This model is the conclusion of several tests and experiments regarding ANIMA LoRA.

• Extracting LoRA: https://civarchive.com/articles/28382/lora-internals-lets-look-in-the-beast-entrails

• ANIMA LoRA training: https://civarchive.com/articles/29508/training-an-anima-lora-one-of-the-thousands-way-to-do-it

• Discussions regarding training Text encoder layers and/or the LLM adapter and checking which layers are included in a LoRA: https://civarchive.com/articles/29594/no-python-no-problem-check-the-names-of-the-layers-of-a-safetensors-easily

• Discussions regarding abliterated version of the Text encoder: https://huggingface.co/DavidAU/Qwen3-0.6B-heretic-abliterated-uncensored

The results of those tests (and associated scripts) are summarized here: https://huggingface.co/n-Arno/ANIMA_extract (you will need a HF account, but the "gate" is automatic).

Now, this LoRA is a concatenation of 3 DiT only LoRA (merged) and the heretic TE LoRA (with different dim/alpha for both part): that's both "megamix" and "qwen-heretic" provided in the HF repo.

Please note that until the loading is fixed for both Comfy and Forge Neo, they may not load the TE keys.

But regardless, i'll be using it so, why not upload it here to link properly my pictures to it 😉

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.