Nodes/BSAI-ComfyUI-vdn-minimax-h3/BSAI VDN-H3 Loader · Weight Load (backbone + linear branch + LoRA)
ComfyUI Node
BSAI VDN-H3 Loader · Weight Load (backbone + linear branch + LoRA)
Load VDN-H3: MiniMax H3 backbone + VDN stage (linear_branch hybrid-attention branch + default/turbo LoRA). Place stage directories under ComfyUI/models/vdn/; see README for layout.
BSAI VDN-H3 Loader · Weight Load (backbone + linear branch + LoRA)
- MODEL
- info
- stage_dir
◄backbone▾►
◄vdn_stagestage-dmd-step-250►
◄merge_linear_branchtrue►
◄software_fallbacktrue►
◄quality_mode🎨 Quality (16 steps + dual LoRA)►
◄merge_default_lorafalse►
◄merge_turbo_loratrue►
◄lora_strength1.00►
◄weight_dtypedefault►
CategoryBSAI/VDN-H3
Inputs (9)
| Name | Type | Default | Description |
|---|---|---|---|
| backbone | COMBO | Backbone transformer weights. The auto option picks a ComfyUI-compatible MiniMax H3 bf16/fp16 backbone from models/diffusion_models (official VDN h3-base shards use diffusers layout which ComfyUI cannot load directly); other entries are single diffusion_models files (e.g. FastH3 4-step distill / 10Eros bf16). | |
| vdn_stage | COMBO | stage-dmd-step-250 | VDN weight stage directory: stage-dmd-step-250 = 8-step turbo (recommended), stage-b-step-2000 = 50-step quality. Place under ComfyUI/models/vdn/. |
| merge_linear_branch | BOOLEAN | true | Merge the VDN linear_branch hybrid-attention branch (needs FA4; software fallback if unavailable). |
| software_fallback | BOOLEAN | true | When FA4 is unavailable, enable software fallback (short-conv + gated hybrid attention), quality near-native. |
| quality_mode | COMBO | 🎨 Quality (16 steps + dual LoRA) | Quality mode: Speed: 8-step sampling + turbo LoRA, fastest but moderate quality Quality: 16-step + weak turbo + default, significantly better (recommended) VDN Pure: relies on linear_branch hybrid attention, 8-step + no LoRA, near-lossless (needs FA4 or software fallback) |
| merge_default_lora | BOOLEAN | false | Merge the 50-step quality LoRA from stage/adapters/default. |
| merge_turbo_lora | BOOLEAN | true | Merge the 8-step DMD2 distill LoRA from stage/adapters/turbo (speed + anti-artifact core). |
| lora_strength | FLOAT | 1.000–2 | LoRA merge strength. |
| weight_dtype | COMBO | default | 4 options: default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2 |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| info | STRING | — |
| stage_dir | STRING | — |