Update - Uncensored V0

Since it was done, i'll publish it before going back to my cave. This was my very first test at "uncensoring" a realistic model and boy did i see some body horror in the very first few epochs. 😆

In the end, in the course of 5 trainings, 29 epochs in total, around 1000 pictures (split into several dataset) for a grand total of 14500 steps, here is a V0 of an uncensored version of this model that should be a proper base for further training. It is far from perfect, has a bias toward my own realistic aesthetic, but it should answer nicely for some pictures (PS: still no ANIMA in there).

Sampler/scheduler: Euler/Simple for clean pictures or ER SDE/Beta for some more variations.

Steps: 32 as-is, i'll try with the Turbo LoRA 😉

CFG: 4 should be ok

Triggers: uncensored, realistic

Base V0

One last model before leaving on a small break (well maybe not, the AI Doll version of Lulu is almost done)

It is well known that ANIMA is not great for realistic stuff. It has been heavily trained for anime stuff, bringing back realistic features would break the training rather fast (i know, i have tried).

BUT, follow me here. ANIMA is based on Cosmos-Predict2 architecture which has been trained on realistic pictures. The only difference is that Cosmos-Predict2 use T5 as Text Encoder and ANIMA use Qwen3-0.6B thanks to an added LLM adapter in the model.

What if someone stupidly tried to add this LLM adapter from ANIMA into Cosmos-Predict2? (and renamed all the keys to match the naming convention used in ANIMA). (What if this someone was yours truly 😁?)

Would it allows having a realistic model that works using the same TE and VAE as ANIMA? Even on-site? With danbooru tags even?

Well, the answer is yes, it works! And the next step would probably be to merge this model, for example using DARE and TIES, into ANIMA to import some realistic features allowing an easier realistic model training for ANIMA. 😉 (aka: i'll be able to cook a proper ANIMA version of UnnamedIXL)

Anyway, regarding licensing:

• This model leverage as-is all the other layers from Cosmos-Predict2, as such, it is a derivative work of Cosmos-Predict2

• This model leverage as-is the LLM adapter layers from ANIMA, as such, it is a derivative work of ANIMA, which is also a derivative work of Cosmos-Predict2

So, it is fully under the current license of ANIMA.

Regarding LoRAs: ANIMA LoRA will not work great (or at all) with this model as it is. Much like an SDXL LoRA and an IXL LoRA. But a merge may help.

PS: Since i didn't want to "obstruct" other creators that may find this model useful, i am releasing this one as-is before testing other things (which is what i'll probably cook during my break).

Have fun everyone, thanks for reading and see you in two month!

Legal

The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial License. Copyright CircleStone Labs LLC. IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.