How to Install Anima-2.9B in ComfyUI for Anime Images

You can now install Anima-2.9B in ComfyUI 0.33.1 or newer without a custom compatibility node. The clean setup is to download the 5.84 GB diffusion model, reuse Anima’s Qwen-3 0.6B text encoder and Qwen-Image VAE, then adapt ComfyUI’s official Anima Base v1 workflow by selecting the 2.9B diffusion file.

This is a larger community derivative of CircleStone Labs’ 2B Anima model. It targets anime and non-photorealistic illustration, not realistic photography. It is also a preview under a non-commercial model licence, so read that restriction before building a paid service around it.

What changed in Anima-2.9B?

The original Anima diffusion model has 28 transformer layers and about 2 billion parameters. The Anima-2.9B author expanded that structure to 40 layers, then trained the additional layers on 1.7 million more anime and illustration samples. The model card gives a July 2026 knowledge cut-off.

That extra depth originally needed a custom ComfyUI node. ComfyUI 0.33.1, released on 13 August 2026, added native support for Anima fine-tunes with extra blocks. Updating ComfyUI removes the need for the creator’s compatibility node on updated installs.

Anima-2.9B is not a new official CircleStone checkpoint. It is a community fine-tune and architecture expansion that uses CircleStone’s licence. That distinction matters when you assess support, stability and commercial use.

What you need

• ComfyUI 0.33.1 or newer.
• About 7.3 GB for the three model files, plus generated images and temporary download space.
• A GPU supported by your current ComfyUI installation.
• The official Anima Base v1 workflow from ComfyUI’s Template Library, used as a starting template for the 2.9B diffusion file.

The 5.84 GB download size is not a minimum-VRAM figure. Inference also needs memory for the text encoder, VAE, latent tensors and ComfyUI itself. The model author does not publish a minimum GPU requirement, so avoid a guaranteed “runs on 8 GB” claim.

The repository also contains a roughly 3.08 GB Anima-2.9B-preview-v1_int8_convrot.safetensors file. Its model card does not document a complete quality, speed or compatibility comparison, so this guide starts with the full checkpoint. Treat the smaller file as an experimental alternative until its trade-offs are better documented.

Step 1: update ComfyUI to 0.33.1 or newer

Check the ComfyUI engine version in its startup console or About information. Version 0.33.1 is the minimum for native Anima-2.9B support, not a recommendation to stay on that older release. Save your working workflow before updating through your installation’s supported route.

• Windows Portable: use update/update_comfyui_stable.bat for the stable release. ComfyUI’s update guide distinguishes this from the development updater and the more disruptive Python-dependency reinstall script.
• Desktop: open the Manage panel and its Update tab to update the ComfyUI engine. Updating the desktop application alone is a separate operation.
• Manual installation: follow the official update guide for your existing ComfyUI Python environment. Updating code without its matching dependencies can leave the template library or frontend behind.

Restart ComfyUI after the update and confirm the version before adding the model. The relevant release note is “Support anima tunes with extra blocks.” If your installation cannot be updated, the creator’s current ComfyUI-Anima-2.9B compatibility node is the documented fallback; use only that official repository and avoid mirrors.

Step 2: download the three model components

Download files only from the two official Hugging Face repositories below.

If you already installed the original Anima workflow, you may already have both shared files. A Qwen-Image workflow may already provide qwen_image_vae.safetensors, but it generally uses a different text encoder. Compare the exact file names and keep a single valid copy rather than downloading duplicates.

Check a suspicious or interrupted download

The full Preview v1 checkpoint is 5,843,204,206 bytes in the creator’s file manifest, checked on 5 September 2026. If ComfyUI cannot read it, compare your file’s SHA-256 with the hash below. From the ComfyUI folder, Windows PowerShell users can run:

Get-FileHash -Algorithm SHA256 '.\models\diffusion_models\Anima-2.9B-preview-v1.safetensors'

Expected SHA-256 for this specific full checkpoint:

0b3020d1b906155f7eb30667622723e87160632c8c7a5f1c93bdce685f2a346d

On macOS use shasum -a 256, or on Linux use sha256sum, followed by the file path. A different hash needs investigation; do not rename an int8 file to make it look like the full checkpoint.

Step 3: place each file in the correct folder

Your model directory should contain this layout:

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── Anima-2.9B-preview-v1.safetensors
    ├── text_encoders/
    │   └── qwen_3_06b_base.safetensors
    └── vae/
        └── qwen_image_vae.safetensors

Do not put the diffusion model in models/checkpoints. The official Anima workflow loads the separate diffusion, encoder and VAE components through standard ComfyUI nodes.

After copying the files, restart ComfyUI so its model lists are refreshed.

Step 4: adapt the official Anima Base v1 workflow

1. Open ComfyUI’s Template Library.
2. Search for Anima Base v1.
3. Open the text-to-image workflow.
4. Open its subgraph if the model selectors are not visible on the main canvas.
5. In the diffusion-model loader, select Anima-2.9B-preview-v1.safetensors.
6. Confirm the text encoder is qwen_3_06b_base.safetensors and the VAE is qwen_image_vae.safetensors.
7. If the workflow exposes a turbo_mode control, keep it turned off for Preview v1.

Comfy Org’s workflow uses the standard UNETLoader, CLIPLoader and VAELoader path. Starting from that template removes a lot of guesswork and makes custom-node conflicts easier to isolate.

Check the controls inside the subgraph before the first run. The current official template JSON contains these defaults:

• Model: the loader starts with the original 2B anima-base-v1.0.safetensors. Selecting the template does not select Anima-2.9B for you.
• Scheduler: KSampler starts with simple. Change it to sgm_uniform to follow the first-run settings below. Leave denoise at 1 for this text-to-image workflow.
• Resolution and sampling: the outer ResolutionSelector starts square. Check the dimensions actually feeding the latent node. If steps or CFG are connected to another control, edit the connected value rather than an inactive sampler field. Keep turbo_mode off.

Step 5: use a conservative first generation

Start with settings drawn from the creator’s documented ranges and sample workflows:

• Resolution: 832 × 1216
• Sampler: Euler
• Scheduler: SGM Uniform
• Steps: 30
• CFG: 4
• Batch size: 1

The model card recommends 28–50 steps, CFG 3.5–5 and several portrait or square resolutions. Do not begin at 1536 × 1536 if memory is uncertain; even the author labels that size “iffy.”

Use a detailed prompt that states the subject, pose, clothing, setting, lighting and composition. Anima understands tag-style prompts and natural language. Descriptive visual language is more reproducible than asking it to imitate a named living artist.

safe, 1girl, solo. An original anime character with short auburn hair and a navy raincoat stands at a tram stop at dusk. Wet pavement reflects warm shop lights, medium shot, balanced composition, clean line work, detailed background.

Confirm the setup before adding anything else

1. Save the workflow under a name such as anima-2.9b-first-run, with the model filename, prompt and settings selected.
2. Queue one image. Confirm that the queue finishes without an error and that the SaveImage node produces a file you can open. Check its dimensions and content; a successful download alone does not confirm generation works.
3. Record the run. Keep the ComfyUI engine version, model filename, GPU/platform, seed, dimensions, steps, CFG and any displayed timing or memory reading with your notes. If a measurement is unavailable, leave it blank.
4. Fix the seed for comparisons. The current template starts with seed behaviour set to randomize. Change that to fixed before comparing prompts or settings. Keep the workflow, models and software unchanged as well; a seed alone does not guarantee identical results across different setups.

Separate the first model-loading run from later runs when comparing speed. Change one setting at a time and save a new workflow copy for experiments. Once this baseline works, the Comfy MCP setup guide covers controlling the workflow from an assistant.

Step 6: compare settings without losing the baseline

Once the first image works, duplicate the workflow or save it under a new name. The model author suggests two useful combinations:

• Euler + SGM Uniform: a balanced starting point for composition and detail.
• Res-Multistep + Linear Quadratic: spends more time at high-noise steps and may improve composition, but takes more experimentation.

Keep the seed and prompt fixed while comparing. Creator-provided sample images are useful for discovering settings; evaluate quality with your own representative prompts and fixed comparison settings.

Common Anima-2.9B problems

Anima Base v1 is missing, or the template shows missing core nodes

Check the engine and dependency update together. ComfyUI distributes its frontend and workflow templates as separate packages, so a code-only update can leave an old library behind. Follow the installation-specific update instructions, restart, then reopen the template. Do not install unrelated custom-node packs to replace missing core nodes.

The model does not appear in the loader

Confirm that the full file finished downloading and is inside models/diffusion_models, not models/checkpoints. Restart ComfyUI after moving it. A partially downloaded file may have a browser-specific temporary extension and should not be loaded.

ComfyUI reports unsupported blocks or tensor shapes

Recheck the ComfyUI version. Native support for the expanded Anima blocks arrived in 0.33.1. Update and restart before changing the workflow. The model card mentions the creator’s compatibility node as a fallback, but it should not be necessary on a current installation.

The output is blank, noisy or fails in an old workflow

Load the official Anima template in a clean canvas and select the three model files again. The model author warns that some other custom nodes may not work with the expanded model. If the clean official workflow succeeds, reintroduce optional nodes one at a time.

The GPU runs out of memory

Reduce resolution, keep batch size at one, close other GPU-heavy applications and restart ComfyUI. Do not assume the checkpoint’s disk size equals its runtime memory requirement. The smaller int8 file exists, but its official documentation does not yet quantify the quality or compatibility trade-off.

Images look flat or the background is empty

Add concrete details about composition, environment and lighting. The creator says short prompts often produce simple backgrounds. Try 30–40 steps and compare CFG 3.5, 4 and 5 with the same seed rather than pushing CFG far outside the documented range.

Realistic faces or text look poor

That is a model limitation, not necessarily an installation error. The base Anima documentation says realism and lengthy text rendering are out of scope. Use a model designed for photography or add text later in an image editor.

Anima-2.9B FAQ

Do I need a custom node?

No, not with ComfyUI 0.33.1 or newer. That release added native support for Anima fine-tunes with extra blocks.

Is Anima-2.9B the official Anima successor?

No. It is a community fine-tune and layer expansion of CircleStone Labs’ Anima model. ComfyUI officially supports its expanded architecture, but that does not make the checkpoint a CircleStone release.

Can I use Anima-2.9B commercially?

The model card places the derivative weights under the CircleStone Labs Non-Commercial License. Commercial or production use is generally restricted apart from the licence’s narrow exceptions; paid hosting, API access and embedding the weights in a monetised product require a separate licence. The licence separately says generated outputs may be used commercially subject to its restrictions and notes inherited NVIDIA Open Model License terms where applicable. Review the current terms and obtain appropriate advice for your use case; do not infer permission from the availability of a public download.

Does it support image-to-image?

The current model card labels the release text-to-image, and this guide covers only that verified path. Do not advertise image-to-image, ControlNet or video support without a current compatible workflow and primary documentation.

Will it run on an 8 GB GPU?

There is no official minimum-VRAM figure for Preview v1. Resolution, attention implementation, encoder loading and other processes all affect memory, so test conservatively rather than treating 8 GB as guaranteed.

Can Comfy MCP control this workflow?

Once the plain workflow works reliably, you can automate it with my guide to control local ComfyUI with Comfy MCP. Keep the first test manual so an MCP configuration problem cannot be mistaken for a model problem.

Should you install Anima-2.9B?

Try Anima-2.9B if you already use ComfyUI, want a current anime and illustration model, and accept a preview checkpoint under a non-commercial licence. Update to ComfyUI 0.33.1, adapt the official Anima Base v1 template, keep turbo_mode off and begin with the full 5.84 GB model at 832 × 1216.

Stay with the official 2B Anima releases if you want more established documentation, smaller files or clearer support boundaries. Use a different model for photorealism, dependable text rendering or commercial model hosting. If you are comparing other local visual workflows, the MiniMax H3 ComfyUI guide covers video generation with native audio.

Primary sources

File manifest, engine-update instructions and official template controls rechecked on 5 September 2026.

• Anima-2.9B Preview v1 model card and files
• ComfyUI 0.33.1 release notes
• ComfyUI expanded Anima-block support change
• Comfy Org Anima workflow documentation
• Original Anima model card, shared files and limitations