ComfyUI Node
Youtu-VL (Advanced)
A ComfyUI node in 🧪AILab/YoutuVL with 13 inputs and 1 output.
Youtu-VL (Advanced)
- image
- text
◄modelYoutu-VL-4B-Instruct►
◄quantizationNone (FP16)►
◄attention_modeauto►
◄deviceauto►
◄preset_prompt🖼️ Describe Image►
◄custom_prompt►
◄max_tokens512►
◄temperature0.10►
◄top_p0.001►
◄repetition_penalty1.05►
◄keep_model_loadedtrue►
◄seed1►
Category🧪AILab/YoutuVL
Inputs (13)
| Name | Type | Default | Description |
|---|---|---|---|
| model | COMBO | Youtu-VL-4B-Instruct | Select the Youtu-VL model. First run downloads weights to models/LLM/Youtu-VL. |
| quantization | COMBO | None (FP16) | Precision vs VRAM. FP16 gives best quality; 8-bit suits 8-16GB GPUs; 4-bit fits 6GB or less. |
| attention_mode | COMBO | auto | auto tries flash-attn v2 when available, falls back to SDPA. |
| device | COMBO | auto | 4 options: auto, cuda, cpu, mps |
| preset_prompt | COMBO | 🖼️ Describe Image | Built-in instruction for how Youtu-VL should analyze the input. |
| custom_prompt | STRING | Optional override - replaces preset template when filled. | |
| max_tokens | INT | 51264–32768 | Maximum number of new tokens to generate. |
| temperature | FLOAT | 0.100.01–2 | Sampling randomness. 0.1-0.4 is focused, 0.7+ is creative. |
| top_p | FLOAT | 0.0010.001–1 | Nucleus sampling cutoff. Lower values keep only top tokens. |
| repetition_penalty | FLOAT | 1.050.5–2 | Values >1 penalize repeated phrases. |
| keep_model_loaded | BOOLEAN | true | Keep model in VRAM after run for faster subsequent inference. |
| seed | INT | 11–4294967295 | Seed for reproducible results. |
| imageopt | IMAGE | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| text | STRING | — |