speach1sdef178/VDN-H3-INT8-ConvRot-ComfyUI

VDN-H3 INT8 ConvRot for ComfyUI

License: This model is a derivative of MiniMax H3 and is subject to the MiniMax H3 Community License Agreement. Please read the included LICENSE before downloading or using these weights.

INT8 ConvRot VDN stage checkpoint prepared for ComfyUI-VDN-H3-24GB.

This repository contains the VDN stage files only. It does not include the MiniMax H3 base model.

Install

Download the complete folder:

stage-dmd-step-250-int8_convrot_comfyui

and place it at:

ComfyUI/models/vdn/stage-dmd-step-250-int8_convrot_comfyui/

The final local structure must be:

ComfyUI/
└─ models/
   └─ vdn/
      └─ stage-dmd-step-250-int8_convrot_comfyui/
         ├─ model_spec.json
         ├─ linear_branch/
         │  └─ model_int8_convrot_comfyui.safetensors
         └─ adapters/
            ├─ default/
            │  ├─ adapter_config.json
            │  └─ adapter_model.safetensors
            └─ turbo/
               ├─ adapter_config.json
               └─ adapter_model.safetensors

Required ComfyUI node

Use with:

ComfyUI-VDN-H3-24GB

Tested 24 GB preset:

apply_turbo_adapter = true
strength            = 1.0
lora_mode           = merge
branch_weights      = stream
retain_buffers      = auto
attention_backend   = grouped

Connect the same H3 latent used by the sampler to auto_memory_latent.

Tested configuration

Validated on an RTX 3090 Ti 24 GB with MiniMax H3 FL2VA INT8 ConvRot at 0.4 MP.

10 second / 8 step release-build test:

2:08 total sampling
16.06 s/it

Notes

Upload the complete checkpoint structure. Do not upload only the linear branch safetensors file.

The base MiniMax H3 model is distributed separately.