79 lines
3.0 KiB
Markdown
79 lines
3.0 KiB
Markdown
# ComfyUI MiniMax H3 Text Encoders
|
|
|
|
One loader node for recovered MiniMax-H3 text encoders.
|
|
|
|
## What it adds
|
|
|
|
One node:
|
|
|
|
`Load MiniMax H3 Text Encoder`
|
|
|
|
Select one package:
|
|
|
|
- Recovered Qwen3-VL-8B BF16.
|
|
- Recovered Qwen3-VL-8B INT8 ConvRot.
|
|
- Physical MiniMax-H3 pruned-24 INT8 ConvRot with balanced ARA.
|
|
- Physical MiniMax-H3 pruned-24 NVFP4/AWQ with balanced ARA.
|
|
|
|
The node returns `CLIP`. Connect it to official `MiniMaxH3ImageToVideo`.
|
|
|
|
No adapter node. No ARA node. No cache node. No barrier node.
|
|
|
|
## Install
|
|
|
|
1. Install this extension in `ComfyUI/custom_nodes/`.
|
|
2. Copy one complete text-encoder package directory into `ComfyUI/models/text_encoders/`.
|
|
3. Restart ComfyUI.
|
|
4. Drag `example_workflows/minimax_h3_recovered_8b_int8_convrot_t2v.json` into the ComfyUI canvas.
|
|
|
|
This is the frontend/canvas workflow. It includes all visible links, including `Load MiniMax H3 Text Encoder` → `MiniMax H3 Image to Video.clip`.
|
|
|
|
`minimax_h3_recovered_8b_int8_convrot_t2v_api.json` is a separate API prompt for `POST /prompt`. Do not drag the `_api.json` file into the canvas; API prompt JSON has no frontend link layout.
|
|
|
|
For recovered 8B, copy the directory as a unit. Do not select, rename, or configure the ARA or adapter separately:
|
|
|
|
```text
|
|
ComfyUI/models/text_encoders/
|
|
└── recovered_8b/
|
|
├── minimax_h3_recovered_8b_manifest.json
|
|
├── qwen3vl_8b_minimax_h3_recovered_int8_convrot.safetensors
|
|
├── qwen3vl_8b_minimax_h3_recovered_bf16.safetensors
|
|
├── ara.safetensors
|
|
└── conditioning_adapter.safetensors
|
|
```
|
|
|
|
The node discovers `recovered_8b/` from its manifest. The manifest gives the internal filenames and SHA-256 values for the selected base, ARA, and adapter. The node verifies and loads all three automatically.
|
|
|
|
## Model files
|
|
|
|
Download complete package directories from [SearchingMan/MiniMax-H3-Text-Encoders](https://huggingface.co/SearchingMan/MiniMax-H3-Text-Encoders/tree/main/text_encoders) and place them under `ComfyUI/models/text_encoders/`.
|
|
|
|
| Package directory | Loader option | Base format |
|
|
|---|---|---|
|
|
| `recovered_8b/` | `recovered_8b_int8_convrot` | INT8 ConvRot |
|
|
| `recovered_8b/` | `recovered_8b_bf16` | BF16 |
|
|
| `pruned24_balanced/` | `int8_convrot` | INT8 ConvRot |
|
|
| `pruned24_balanced/` | `nvfp4_awq` | NVFP4/AWQ |
|
|
|
|
Install one package at a time if disk space is limited.
|
|
|
|
## Scope
|
|
|
|
Text-only MiniMax-H3 T2V. Image, first-frame, last-frame, and reference inputs are not supported by the recovered 8B package.
|
|
|
|
The recovered 8B node loads the required trained 4096-to-5120 adapter inside the loader. The adapter is not a visible node.
|
|
|
|
## Required official files
|
|
|
|
Get these from `Comfy-Org/MiniMax-H3`:
|
|
|
|
- `diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors`
|
|
- `vae/minimax_h3_video_vae_fp16.safetensors`
|
|
- `vae/minimax_h3_audio_vae_fp32.safetensors`
|
|
|
|
## License
|
|
|
|
Code: MIT.
|
|
|
|
Weights: Apache-2.0 modified derivatives of Qwen3-VL. Read the model repository `LICENSE` and `NOTICE`. The model repository does not contain MiniMax-H3 diffusion-model or VAE weights.
|