Added V1.5
Merged with the latest Krea 2 SNOFS v1.5
A merged of Krea 2 with the Krea 2 Identity Edit lora v1.2 adaptation by conradlocke and SNOFS v1.4 LOKR by Ashen [Huggingface] Both baked directly into the model so the separate loras are not required.
Requires the ComfyUI-Krea2Edit node pack, the edit lora is trained with dual conditioning (in-context VAE tokens + image-grounded Qwen3-VL encoding) that stock nodes don't provide. Example Workflow is attached on the Optional File.
Change the GGUF loader node to Load Diffusion Model node if you are using BF16, Int8 or Int4 model.
GGUF loader : https://github.com/molbal/ComfyUI-GGUF
All example images were created using the Q4_0 / Q4_KM GGUF variant at 0.7 MP resolution with a 3:4 aspect ratio. The BF16, Int8_convrot, int4_convrot variants haven’t been tested extensively, so please share your results here.
Precision variants
Several precision variants are available to let you choose the best balance between quality, VRAM usage, and performance:
BF16 —> The highest quality, highest precision, high VRAM usage.
INT8 ConvRot —> Higher precision with lower quantization loss. Requires more VRAM than the INT4 variant.
Q6_K GGUF —> Higher-quality 6-bit quantization, Mostly identical to Q8 with lower vram usage.
Q5_1 GGUF —> High-quality 5-bit quantization with a good balance between quality and VRAM usage.
Q4_0 GGUF / Q4_KM —> More aggressive 4-bit quantization with lower VRAM usage while retaining good image quality (The best balance for low vram user).
INT4 ConvRot —> The most aggressive option, designed for lower VRAM usage and faster inference on supported hardware. Quality is slightly below the Q4 GGUF variants (The fastest).
In general, expect the precision trade off to be approximately:
BF16 > INT8 ConvRot > Q6_K > Q5_1 > Q4_KM > Q4_0 > INT4 ConvRot
Higher precision generally provides better quality, while lower precision reduces VRAM requirements.
Credits
Krea 2 -> Original base model by Krea AI
Krea 2 Identity Edit -> Adaptation by conradlocke
SNOFS -> by Ashen [Huggingface]
This is an unofficial community modification and is not affiliated with or endorsed by Krea.ai.
Description
initial release
FAQ
Comments (12)
Jutst one word for this model... woooow. Short promt or natural language. Does it all, and does it well.
Thanks for your time and effort to make this.
Keep up the good work
you're welcome
First and foremost, Ashen3 is a legend. Instant download as soon as I see something new drop. Thanks for your effort!
Works very well maintaining identity and likeness. It does great at incorporating the subject into new poses in the current or new settings.
Oddly, my biggest usage issue has been maintaining the subject's original pose and position. Sometimes it will make a seated subject stand or place their arms/hands into unprompted poses.
Any suggestions for prompts to hold the subject in their original position or making no changes to the original image?
Try the CFG more than 1 or use higher ref_boost like 3-4
Great checkpoint, thank you. it is amazing
Completely insane model. Feed it anything. Prompt. And the magic happens. Wow! Qwen Edit, Flux, etc. are all light years away behind Krea2 and what can be achieved with it as a base for training.
This Checkpoint is insanely efficient. Even the "NSFW" part feels so much better than other checkpoints.
Nothing short of amazing.
Thank you, glad you liked it.
Excellent checkpoint! It understands prompts perfectly! The POV (Point of View) effect that distorts photos doesn't appear when generating photos without a request! In Forge NEO, the INT8convrot format works perfectly!
Nice, if you can post your int8 result here that would be great 👍
What is the grounding_px sweetspot for maintaining editing adherence and identity likeness?
my sweetspot is 1024px, lower it if you have double image issue.
Do we need anything like Krea2 TextFusion Refusal-Reduction LoRA since it is in the example workflow (but disabled)? And I have found that changing the output resolution completely messes up the result - is that just me?
Btw. int8-convrot seems to perform the same as Q4.