Krea 2 Base & Turbo — NVFP4 / FP8 / MXFP8 / INT8 / ConvRot INT8
Quantized versions of Krea 2 Base and Krea 2 Turbo, converted using silveroxides/convert_to_quant (ctq CLI) with ComfyUI-compatible metadata.
Which Model Should I Use?
Train on Raw, run on Turbo. LoRAs trained on the Raw base model transfer well to Turbo for inference. Raw is the undistilled base checkpoint — diverse, malleable, ideal for LoRA training and fine-tuning. Turbo is an 8-step distilled checkpoint built for fast, high-quality generation at 1k–2k resolution.
Recommended Inference Settings
Turbo — fast inference, high quality:
- Steps: 8
- CFG: 0.0 (disabled)
- Mu: 1.15
- Resolution: 1024–2048px
Raw — undistilled base model:
- Steps: 52
- CFG: 3.5
- Resolution: up to 1024px
Text Encoder
Krea 2 requires the Qwen3-VL 4B text encoder. Load it in ComfyUI's CLIPLoader with type set to krea2.
Download: qwen3vl_4b_fp8_scaled.safetensors
VAE
Krea 2 uses the same VAE as Anima:
Download: qwen_image_vae.safetensors
Available Formats
| File | Format | Notes |
|---|---|---|
krea2_base_fp8.safetensors |
FP8 Tensorwise | Recommended for RTX 40xx |
krea2_base_mxfp8.safetensors |
MXFP8 | Recommended for RTX 50xx Blackwell |
krea2_base_nvfp4.safetensors |
NVFP4 | RTX 50xx Blackwell only |
krea2_base_int8.safetensors |
INT8 Row-wise | Recommended for RTX 30xx |
krea2_base_convrot_int8.safetensors |
INT8 + ConvRot | Best quality INT8, RTX 30xx |
krea2_turbo_fp8.safetensors |
FP8 Tensorwise | Recommended for RTX 40xx |
krea2_turbo_mxfp8.safetensors |
MXFP8 | Recommended for RTX 50xx Blackwell |
krea2_turbo_nvfp4.safetensors |
NVFP4 | RTX 50xx Blackwell only |
krea2_turbo_int8.safetensors |
INT8 Row-wise | Recommended for RTX 30xx |
krea2_turbo_convrot_int8.safetensors |
INT8 + ConvRot | Best quality INT8, RTX 30xx |
Which Format Should I Use?
- RTX 30xx → INT8 ConvRot for best quality, INT8 for fastest
- RTX 40xx → FP8
- RTX 50xx Blackwell → NVFP4, MXFP8, or FP8, your choice
Conversion Details
All variants converted with:
--simple(round-to-nearest, no learned rounding)--comfy_quant(ComfyUI-compatible layer names and metadata)--save-quant-metadata(quantization info embedded in file)--low-memory(reduced peak RAM usage during conversion)--flux2profile
ConvRot variants additionally use Hadamard rotation (--convrot) prior to
quantization to reduce outliers and improve accuracy.
Usage in ComfyUI
Load via the standard diffusion model loader node. Requires a ComfyUI build with comfy_quant support. Recommended nodes: Bob Johnson's Nodes
Original Model
Krea 2 is developed by Krea.ai.
Original model: krea-community/krea-2
License
These quantized weights are derived from Krea 2 and are subject to the Krea 2 Community License Agreement (v.1, June 22, 2026).
Key terms:
- Free for personal and commercial use under $1M USD annual revenue
- Over $1M annual revenue requires an Enterprise License from Krea
- Derivatives (including these quantizations) must be distributed under the same license with attribution
- You must include "Krea" at the beginning of any derivative model name
- Content filtering required for deployed applications
Full license text: krea.ai/krea-2-licensing
Attribution: Krea 2 is licensed under the Krea 2 Community License Agreement.
Quantized by
Winnougan
ComfyUI node developer & AI content creator.
Ko-fi: ko-fi.com/winnougan
YouTube: youtube.com/@Winnougan