ComfyUI Node
AuK Models Loader (Low VRAM)_Doc
Select local models. Low VRAM uses CPU Qwen and VAE; GPU DiT uses half precision. For 8 GiB GPUs use low_vram; see the included validation report for tested limits.
AuK Models Loader (Low VRAM)_Doc
- engine
◄model_name▾►
◄qwen_name▾►
◄memory_modelow_vram►
◄dtypebf16►
◄device▾►
◄sequential_cfgtrue►
CategoryAuK_Doc
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| model_name | COMBO | 1 options: No AuK models found | |
| qwen_name | COMBO | 1 options: No Qwen2.5-Omni models found | |
| memory_mode | COMBO | low_vram | low_vram: Qwen + VAE on CPU, half-precision DiT on GPU. balanced: Qwen and DiT alternate on GPU, VAE on CPU; requires more VRAM and is not recommended for 8 GiB. |
| dtype | COMBO | bf16 | 2 options: bf16, fp16 |
| device | COMBO | 1 options: CUDA unavailable | |
| sequential_cfg | BOOLEAN | true | Evaluate CFG branches sequentially, reducing activation memory while taking more time. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| engine | AUK_ENGINE_Doc | — |