ComfyUI Node
Yogurt Sa2VA Model Loader
Loads a Sa2VA model once and returns a reusable model object. Connect this output to multiple Yogurt Sa2VA segmentation nodes to avoid reloading the same model.
Yogurt Sa2VA Model Loader
- sa2va_model
◄model_nameByteDance/Sa2VA-Qwen3-VL-4B►
◄use_8bittrue►
◄use_flash_attntrue►
◄force_reloadfalse►
CategoryYogurtSa2VA
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| model_name | COMBO | ByteDance/Sa2VA-Qwen3-VL-4B | 7 options: kumuji/Sa2VA-i-1B, ByteDance/Sa2VA-Qwen3-VL-4B, ByteDance/Sa2VA-InternVL3-2B, ByteDance/Sa2VA-Qwen2_5-VL-3B, ByteDance/Sa2VA-Qwen2_5-VL-7B, ByteDance/Sa2VA-InternVL3-8B, +1 |
| use_8bit | BOOLEAN | true | Use 8-bit quantization to reduce memory usage (requires bitsandbytes). |
| use_flash_attn | BOOLEAN | true | Use flash attention for faster processing (requires flash-attn). |
| force_reload | BOOLEAN | false | Unload and recreate the cached model for this model/quantization/attention configuration. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| sa2va_model | YOGURT_SA2VA_MODEL | — |