Anima 2.9B preview v1 convrot quantization

Convrot quantization of Anima 2.9B preview v1. All credits to @gazingstars321 for creating that model. Anima 2.9B is a diffusion model based on Anima-base_v1 by circlestonelabs and is thereby a derivative of Anima. It's a self-funded finetune, so check out the huggingface page, if you want to support it. It has additional layers and was trained on 1.7M images additional to the dataset of anima base with knowledge cutoff July 2026. Can be used with Forge Neo or ComfyUI with latest versions. Also LoRAs trained with anima-base can be used with this model in Forge Neo, though they may not all be working correct. I made two different quantized versions. The first one is made using svd learned rounding for less rounding errors. The second one keeps more layers in bf16 precision and is therefore slightly larger. The differences between them and the original full bf16 version are usually subtle except for very short prompts. Due to the additional layers Anima 2.9B is slower in inference than Anima-base, though using the int8convrot quantization makes it run at a similar speed.

recommended quality tags: masterpiece, best quality, highres, absurdres, newest, year 2025
negative prompt: lowres, bad quality, worst quality, score_1, score_2, score_3, jpeg artifacts, signature, watermark
Sampler: ER-SDE
Scheduler: Beta
Steps: 30-50
CFG: 3-5

Showcase images were made with ~1.5k resolution mostly, but I think the model generates slightly better images with 1k resolution. Look at the images directly to get a grasp of what the model is capable of, in the overview the images are scaled to 450px width and may look blurry.

Attribution Notice

“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.”