Introduction
Anima is a 2 billion parameter text-to-image model created via a collaboration between CircleStone Labs and Comfy Org. It is trained with Cosmos-Predict2-2B-Text2Image as a base using mostly anime images with an additional large batch of non-anime artistic images.Currently only preview versions (Anima Preview 3 is now available) have been published, but it is already showing promising results. Anima Base model is now available. For more information about Anima, see the anima huggingface page and the discussion page therein. Also have a look at the anima checkpoint model pages on civitai and at these Anima related articles:
https://civitai.com/articles/26217/anima-what-is-anima
https://civitai.red/articles/25969/anima-model-comparison
https://civitai.red/articles/27001/anima-may-save-people-left-behind
https://civitai.red/articles/28921/prompting-for-tag-weights-in-anima-what-works-what-doesnt-and-why
If you want to train your own LoRA for anima, Citron made an easy trainer app for that:
https://civitai.red/articles/28641/anima-lora-trainer-app
If you want to generate images with Anima locally, you can choose between using ComfyUI or SD WebUI Forge Neo. This is a guide for setting up Forge Neo using Stability Matrix to generate images with Anima.
Forge Neo is a tool for generating images (like ComfyUI) using various models, but compared to comfy it is mostly considered easier to use.Unfortunately it does not officially support Linux, macOS, AMD and Intel (GPUs). On the github page the note regarding not supporting Linus, macOS, AMD and Intel has been removed and instead there is a note saying "Improved non-Nvidia support". However I haven't tested it, so this guide is mainly directed at Windows users with an Nvidia GPU (may work with other OS and GPU but no guaranty). It should work for any system with Windows 10/11 and RTX 30XX or newer GPU (it works for me with RTX 3080). Also older Nvidia GPUs should in principle be supported, but I am not sure if Stability Matrix will install the correct earlier pytorch version, which might be required.
I don't know if there really is a demand for this guide, but I was requested to do this, so hopefully it helps some people and maybe can be of use for future reference. This is the first such guide I am writing, so please bear with me. If you have any suggestions or questions be sure to leave a comment.
Installing Forge Neo using Stability Matrix
Stability Matrix is a multi-platform package manager for image generative ai tools. It basically serves as a wrapper application for image inference tools like ComfyUI, SD webUI or Forge and training tools like kohya_ss gui, OneTrainer or AI Toolkit. It also provides the possibility for downloading models from civitai via API key and other functions. As a wrapper for image generation tools it handles the dependencies for each tool (i.e. python packages, pytorch, cuda etc.), so you don’t have to worry about them (at least in most cases). You can find the download links for Stability Matrix on the github page. Here is the download link for the latest Windows 10/ 11 version:
https://github.com/LykosAI/StabilityMatrix/releases/latest/download/StabilityMatrix-win-x64.zip
It is also available for Linux and MacOS.
On Windows you have to extract the downloaded .exe file and run it. You have to choose a data directory, where packages are installed and other data is stored. You should choose a directory you can easily find again later.
Next you are prompted to choose a preferred interface, but you can skip this dialogue.
After skipping you should see the packages overview, which is empty, since you haven’t installed a package. If you are not in the packages overview, you can go there by clicking the package symbol at the top of the sidebar. To install a package, you have to click the "Add Package" button at the bottom.
You should now see the packages available for installation. To install Forge Neo click on the preview named "Stable Diffusion WebUI Forge – Neo".
Next you can change some settings for the installation, but the defaults should be fine in most cases. I had to change the display name/installation directory, since I already have an instance of Forge Neo installed. To start the installation you just have to hit the "Install" button.
Next you will see a dialogue window, showing all the dependencies downloaded and installed, which you need to run Forge Neo. The download and installation may take some time (for me it was rather quick since the dependencies already existed for the other Forge instances).
After installation is finished the dialgue window will close and you should see now the new installed Forge Neo package in the package overview. In my case there are other packages already installed, so the new package shows up on the bottom right of the package overview.
To start Forge Neo just click the launch button.
Next you’ll see a console interface showing the startup process. The configuration may take some time on the first startup.
After configuration is finished, the user interface will be opened in a browser window. This doesn’t mean you are connecting to a server, Forge Neo just uses Gradio, which runs in your browser. If the user interface does not open on itself, you have to click Open Web UI at the top.
Setting up Forge Neo for Anima
If you want to use Anima or any other model, you need to download the model and put it into the appropriate model folder of Stability Matrix. You can either download Anima-preview from civitai or from the huggingface page: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/diffusion_models
Other Anima based checkpoints can be downloaded on civitai from their model page.
In the data directory you have set at the beginning there is a "Models" folder. In the Models folder you have to place the downloaded Anima checkpoint into the folder called "StableDiffusion".
For using Anima with Forge Neo, you also have to download the text encoder and the VAE from the huggingface page.
Anima text encoder: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders
Anima VAE: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
Place the text encoder file in the folder "TextEncoders" located in the "Models" folder and the VAE file in the "VAE" folder.
After downloading the checkpoint, text encoder and VAE and placing them in the correct folder, you have all in place to use Anima with Forge Neo.
Go back to the Forge Neo user interface (your browser window) and hit the refresh button right next to the VAE / Text Encoder field at the top. Now you should be able to select the Anima checkpoint from the Checkpoint dropdown menu.
Next you have to individually select the text encoder and the VAE from the VAE / Text Encoder dropdown menu. The text encoder is named "qwen_3_06b_base" and the VAE is named "qwen_image_vae".
For using Anima you have to apply anima in the UI Preset in the top left corner. So now your settings at the top of the user interface should look like this:
Beneath the setting fields for checkpoint etc. you can see the fields for your prompt and negative prompt you want to use in your image generation. On the right there is a button to start the generation of your image.
Under the prompt field you have the panel for your actual generation settings. These are the default settings I am using for Anima.
You can change the image width and height of course, but it is recommended to not use a larger resolution than approximately 1 MP with anima-preview (preview3 can be used at slightly larger image size). You can also try using different Sampling Methods or Schedule Types. I made some images comparing different Sampling Methods, you can find the image post here, and different Schedule Types, you can find the images here. The CFG Scale should be 4-5 and the Sampling Steps 30-50. Setting the seed to -1 gives you a random seed. On the right of the settings panel is a field, where a preview of your image during generation and the final image will be shown.
You can use LoRas after you downloaded them and placed them in the "Lora" folder in your "Models" directory. After placing them into the "Lora" folder click the Lora tab under your negative prompt field and hit the refresh button at the top right of your Lora tab. To use a LoRa you have to click on the model card. This will add the lora activation line to your prompt.
The LoRa weight can be controlled by changing the number after the colon.
If you want to use danbooru tags in your prompt you should respect the tag order recommended on the anima huggingface page, that is:
[quality/meta/year/safety tags] [1girl/1boy/1other etc] [character] [series] [artist] [general tags]
You can also use natural language or a mix of natural language and tags. Be sure to check the other prompting suggestions on the huggingface page regarding quality tags, artist tags, etc. and other prompting tips.
Summary and Remarks
In this guide I provided a step by step advice for installing SD WebUI Forge Neo using Stability Matrix on Windows and a basic set-up for using an Anima based checkpoint with Forge Neo. There are a lot of other models you can use with Forge Neo and more elaborate options built in than covered in this guide.
If you migrate from another Forge or A1111 version, your previously used extensions might not work correctly or you need to find a fork designed for Forge Neo.
If you find any spelling mistakes or weird formulations, let me know, so I can correct them.
I hope you will be able to try using Anima following this guide.