Converted from https://huggingface.co/Comfy-Org/Krea-2/tree/main/diffusion_models
All credit belongs to the original model author. License is the same as the original model.
Quantized qwen3vl text encoder: link.
This quantized ckpt uses mixed precision. Instead of quantizing the entire model to targeted format. This ckpt keeps important layers in high precision for better quality. (Similar to M and XL suffixes in GGUF)
All txtfusion layers are in high precision. You don't needs to worry the precision when applying "Refusal-Reduction" LoRA (or whatever new LoRAs that try to nuke the txtfusion layers and the safety filter).
Scaled fp8 and nvfp4 are deprecated.
Use int8convrot or w4a8 instead. Faster, higher precision.
Description
FAQ
Comments (5)
any comparisons with same seed?
stuck by civitiai potato scanner server
Hands down the best version of Krea 2 Turbo on site (I'm using NVFP4 version). Works excellent with LoRAs and Krea 2 Bypass works with just 1 strength! Thank you so much!
Do you have any further plans for this model? This model claims to be trained without any synthetic images, which should suit your preferences perfectly.
The model is quite large at present, yet the difference between training with full FP8 and BF16 seem no big. Could mixed quantization (retaining BF16 for text processing layers) further narrow these gaps?
Yes, I have further plans.
Krea 2 looks extremely promising so far. Except that content filter, which prevents all kinds of characters interactions. I'm sure it will be nuked by the community.
_ further narrow these gaps
Yes.