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

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

These are the pruned MiniMax-H3 FL2VA and Ref2VA transformers from Comfy-Org/MiniMax-H3 with the full-rank PDMD 4-NFE 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_pdmd_4step_int8_convrot.safetensors	FL2VA	minimax_h3_fl2va_pruned_int8_convrot.safetensors
minimax_h3_ref2va_pruned_pdmd_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: PDMD was trained on the text-to-audio-video transformer, and the Ref2VA weights differ from it. The PDMD authors don't publish results for the Ref2VA partition.

Examples

FL2VA

Steps	CFG	Size	Frames	Seed	Sampler	Shift (video / audio)
4	1.0	1344×768	124 at 24 fps	42	PDMD	12 / 3

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	43	PDMD	12 / 3

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. This render used the same setup as the FL2VA example.

Sampling

Use the PDMD student's own operating point: 4 steps, CFG 1 (no classifier-free guidance) and the base model's released scheduler configuration: video time shift 12 and audio shift 3.

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 (FL2VA: revision d6cfb4e5; Ref2VA: revision 912b6e38).
   • LoRA: lora_model_0.safetensors from pdmd2026/pdmd_4NFE_lora (revision 9913253b), with no rank reduction: rank 128, alpha 128, in diffusers layout.
   • Layout reference: diffusion_models/minimax_h3_<variant>_pruned_int8_convrot.safetensors from Comfy-Org/MiniMax-H3 (FL2VA: revision d6cfb4e5; Ref2VA: revision 912b6e38).
2. Convert to the Comfy-Org layout: the 312 diffusers adapters become 208. to_q, to_k and to_v are fused into one attn.qkv_proj adapter of rank 384: the down projections are stacked and the up projections are placed block-diagonally, so each output band only sees its own adapter. ff.net.0.proj becomes mlp.fc1 with the two halves of its up projection swapped, because the Comfy-Org SwiGLU weight stores the gate half first. attn.to_out.0 and ff.net.2 map to attn.out_proj and mlp.fc2 unchanged. Alpha stays equal to the rank.
3. 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.
4. 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.
5. 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 0.9967 for FL2VA and 0.9966 for Ref2VA. All 200 block layers are within 1 ± 0.1. The 8 token_refiner layers fall between 0.48 and 0.77: their PDMD delta is only about 0.1% of the weight, which is below BF16 rounding, so part of it rounds away when they are stored in their Comfy-Org dtype.

Credits & license

• Base model: MiniMaxAI/MiniMax-H3, repackaged by Comfy-Org/MiniMax-H3
• Distillation: pdmd2026/pdmd_4NFE_lora (paper, project page)

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

Powered by MiniMax H3.