Views
No views yet
int8_tensorwise)UNETLoader) node — no OTUNetLoaderW8A8 /
ComfyUI-INT8-Fast custom loader required.| File | Size | Source BF16 | Notes |
|---|---|---|---|
Krea2-Turbo-int8-ConvRot.safetensors | ~13.2 GB | krea2_turbo_bf16.safetensors (Comfy-Org/Krea-2) | 8-step distilled |
Krea2-Raw-int8-ConvRot.safetensors | ~13.2 GB | krea2_raw_bf16.safetensors | Undistilled base |
ComfyUI/models/diffusion_models/..comfy_quant JSON)1{
2 "format": "int8_tensorwise",
3 "orig_dtype": "torch.bfloat16",
4 "convrot": true,
5 "convrot_groupsize": 256,
6 "per_row": true
7}{"convrot": true, "per_row": true} (no "format": "int8_tensorwise")
do not load in stock ComfyUI ≥ 0.27 — this repo replaces those.int8_tensorwise + ConvRot).safetensors into models/diffusion_models/UNETLoader), weight_dtype: defaulttype: krea2) → CLIPTextEncode → KSampler / FLS → VAEDecodeLoraLoader on the MODEL output. Prefer
clip_strength = 0 for Krea UNet-only LoRAs so text encode can cache.1ctq -i krea2_turbo_bf16.safetensors \
2 -o Krea2-Turbo-int8-ConvRot.safetensors \
3 --int8 --convrot --convrot-group-size 256 \
4 --scaling_mode row \
5 --comfy_quant --save-quant-metadata --krea2 \
6 --simple --low-memory --device cudakrea2_raw_bf16.safetensors). --scaling_mode row is mandatory.1from safetensors import safe_open
2import json
3with safe_open("Krea2-Turbo-int8-ConvRot.safetensors", framework="pt") as f:
4 raw = f.get_tensor([k for k in f.keys() if k.endswith(".comfy_quant")][0]).tolist()
5 print(json.loads(bytes(raw)))
6# Must include: format=int8_tensorwise, convrot=True, per_row=True, convrot_groupsize=256ctq)--krea2 (sensitive first/last/modulation layers kept high precision)