ComfyUI Node
WINT8 Diffuser Loader
A ComfyUI node in WINT8 with 7 inputs and 1 output.
WINT8 Diffuser Loader
- model
◄unet_name▾►
◄model_typeflux2►
◄weight_dtypedefault►
◄on_the_fly_quantizationfalse►
◄enable_quarotfalse►
◄sage_attentionfalse►
◄quant_modetensorwise►
CategoryWINT8
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| unet_name | COMBO | 0 options: | |
| model_type | COMBO | flux2 | 9 options: flux2, z-image, chroma, wan, ltx2, qwen, +3 |
| weight_dtype | COMBO | default | 4 options: default, fp8_e4m3fn, fp16, bf16 |
| on_the_fly_quantization | BOOLEAN | false | Quantize fp16/bf16 weights to INT8 at load time. Only needed if your checkpoint is NOT already INT8. |
| enable_quarot | BOOLEAN | false | Apply Hadamard (QuaRot) rotation to reduce outliers. Improves quality for heavily quantized models. |
| sage_attention | BOOLEAN | false | Replace ComfyUI's attention kernel with Sage Attention for reduced VRAM usage. Requires sageattn package. |
| quant_mode | COMBO | tensorwise | tensorwise: one scale per weight tensor (faster load, lower memory). blockwise: one scale per 128×128 tile (finer granularity, better quality). |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |