Compared to the original model
Slightly more realistic image generation
Improved prompt adherence
Better overall color reproduction and sharpness
Text generating capability has been preserved
This model sits somewhere between Turbo and Raw, leaning around 90% toward Turbo. It still performs well at 8 steps, but increasing the sampling to 10–12 steps can produce additional detail in some cases.
Several layer weights have been adjusted. The model has also been merged with a small selection of LoRAs to enhance realism and expand its uncensored capabilities while maintaining its original strengths.
This is a purely experimental release and may behave unpredictably in certain scenarios. Feedback and testing are greatly appreciated.
Recommended Settings
clownsharksampler
Scheduler: ddim / beta57
Steps: 8 (fast) or 10–12 (higher detail)
Any resolution between 1024-2048 good, between 1536-2048 best
Description
FAQ
Comments (12)
Please add comparison with vanilla model. dont understant why I need it.
The difference between this model and the vanilla version is mentioned in the description. Besides, this is an experimental model using various techniques, so you don't need it as your primary model. However, I'd appreciate it if you could test it and let me know the results
@saalde Ok but what is your end goal? Are you trying to create an NSFW model because the base Krea 2 is completely censored. I would be happy to test it for you if that's your purpose, but if it's like something artistic then probably an artist should test it :)
@ferrrett33 As I stated in the description, 'Slightly more realistic image generation,' and I also tagged the model as 'photorealistic,' so its purpose is quite obvious: 'photorealism.' It doesn't have a specific niche; as a base model, it should be able to do everything, but it features improvements in realism compared to the original model. Additionally, the model is uncensored, meaning it can generate NSFW content.
@saalde Ok fair enough, thanks. I'll test it along with other models and get back to you in a few days time.
Slightly more realistic image generation
Improved prompt adherence
Better overall color reproduction and sharpness
Text generating capability has been preserved
tips：More NSFW。THE Best one。So soft skin，nice light！
Good realism, but likely overbaked: compared to other Krea-2 models I am getting pretty much the same face ("sameface issue") generated all the time, with very little variance. Likely one or more LoRAs baked too strong.
Could you be referring to the situation in the comparison image I recently added? It seems that the changes I made regarding prompt following are causing this. To get a different result, you need to make fundamental changes to the prompt rather than just changing the seed; otherwise, like in the example, minor changes only lead to minor variations. The LoRA ratios I added were truly minimal, so I don't think that's the cause, but I will investigate it. Thank you for your feedback.
@saalde No, I ran it to generate a bunch of images. I pushed it hard (have wildcards describing the various shapes/facial features), but no matter what I do it's still (almost) the same face with different haircuts. Same issue was plaguing realistic Pony and then realistic Illustrious models for years.
Nice work! Could you add a fp16 version for my big GPU <3 ?
AMAZING!!!!