ComfyUI Node
H3 PDD Acceleration (8-Step)
Loads Alibaba PAI's MiniMax-H3 PDD acceleration file directly, including its displacement heads, and outputs the exact eight-step sigma schedule. Use Euler and CFG 1.0. Supports the FL2VA/Ref2VA hybrid checkpoints by taking only their late AdaLN adapters from the Ref2VA PDD file.
H3 PDD Acceleration (8-Step)
- model
- model
- sigmas (8 NFE)
◄pdd_lora_name▾►
◄hybrid_modeOff (matching full model)►
◄hybrid_ref2va_pdd_lora[None]►
◄manual_hybrid_start_block20►
◄strict_scheduletrue►
◄allow_experimental_shiftsfalse►
CategoryMiniMax H3/model_patches
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | Native ComfyUI MiniMax-H3 MODEL. | |
| pdd_lora_name | COMBO | Alibaba PDD file matching the base: FL2VA for FL2VA and hybrid checkpoints; Ref2VA for a full Ref2VA base. Copy the original .safetensors into models/loras. | |
| hybrid_mode | COMBO | Off (matching full model) | For smhfacct hybrid checkpoints, select the checkpoint's same block boundary. Off loads one matching PDD file. Hybrid mode keeps FL2VA adapters/heads everywhere except the selected late block AdaLN adapters. |
| hybrid_ref2va_pdd_loraopt | COMBO | [None] | Required only in hybrid mode: the Alibaba Ref2VA PDD file. It is read only for late AdaLN adapters. |
| manual_hybrid_start_blockopt | INT | 200–49 | Used only when Hybrid Mode is Manual. |
| strict_scheduleopt | BOOLEAN | true | Recommended: reject a sampler schedule that differs from the exact eight PDD blocks. Turning this off permits experiments but can reduce quality. |
| allow_experimental_shiftsopt | BOOLEAN | false | Alibaba's released PDD heads were distilled at video/audio shifts 12/3. Leave disabled for production. Enabling permits other shifts but cannot retarget the learned heads and may substantially reduce quality. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |
| sigmas (8 NFE) | SIGMAS | — |