The continuing journey of UnNamedA will be over HERE.

First ANIMA 2.9B version made using this method.

Hotfix: I just uploaded a new version of the safetensors file, annoted BF16. This version is full BF16, no mixed FP16 tensors that failed with Comfy :D

It is HERE! A major update for UnNamedA, my realistic model based on ANIMA. This time, everyone should get good realistic pictures, both using NLP or tags! 😍

This version has my Turbo V2.9 baked in because it really gives better results and this way, it will cost less (no need to include a LoRA on-site): if you want to unbake it (for training for example), it was added around ~0.7 strength, so -0.7 should be enough.

Usage (also check the showcase and below posts):

• ER SDE / SGM Uniform

• Shift 3

• Steps 16

• CFG 1.2

• Negative prompt: worst quality, low quality, blurry, anatomical nonsense, text, watermark (don't add scores)

• Positive prompt:

  • Tags: masterpiece, best quality, realistic, <your prompt>

  • NLP: Either nothing, realistic, <your prompt> or Digital photography, <your prompt> for example

• Adetailer: NOT REQUIRED ANYMORE 😉

• Hires.Fix: denoise 0.3, same number of steps and CFG

It will be available on-site in around 3h and will replace V1.

Previously, V1:

Here is my very first version of UnNamedIXL for ANIMA.

Based on my own finetune of Cosmos2 plus @Felldude incredible work on Final Cut, with some more additional training to bring some of UnNamed original aesthetic.

Disclaimer: I am making it early access at low cost to avoid automatic scraping download for the first day. As I usually tip 100 Buzz to the first few pictures made using my models, "buying" it in EA will most probably be refunded 😉

Two versions provided:

• pruned: That's the base diffusion model.

• full: All In One, VAE+TE+Turbo LoRA included.

Usage (what i use, feel free to adapt):

• sampler/scheduler: ER SDE/Beta

• 32 steps

• CFG 4

• Flow shift 3

• Positive: masterpiece, best quality, realistic, detailed skin, highres, <your prompt> (optionally, you can add uncensored)

• Negative: worst quality, low quality, blurry, anatomical nonsense, jpeg artifacts, bad anatomy, extra limbs, deformed, watermark, text, signature, copyrights name, bad hands, missing fingers, extra digit, fewer digits, artistic error, ye-pop, deviantart, patreon logo

Now, it works nicely as a standalone, but it gives even better results (IMHO) with the Turbo LoRA, in this case i switch to:

• 16 steps

• CFG 1.2

You can check in the showcase some pictures made with and without.

Recipe (how it was made):

Both FinalCut and my uncensored Cosmos2 were added to base ANIMA v1 + some LoRA for good measure:

(venv) root@nyarlathotep:~/ANIMA# python
>>> import sd_mecha as sdm
>>> models = [
...   "animaFinalcut.safetensors",
...   "uncensored_cosmos2.safetensors",
... ]
>>> models = list(map(sdm.model,models))
>>> recipe = sdm.add_difference_ties(sdm.model("anima-base-v1.0.safetensors"),*models,alpha=0.6)
>>> sdm.merge(recipe,output="unnamed-init.safetensors")

(venv) root@nyarlathotep:~/ANIMA# python merge_lora.py unnamed-init.safetensors nArnima_lora_V2.safetensors 0.3 unnamed-v0.safetensors

---
Here, 12 epochs (6312 steps) of further training was done on v0 to make an upgrader lora.
---

(venv) root@nyarlathotep:~/ANIMA# python merge_lora.py unnamed-v0.safetensors unnamed-upgrader.safetensors 0.6 UNNAMED_ANIMA_V1.safetensors

Thanks everyone! 🥰🥰🥰

PS: the next step will probably some distillation from a ZiT model as i want to test build a student/teacher workflow from scratch (for those who didn't know, i do most of this to learn how to handle AI low level stuff for work, at one point i'll probably try making a DiT model from scratch ha ha)

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.