A solo developer running an 8× RTX 5080 cluster just published Anima-2.9B, and within three days it was the top trending text-to-image model on Hugging Face with nearly 17,000 downloads. It is a layer-expanded fine-tune of CircleStone Labs’ Anima, the 2B anime and illustration model built with Comfy Org. The weights are free, and the smaller build is a 3.08 GB file.
If you draw, storyboard or make visual novels and want a local model with no subscription attached, this is worth an afternoon. If you were planning to wrap it in a paid product, skip to the license section — there is a hard line there, and it is not the one most people assume.
Anima-2.9B at a Glance
- What it is: a 2.9B-parameter text-to-image model for anime, illustration and non-photorealistic art, expanded from 28 to 40 transformer layers on top of Anima-base.
- What’s free: everything. Open weights on Hugging Face, a ComfyUI loader node, and a LoRA trainer with a web GUI. No account, no credits, no API key.
- What it costs: disk and VRAM. 5.84 GB for the bf16 file, 3.08 GB for the int8 build, plus 1.44 GB of shared text encoder and VAE.
- The catch: the weights are non-commercial. Your generated images are not — that distinction matters and we break it down below.
- Status: labelled preview v1, training still in progress. Only the newly inserted layers have been trained so far.
What Actually Changed From Anima-base
Base Anima is 2B parameters, built on NVIDIA’s Cosmos-Predict2-2B-Text2Image, trained on several million anime images plus ~800k non-anime artistic images, with an anime knowledge cutoff of September 2025. It has 181,000 downloads and 2.2 million on-site generations on CivitAI.
Anima-2.9B does two things to that foundation. First, depth expansion: twelve new transformer layers inserted by deep-copying neighbouring weights with zeroed-out output projections, a trick from the LLaMA Pro block-expansion paper, so the bigger model starts functionally identical to the smaller one. Second, 1.7M additional anime and illustration samples with a July 2026 knowledge cutoff — ten months newer than base. That is the practical headline: newer characters, series and artist tags.
Training ran on the Muon optimizer across eight RTX 5080s, with about 70% of compute spent at 1024px. Captions mix Danbooru-style tags and natural language. The new data carries no score tags, though the model still understands them from base training.
The File List, and What You Actually Need to Download
This is where most people get stuck, because the Anima-2.9B repository contains only the diffusion model. The text encoder and VAE live in the base repo. You need three files:
Anima-2.9B-preview-v1.safetensors— 5.84 GB (bf16), orAnima-2.9B-preview-v1_int8_convrot.safetensorsat 3.08 GB →ComfyUI/models/diffusion_modelsqwen_3_06b_base.safetensors— 1.19 GB, from the base Anima repo →ComfyUI/models/text_encodersqwen_image_vae.safetensors— 0.25 GB, same repo →ComfyUI/models/vae(this is the standard Qwen-Image VAE; you may already have it)
The model card publishes no VRAM figure, so here is the arithmetic rather than a claim: the int8 route puts about 4.5 GB of weights on disk in total. With ComfyUI’s default offloading of the text encoder, an 8 GB card is a realistic floor at 1024px, and 12 GB is comfortable at the 1536² ceiling. That is our estimate from the file sizes, not a number the author has published or that anyone has independently benchmarked. If you are running a 6 GB card, expect to drop resolution before you expect it to work.
How Anima-2.9B Compares
| Model | Diffusion file size | Free weights? | License | Best at |
|---|---|---|---|---|
| Anima-2.9B (preview v1) | 5.84 GB bf16 / 3.08 GB int8 | Yes | CircleStone non-commercial + NVIDIA Open Model | Anime & illustration, mid-2026 character knowledge |
| Anima-Aesthetic v1.1 (base family) | 4.18 GB | Yes | Same | Cleaner default style, most stable out of the box |
| Anima-Turbo v1.0 | 4.18 GB | Yes | Same | Fast iteration at CFG 1, 8–12 steps |
| Krea 2 / Kroma (12B) | ~24 GB class | Yes | Open, commercial-friendly | General-purpose realism and photography |
If you want photoreal output, the Anima family is the wrong tool — the base card says plainly it will not work well at realism. Read our breakdown of Kroma and Krea 2 instead. Anima competes on being small, fast, current on anime concepts, and trainable on a single consumer GPU.
Try It Free: The Exact Setup
Path 1 — no GPU, no install. The base Anima versions are hosted free on CivitAI and TensorArt, both with free daily generation credits. To be clear: these host the 2B base family, not the 2.9B expansion. It is the fastest way to decide whether the art style suits you before downloading 5 GB.
Path 2 — ComfyUI locally. Anima is natively supported in ComfyUI. Update to 0.33.1 or later and the 2.9B file loads with no extra work. If you are on an older build, clone the loader into your custom nodes folder:
cd ComfyUI/custom_nodes
git clone https://github.com/gazingstars123/ComfyUI-Anima-2.9BIt patches the ComfyUI model loader at start-up — no workflow node to add. The repo README warns it can conflict with other custom node packs, so test on a clean profile. The base Anima card ships a drag-and-drop example workflow image you can reuse unchanged.
Path 3 — train your own LoRA. The Anima Standalone Trainer is a decoupled sd-scripts build with a web GUI, and it explicitly supports 2.9B LoRA training. You need Python 3.12, Node.js and CUDA 12.7+:
git clone https://github.com/gazingstars123/Anima-Standalone-Trainer.git
cd Anima-Standalone-Trainer
./setup_env.sh # or .\setup_env.bat on Windows
./training-ui/start_linux.shOpen http://localhost:3000, click the gear icon, and set three paths: DiT model, VAE and text encoder — the same three files as above. There is also an sd-scripts fork for command-line training.
Settings That Actually Work
- Sampler + scheduler: the author’s default is
euler + sgm_uniform.res_multistep + linear-quadraticspends more time at high noise and visibly improves composition. - Steps and CFG: 28–50 steps, CFG 3.5–5. No universal winner between 3.5 and 5 — run both on the same seed.
- Resolution: the card lists 812×1216, 1152×1536 and 1536×1536, flagging square 1536 as “iffy.”
- Prompt length is not optional. Short prompts produce bland, empty backgrounds.
- Artist tags need the @ prefix. Write
@artistname; without it the effect is very weak. Tag order: quality → character count → character → series → artist → general.
The Honest Trade-Offs
It is a preview, and the author says so. The card reads “Status: Training in progress,” and as of preview v1 only the newly added layers have been trained. Pretraining on 10M general samples is listed as the next step; prompt understanding outside anime subjects is the known weak spot.
No benchmarks exist. No leaderboard entry, no independent evaluation, no published side-by-side. Everything here comes from the model card and the file listings. The base repo does ship an anima_comparison.json ComfyUI workflow that generates a model-vs-model grid — the closest thing to an objective test you can run yourself.
The license is genuinely restrictive on weights — and genuinely permissive on outputs. The CircleStone Labs Non-Commercial License forbids hosting the model behind a paid API, putting it on a paid generation platform, or embedding the weights in a monetized game. But the restriction applies to the model, not the images: selling artwork, paid commissions, and using generations as concept art or assets in a commercial game are all explicitly allowed. Because Anima derives from Cosmos-Predict2, the NVIDIA Open Model License also applies.
Weak text rendering, and a safety-tag caveat. Single words and short phrases render acceptably; longer text does not. Short or vague prompts can also produce content you did not ask for — put safe in your positive prompt and the corresponding tags in your negative.
Five Workflow Changes Worth Making
- Download the int8 build first (3.08 GB). If quality holds for your style, you have saved 2.76 GB and real VRAM headroom.
- Prototype on Anima-Turbo at CFG 1 and 8–12 steps, then re-run the winning seed through 2.9B at 40 steps. Iterating at 50 steps is wasted time.
- Do not train the LLM adapter. Set
llm_adapter_lr=0. It holds a surprising amount of knowledge and degrades easily. Start a rank-32 LoRA at 2e-5. - Build a prompt template, not prompts. Lock quality, year, safety and artist tags as a prefix; vary only the subject block. It makes A/B comparison meaningful.
- Keep the base repo installed alongside. Shared text encoder and VAE let you hot-swap Base, Aesthetic, Turbo and 2.9B with one node change.
FAQ
Is Anima-2.9B free?
Yes. The weights, the ComfyUI loader and the LoRA trainer are all free downloads with no account required. The only cost is hardware, or free credits if you use a hosted platform for the base versions.
Can I use Anima-2.9B commercially?
You can sell the images. You cannot commercialise the model itself. Paid commissions, selling prints, and shipping generated art inside a paid game or visual novel are permitted. Hosting the weights behind a paid API or bundling them into a monetized product is not, without a separate license from CircleStone Labs.
What are Anima-2.9B’s system requirements?
ComfyUI 0.33.1 or newer, plus roughly 4.5 GB of weights for the int8 route (3.08 GB model, 1.19 GB text encoder, 0.25 GB VAE) or 7.3 GB for bf16. No official VRAM figure has been published; on our reading of the file sizes, 8 GB is a workable floor at 1024px. LoRA training needs Python 3.12, Node.js and CUDA 12.7+.
Is Anima-2.9B better than the base Anima model?
It is newer, not conclusively better. The July 2026 cutoff means it knows characters and artists the September 2025 base does not. But it is a preview with partial training, and the author calls the style improvement over base slight. Run the comparison workflow yourself.
Does Anima-2.9B work outside ComfyUI?
The author states it is officially supported in ComfyUI and Forge-Neo. We verified the ComfyUI path against the loader repository; the Forge-Neo claim is the author’s and we have not tested it. There is no Diffusers pipeline listed.
Worth Your Afternoon?
If you make anime or illustration work, yes — a 3 GB file with 2026 character knowledge and a working LoRA trainer is a lot of capability for a free download. If you need realism or commercial model hosting, no. It is the wrong model, and the license says so.
For more on where small open image models fit next to the hosted giants, see our comparison of Qwen-Image-Flash on a 12GB GPU and our look at LTX-2.5 and the real VRAM bill for open video. And if you want to see how ten text-to-image models stack up head to head, watch MAI Image 2.5 Pro vs 10 Other Text To Image AI Models on the Creative AI Show channel.
Accurate as of August 15, 2026 — model cards, file sizes and licensing terms move fast, and Anima-2.9B is explicitly still in training.
No responses yet