OpenVDN MiniMax H3 for ComfyUI
This repository is a ready-to-place model bundle for the OpenVDN workflows in MiniMax H3 Audio T8. It contains the exact local model set used for the project's corrected DMD8 validation.
T8star links
- Bilibili: T8star's channel
- YouTube: @T8star-Aix
- Seedance API: sign up
- Online AI apps: T8star on RunningHub
- ComfyUI all-in-one package: Quark Drive download
- Hugging Face: t8star
The Seedance and RunningHub URLs above contain referral or invite codes supplied by T8star.
Install
-
Install or update the custom node:
cd ComfyUI/custom_nodes git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8 -
Log in to Hugging Face after your access request is approved, then download this repository directly into
ComfyUI/models:hf auth login hf download t8star/Vdn-Minimax-H3-Comfy --local-dir ComfyUI/models -
Restart ComfyUI and load one of the OpenVDN workflows from
examples/workflows/10-speedin the custom-node repository.
The repository already uses ComfyUI's folder names, so no file rearrangement is needed.
Included files
| ComfyUI folder | File | Size |
|---|---|---|
diffusion_models | minimax_h3_fl2va_int8_convrot.safetensors | 34.04 GB |
diffusion_models | minimax_h3_fl2va_pruned_int8_convrot.safetensors | 20.97 GB |
diffusion_models/OpenVDN/vdn-minimax-h3 | OpenVDN DMD/Stage-B configs, 50-block branch, native adapters, and T8 curve-projected Turbo adapter | 6.62 GB |
text_encoders | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | 15.69 GB |
vae | minimax_h3_video_vae_fp16.safetensors | 5.21 GB |
vae | minimax_h3_audio_vae_fp32.safetensors | 0.61 GB |
Exact byte sizes and SHA-256 hashes are listed in MODEL_MANIFEST.json.
Full and supported pruned bases
OpenVDN DMD8 can use either minimax_h3_fl2va_int8_convrot.safetensors or
minimax_h3_fl2va_pruned_int8_convrot.safetensors from this bundle. Keep the same
workflow and select the base in UNETLoader; the T8 Composer detects its structure and
chooses the matching adapter automatically.
The published Turbo adapter has 51 AdaLN LoRA targets trained for a 2688-column input.
For the bundled eight-column pruned base, T8 supplies a separately identified
curve-projected adapter under
stage-dmd-step-250/adapters/turbo_pruned_curve_fl2va/. It retains 208 directly
compatible modules and represents the 51 AdaLN modules as eight-column LoRA factors plus
51 bias residuals, for 310 applied patches. No extra LoRA node is needed.
Selection is bound to the raw adaln_t_table content hash, not to a filename. A pruned
checkpoint with an unknown curve table fails before sampling instead of silently using
the wrong adapter. Renaming an unsupported checkpoint does not bypass this check.
Supported workflows
The custom node provides formal Advanced workflows for:
- T2VA
- first-frame I2VA
- last-frame L2VA
- first-and-last-frame FL2VA
- single-image and multi-image Ref2VA
- reference video with its audio
- standalone reference audio
- first-frame plus reference-audio Hybrid
OpenVDN upstream documents T2VA. The other modes are extensions implemented and tested through ComfyUI's native MiniMax H3 conditioning layout by the T8 custom node.
For DMD, the node applies the OpenVDN default and turbo adapters and fixes inference at 8 Euler/native-flow steps with video/audio shifts 12/3. Stage B uses the default adapter at 50 steps. Do not stack another Turbo/EMA LoRA, SLA, VSA, Sol-Attn, BlockCache, or a second model/attention patch on the OpenVDN model.
Validation notes
The full-width validation ran each multimodal workflow separately at 512x288x39 on an RTX 4060 Ti 16 GB. A second strictly serial matrix used the bundled pruned INT8 base at 320x192x39 and covered T2VA, first frame, last frame, first+last, single/multiple reference images, reference video plus audio, standalone reference audio, and first-frame+audio Hybrid. All 9/9 pruned runs applied 800 branch tensors, 104 default patches and 310/310 curve-Turbo patches (259 logical targets plus 51 bias residuals), logged zero LoRA errors, and passed strict H.264/AAC video, audio, and combined decoding with finite non-clipped PCM.
Minimum free VRAM in the pruned matrix was 290 MiB, and only T2VA and I2VA cleared the project's 512 MiB exact-run margin. This proves compatibility, not universal 16 GB safety.
These are mechanical compatibility checks, not a promise that every prompt, image, sound, GPU, or driver will produce the same quality or memory use. The local INT8/ConvRot base is structurally compatible with the OpenVDN release, but this repository does not claim it is byte-identical to OpenVDN's declared upstream BF16 base.
Sources and licenses
- OpenVDN weights and configs: OpenVDN/vdn-minimax-h3, pinned by the custom node to revision
18be6bcc4ee72585eee322ba28b5ccac2cf85ef0. - OpenVDN source implementation: NVlabs/OpenVDN, Apache-2.0.
- Base model: MiniMaxAI/MiniMax-H3.
- Text encoder architecture: Qwen3-VL, Apache-2.0.
- Required custom node: T8mars/comfyui-minimax-h3-audio-T8.
The MiniMax H3 model and derivatives are governed by the MiniMax H3 Community License
Agreement in LICENSE. Its Applicable Territory excludes the European Union, the United
Kingdom, the Republic of Korea, and the United States of America. Distribution and use
outside the Applicable Territory are not authorized by that agreement. Read the full
license and Acceptable Use Policy before requesting access, downloading, redistributing,
or running these files.
The Apache-2.0 texts for OpenVDN source and Qwen3-VL are included under licenses/.
The required MiniMax notice and a description of the packaged conversions are in NOTICE.