all 50 blocks and both token_refiner blocks. Each merge computes W' = W + (alpha/rank) · up @ down in FP32.
3. Quantize: the 200 block linears are rotated with a group-256 Hadamard ConvRot, then quantized to INT8 with per-row scales, directly from the FP32 merge result. The 8 merged token_refiner weights go back to BF16, because that layer stays unquantized in the Comfy-Org file.
4. Copy everything else: tensors the LoRA doesn't touch are copied bit for bit from the layout reference.

Checks

• Against the Comfy-Org INT8 file, byte for byte: all 932 tensors match in name, order, dtype, shape and offset. Only 408 tensors have different bytes: the 200 INT8 weights, their 200 scales and the 8 token_refiner weights. The other 524 are identical, including all 200 .comfy_quant tensors.
• LoRA presence: this projects the change from the base onto the LoRA delta, where 1 means fully present. The overall result is 0.9977. Of the 208 layers, 205 are within 1 ± 0.1. The other three are token_refiner.blocks.0.mlp.fc2 (0.647), token_refiner.blocks.1.attn.qkv_proj (0.896) and token_refiner.blocks.1.attn.out_proj (0.899).

Credits & license

• Base model: MiniMaxAI/MiniMax-H3, repackaged by Comfy-Org/MiniMax-H3
• Distillation: ZhengmingYu/DMAD (paper, code)

This derivative is released under the MiniMax H3 Community License Agreement, which includes territorial restrictions. See LICENSE and NOTICE.

Powered by MiniMax H3.