MiniMax H3 Ref2VA PDD-Acc Baked INT8 ConvRot

Quality-oriented INT8 ConvRot build of MiniMax H3 Ref2VA with the official MiniMax-H3-Ref2VA-Acc-8Step PDD-Acc trunk baked into the model.

The PDD acceleration head bank is not baked into the main checkpoint. It remains a runtime component and must be loaded with ComfyUI-MiniMax-H3-PDD-Acc.

Model

Filename: minimax_h3_ref2va_pdd_baked_int8_convrot.safetensors
Architecture: MiniMax H3 Ref2VA
Base: official pruned MiniMax H3 Ref2VA INT8 ConvRot checkpoint
PDD source: alibaba-pai/MiniMax-H3-Acc-LoRAs
PDD source file: MiniMax-H3-Ref2VA-Acc-8Step.safetensors
PDD bake strength: 1.0
Model size: 19.53 GiB
Total tensors: 932

Quantization and precision

The model was built with the priority of preserving video-generation quality rather than aggressively quantizing every component.

Main Transformer

50 Transformer blocks × 4 target modules per block:

• attn.out_proj
• attn.qkv_proj
• mlp.fc1
• mlp.fc2

Total: 200 INT8 ConvRot-256 modules.

Every quantized module keeps:

• weight
• weight_scale
• comfy_quant

comfy_quant:

format: int8_tensorwise
convrot: true
convrot_groupsize: 256

Preserved higher precision

• Token Refiner parameters are kept in higher precision.
• AdaLN projection parameters are FP16.
• adaln_t_table is FP32.

Final audited dtype inventory:

torch.bfloat16 : 220 tensors
torch.float16  : 102 tensors
torch.float32  : 210 tensors
torch.int8     : 200 tensors
torch.uint8    : 200 tensors

Total: 932 tensors.

PDD bake method

The official PDD-Acc checkpoint is not treated as a plain LoRA. The rank-64 PDD trunk is baked with the PDD-aware bake pipeline.

For INT8 ConvRot weights the merge path is conceptually:

INT8 ConvRot
    ↓
dequantize
    ↓
high-precision PDD trunk merge
    ↓
INT8 ConvRot-256 re-quantization

The PDD final-layer head bank remains separate and is applied at runtime.

Runtime

Pair this checkpoint with the official Ref2VA PDD-Acc file through MiniMaxH3PDDAccApply.

Because the trunk is already baked:

nfe           = 8
lora_strength = 0.0
head_strength = 1.0

The runtime PDD head bank is still required.

Recommended sampling recipe

NFE / sampler steps : 8
Sampler              : Euler
CFG                  : 1.0
Video Sigma Shift    : 12.0
Audio Sigma Shift    : 3.0
Denoise              : 1.0

Use the sigma output from the PDD Apply node in the advanced sampler path so the trained PDD block-boundary sigmas are preserved.

Do not add other distillation LoRAs (for example turbo/lightx2v) on top of the PDD recipe.

Quality-oriented design

The release deliberately keeps quality-sensitive components in higher precision instead of forcing a full-model INT8 conversion.

high visual quality
      +
INT8 VRAM / storage efficiency
      +
PDD 8-step acceleration

The objective is not minimum file size. It is a practical production checkpoint that keeps H3 visual behavior as closely as possible while gaining INT8 ConvRot efficiency and PDD acceleration.

Verification

Structural audit

932 / 932 tensors present
200 / 200 expected main Transformer ConvRot modules present
ConvRot triple errors: 0
NaN / Inf scales: 0
Non-positive scales: 0

Every quantized main-module triple was verified as:

weight         = INT8
weight_scale   = FP32
comfy_quant    = UINT8

AdaLN

50 blocks × (weight + bias) = 100 tensors
dtype = FP16

adaln_t_table:

dtype = FP32
shape = [1025, 8]

PDD bake verification

The bake pipeline performed a post-write comparison against the exact high-precision merge. The audit sampled 8 / 258 baked modules.

Worst relative error:

7.378e-03

approximately 0.738%.

This is the measured INT8 re-quantization error relative to the exact merge for the sampled module with the largest deviation. It is not a claim that visual video quality decreases by 0.738%.

File integrity

Model SHA256:

03b5e498fa3ffd2d5f1befbb9fc7b1ed5771cc2f160378ad4ad82b8c36294f32

Checkpoint metadata records:

pdd_acc_baked: true
pdd_acc_bake_strength: 1.0
pdd_acc_bake_source: MiniMax-H3-Ref2VA-Acc-8Step.safetensors

Important: runtime head bank

This checkpoint is not intended to be used as a standalone PDD model without the runtime head bank.

Load it as:

MiniMax H3 Ref2VA PDD baked checkpoint
        +
MiniMaxH3PDDAccApply
        +
official Ref2VA PDD-Acc 8-step file

Recommended Apply settings:

nfe           = 8
lora_strength = 0.0
head_strength = 1.0

Sources

Official PDD-Acc release:
https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs

Official Ref2VA PDD-Acc file:
https://huggingface.co/alibaba-pai/MiniMax-H3-Acc-LoRAs/resolve/main/MiniMax-H3-Ref2VA-Acc-8Step.safetensors

ComfyUI PDD-Acc runtime:
https://github.com/Jalen-Brunson/ComfyUI-MiniMax-H3-PDD-Acc

Base MiniMax H3:
https://huggingface.co/Comfy-Org/MiniMax-H3

Notes

Experimental community derivative for ComfyUI workflows.

Follow all applicable upstream model licenses and terms.

For reproducible A/B tests, keep seed, references, prompt, dimensions, NFE, sampler and sigma settings identical.