I took the Krea2 BF16 Safetensors from <a target="_blank" rel="ugc" href="https://huggingface.co/krea/Krea-2-Turbo/tree/main">https://huggingface.co/krea/Krea-2-Turbo/tree/main</a> and made GGUF's Q8_0 and Q4_0, I also have the FP8Mixed version. I tested both, and they worked with GGUF - I used this particular GGUF loader, but there are several out there for Krea, but this version I installed for Ideogram 4 GGUF and it just happens to work for Krea2 as well - <a target="_blank" rel="ugc" href="https://github.com/molbal/ComfyUI-GGUF">https://github.com/molbal/ComfyUI-GGUF</a>I do also have Krea2 RAW GGUF, but trust me it doesn't really make images, I only have it in case I can use it for Trained.

Krea2 Tubro Q8 from BF16

Converted from the full FP16 Krea2 model directly from Krea2 Hugginface repository.FP8-Mixed models are specialized models that keep FP16 for the important weights, but quantize the "useless" ones. Making the model smaller, but still almost "full strenght"You need to make the model compute_dtype FP16 to use otherwise it'll kick back errors. I use KJNode's "Checkpoint Loader" which allows you to set the set the compute_dtype to FP16. (It really does need to be FP16, or FP32, it will not work with BF16) The downside - it's very "polished" look, almost always digital illustration.

Files with SHA256: e9aa38705d40f9b5faeb49d1a01af571cafc10e45b8066b4c82d5a6f654bcfb7