This is one Spicy model; it massively improves Kreas' NSFW capabilities.
I like er_sde sampler and 6-12 steps
Seems to has reduced art style abilities somewhat.
Added int8_convrot Version: https://civarchive.com/api/download/models/3102188?fileId=2985178 It seems slightly better quality and more natural look.
Description
Initial release, unlocks full NSWF Anatomy and many poses.
FAQ
Comments (20)
So going by the examples, I take it this is a finetune of Krea 2 to make it look like Flux.1 dev?
I hope to add back a bit of analogy photography/amateur look in the next version, but training NSFW almost always causes some skin realism degradation, probably I haven't found the best generation settings yet, adding more film grain can help.
@J1B My remark was more in relation to the Flux Chin that appears in most examples, and general Flux Girl appearance, and exaggerated/contrasting lighting, not just skin texture
@kangaru861 Noted. But about 15% of people have a cleft chin (including myself) so I don't mind it as much as some people, it seems to be in less than %50 of faces, much less than Flux Dev.
Personally, there is very little variation in faces, which looks like classic 1girl. I suspect the finetune dataset mostly consists of Flux/SDXL (Illustrous) gens?
@redhome Yeah fair point, I want to improve that in my next version, I rushed this one out as quite a few people were demanding I release what I had. But every model has its preferred faces in my experience.
@J1B I really liked your work on the older models (for Illustrous), but I think Krea 2 gives us much more quality in terms of photorealistic imagery, so the fine-tune input need to be of much higher quality....
The workflow is really interesting! Got a lot of creativity out of it. SeedEnhancer?
I think SeedVarianceEnhancer node is great, and is a must for some models. between 3-8 seems good for Krea 2
There is something very particular about this workflow. With 1920x1088, it has a low res feel. Like it's pixelated somehow. It's better with seedvr2 upscale but there is something "rough" about it. Any ideas? I can't see any specific setting that is an issue.
hmm, it could be the custom Sigmas I made being quite resolution depended you could try a basic Scheduler with Simple or Beta Schedulers , I will do some testing, I think I have seen what you are seeing when using some loras or on particular backgrounds.
@J1B Seems possible. I don't know. I tried Euler and Beta, it was better but there was an issue with the number of steps. Almost like at 24 steps with Raw+Turbo it was worth than 12 steps. But then other images it went back again. It's like it's always a hassle to find the right settings with enough imagination and not low res splotchy.
@brnfd24434343d Strange as I have been testing tonight and my model is way more smooth that all the other models I tried and doesn't show this pixelated noise like even the base model does. I am liking er_sde sampler with Kera 2.
@J1B So... When I started using your workflow the model wasn't out. I have some testing with your model but not a ton. Mostly I've been using raw+turbo. I can confirm issues with your model match the others. I didn't think it would matter much. Something is clearly "off" on many of these generations. It's like a moving target though, something is wrong but when you fix it, something else is fucked up.
I had this too. His sharpening and image "correction" is WAY too much. That's why it blows out only some images, those that are already sharp get fried, those that are bad get a lot of help. Turn all that shit off and see if it helps.
Great model, truly nice work.
The workflow looks great, but is a bit of a beast to get running. One issue though, the upscalers are looking for an image name.png, even before I generate a single pixel. What am I missing?
Any help is appreciated.
I just generated a batch of small/normal sized images, with the upscale sections disable and then upscale just the best one, you can pipe it though so it upscales every image, but I find it inefficient.
@J1B Thanks for the reply, I understand your thinking, thanks.
I get better results using your workflow/model with krea2filterbypass set at 1
Can't wait for V2 (Ghost pepper 🌶️) maybe you can combine both Realism Engine Krea 2 and [KREA 2] Mystic XXX (I think this Lora has the bypass filter built in). Great work as always.
Details
Files
jibMixKrea_v10JalapeO.safetensors
Mirrors
Jib Mix Jalapeno INT8 Convrot - Krea2 - er_sde,6-12steps.safetensors
Jib Mix Jalapeno INT8 Convrot - Krea2 - er_sde,6-12steps.safetensors
Jib Mix Jalapeno INT8 Convrot - Krea2 - er_sde,6-12steps.safetensors
Jib Mix Jalapeno INT8 Convrot - Krea2 - er_sde,6-12steps.safetensors
Jib Mix Jalapeno INT8 Convrot - Krea2 - er_sde,6-12steps.safetensors
Jib Mix Jalapeno INT8 Convrot - Krea2 - er_sde,6-12steps.safetensors