ComfyUI Node
Idol Precision
How precisely the denoiser computes. Keep it in the graph at default too: it also makes idol_fp32.safetensors compute in the dtype chosen here instead of always in fp32. fp32 needs idol_fp32.safetensors to gain anything.
Idol Precision
- model
- MODEL
◄precision▾►
Categorymodel/patch
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| precision | COMBO | default: ComfyUI's choice (fp16 on most GPUs, the precision the Idol SDK runs in), 24 s on an RTX 4090. fp32, fp16 attention: everything in fp32 except attention, 42 s, about 7x closer to the fp32 result than fp16. fp32: the full-precision result, 112 s. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |