Anima-2.9B

First thing first, I want to say a thank you to everyone who's been trying out, sharing, and supporting the model thus far. In less than 12 hours after release, Anima-2.9B has already received native support on major platforms including ComfyUI and Forge-Neo.The model is also available to download from Civitai, where you can share your image and what people had managed to do even with this Preview version, like the RDBT Distilled Turbo LoRA for instance.Anyway, LoRA training is now supported via my training GUI and you can train + load LoRAs natively in both ComfyUI and Forge-neo. A fork of regular sd-scripts is provided here, and a PR has also been created.I've been also listening to your valuable feedback, and where the model can be improved even further in the next iterations, and I will explain what this version is and isn't, and what it aims to be in the future.What v1 preview actually isRegarding this version, perhaps it's more appropriate to name it v0.1, I guess? The released version was trained using Muon optimizer, on approximately 2.5 Epochs, but the true number is closer to 5 due to extensive use dynamic repeats on both new and old characters, with the dataset focus on post September 2025.This version's main aim is to be more knowledgeable and be the most up-to-date anime model at release, v1 preview is not a "full" finetune (the whole original weights are frozen), and many pros and cons from the original Anima also carry over. By itself, v1 has its own strengths and weaknesses as well. It's not a model trained for aesthetic or with any RLHF. It's a model with a slight bias toward modern East-Asian anime style illustration, this is in fact intended.It's also soon and easy to realize that 1.7M is not a very large number of samples (it's a decision I have to make at the time based on time and money constraint), and the new expanded layers has a lot of room for much, much more information. In other words, a lot more samples are needed, and that's why there is a 10M (pretraining non-anime focus) samples floating around. Unfortunately it's not a cheap or fast task to accomplish, and for that reason, your supports are greatly appreciated. Even without actual monetary support, it can still be achieved, just won't be quick, nor reaching the full vision/potential that I had for the final model.Prompting guideFinally, if you're struggling to prompt your desired results, here are several very important points for consideration:- Characters should (think "must" in this case) be follow by their series/copyrights, (think of these like anchors, they always tag along) follow by their appearances (the more the better). Simple or very short prompt won't do as well.- Don't use underscore except for score tags.- Several metadata tags are very good to keep, I always recommend include highres and absurdres, following by the year tags (this has very strong influence on the generated image), score tags may not needed but you can still use them. You can throw away garbage such as "raytracing" or "4k" and "8k", these has never done anything and will just poison your output.- Always recommend using artist tags, same as Anima-base, and you can mix them as I often do with proper prompt weighting, but don't expect it to be the same as sdxl.- Prompt weighting and negative prompt are very important as well, this is something very easy to be underutilized.- Prompting the background is also important if you want it to be more dynamic. Additionally, use keywords such as "cinematic composition" and "dynamic angle" can improve your image significantly.Thank you once again, I will await your feedback.You may need to install ComfyUI-Anima-2.9B to the custom node folder. Plug and play, there is no workflow node needed. Sometimes may not work with other custom nodesOverviewAnima-2.9B is a fine-tune and layer-expansion of circlestone-labs/Anima. The base Anima model targets anime, illustration, and non-photorealistic art; this release continues training on that foundation with an expanded architecture. The model is trained on an additional 1.7M anime/illustration samples, with knowledge cutoff in July 2026, making Anima-2.9B one of the most up-to-date anime/illustration model at release.VersionsAnima-2.9B-preview-v1: initial releaseTraining/DatasetTrained using Muon optimizer on a 8x 5080s cluster, with earlier steps trained locally on my PCAs of preview v1, only new layers have been trained, with roughly 70% of the compute spent on 1024pxKnowledge cutoff in July 2026, training data included both new and old samples prior to September 2025Mixed captioning, including both tags and natural languages, using a mix of Gemini 3.1 Flash-Lite, Gemini 3.5 Flash-Lite, and Claude Sonnet 5NO scoringArchitectureTransformer depth expansion: expanded from 28 transformers layers to 40, growing the model to ~2.9B parameters. Each new layer is added by deep-copying its neighboring layer's weights, using interleaved insertion with zeroed-out output projections, making the new model functionally identical to Anima-base at initialization.Prompting tips :Follow Anima prompting tips: quality tags, year/period tags, @artist tags, character count (1girl, 1boy), character tags (follow Danbooru and Gelbooru tags), series/copyrights, base appearance.Character name/tags should be follow with series/copyrights tags or else the model might confuse.For multi-character images, attribute the character and names with their respective tags/appearance.The model does improve the base art style slightly, but I'd still recommend using artist tags.The dataset does not include score in its captions, however, you can still use them. Also HIGHLY recommend using year tag.(IMPORTANT) THE MORE DETAILED THE PROMPT, THE BETTER, short prompt will often generate a bland simple background, and may not able to produce the desire resultsGeneration (Recommendation)Sampler: Euler/Res-multistep/Er-sdeScheduler: sgm-uniform/beta/beta57/linear-quadraticResolution: 812x1216, 1152x1536, 1536x1536Steps: 28-50CFG: 3.5-5My personal usage are euler + sgm-uniform, which has a good balance between composition and fine details. Additionally res-multistep + linear-quadratic spend more time at high noise steps, which does lead to visibly better composition. My recommendation for the highest quality is 50 steps, CFG 3.5. You may need to use slightly lower CFG than regular Anima. Experiment yourself!LicenseModel weights are released under the CircleStone Labs Non-Commercial License, falling under derivative model category.Not affiliated with CircleStone LabsAcknowledgementsBuilt on nvidia/Cosmos-Predict2-2B-Text2Image and circlestone-labs/Anima.LLaMA Pro: Progressive LLaMA with Block Expansion.Training infrastructure built on sd-scripts.