How viable would a Krea 2 finetune be?
Krea 2 is a new open weights model, and while it does have a safety filter baked in it's pretty easy to bypass in various ways.
Its license is much less restrictive than Ideogram 4's, and I admit that I didn't actually read the license so I can't say for certain but it doesn't look like there's much in the way of trying out an Anima 2 of sorts on that model.
I've seen people training NSFW artstyle LoRAs on Civit and they work pretty well, plus LoRAs can be combined with ease, so I don't have much reason to doubt that a full proper finetune would work well too.
Only if you can handle 24GB weights. Roughly it would require ~48GB VRAM at least for finetune the model.
The Krea2 model has a very good ability to produce high-resolution images, but using Loras trained with "small" resolutions (such as 1024 * 1024), images generated with high-resolution (2 million pixels) will result in extremely smudging strokes.(The quality decline caused by NSFW is similar, but I think it is particularly strong for thick painted illustrations) It means a trade-off needs to be made between maintaining the model's advantage and training costs
In addition, I tried to mix the anime style that I trained myself.(Involving various specific styles, such as standing illustrations, copperplate prints, pop art, pencil sketches, and celluloid) The feel of using lora mixture is similar to ANIMA default artist mixture: most seeds (not case) have only one dominant style, and a few seeds can show patchwork fusion(mix weight keep at 0.5+0.5)
Although it is still possible to perform two styles of high-frequency and low-frequency special mixing by constraining Lora, the VRAM consumption is huge and the range is limited（only p6k can do，If Comfyui did not optimize the constraint Lora
Because single 5090 and pro6000 seems to weak and my practical work， I didn't have the enough time to conduct further experiments. I wonder if this is also related to the 12 layers of attention
Anima only has 2B parameters, while Krea2 has 12B—which is six times the size of Anima. Circlestone-Labs was already running low on funds when training Anima, so who is going to give them the money to train a model six times larger?
Krea’s license prohibits removing the safety filter from the weights, which could be interpreted as “You cannot train a custom model for nsfw on this”
not actuslly sure here, someone legal clarify for me
Krea’s license prohibits removing the safety filter from the weights, which could be interpreted as “You cannot train a custom model for nsfw on this”
So does the Nvidia Open Model License used by Cosmos. Technically, Anima could be said to have broken 2.1 specifically for training NSFW content, but I absolutely do not see Nvidia's legal teams chasing CircleStone down over that for any Cosmos model whatsoever.
Krea 2 looks like chat gpt 2.0 image all these models looks bad I can see faster if it's Ai image or not weird colors and Blah blah