Anima LoRA Training: Settings, Samplers, and Local Setup
How to configure Anima for image generation, pick the right sampler, and train a LoRA locally with verified settings and hyperparameters.
On this page
Anima LoRA Training: Settings, Samplers, and Local Setup
TL;DR
- Anima is a 2 billion parameter text-to-image model by CircleStone Labs and Comfy Org, built on NVIDIA Cosmos-Predict2, and works inside ComfyUI with three separate model files.
- For generation, use 30-50 steps at CFG 4-5 with a resolution between 512² and 1536²;
er_sdeis a solid default sampler. - Train LoRAs on Anima-Base using kohya-ss/sd-scripts or a local GUI trainer, starting with a learning rate of 2e-5 for rank 32, and never train the LLM adapter.
What Anima Is
Anima is a text-to-image diffusion model focused on anime concepts, characters, and styles. The model uses the CircleStone Labs Non-Commercial License for non-commercial use; generated images can be used commercially. Anima does not perform well at realism; this is by design.
The model ships as three separate files for ComfyUI: the diffusion model in models/diffusion_models, a Qwen3 0.6B text encoder in models/text_encoders, and a Qwen-Image VAE in models/vae. The Qwen-Image VAE is 16-channel and handles image decoding. Anima is a derivative of NVIDIA's Cosmos-Predict2-2B-Text2Image and is subject to the NVIDIA Open Model License Agreement for Derivative Models.
The training dataset covers several million anime images and about 800k non-anime artistic images from sources including Danbooru, Gelbooru, DeviantArt, and LAION-POP, with no synthetic data used. The anime training data cutoff is September 2025.
Generation Settings That Work
Anima operates at resolutions between 512² and 1536² pixels. The recommended range is 30-50 denoising steps with a CFG scale of 4-5. Quality tags boost results: use a positive prefix of masterpiece, best quality, score_7, safe, and a negative prompt of worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration.
The model was trained on Danbooru-style tags, natural language captions, and combinations of both. When using tags, place them in this order: quality and meta tags first, then subject tags (1girl/1boy), character, series, artist, and general tags. Artist tags must be prefixed with @, without which the effect is very weak. The model was trained with random tag dropout, so you do not need to include every relevant tag at inference.
| Setting | Value | Source |
|---|---|---|
| Resolution range | 512² to 1536² pixels | HuggingFace model card |
| Steps | 30-50 | HuggingFace model card |
| CFG scale | 4-5 | HuggingFace model card |
| Positive prefix | "masterpiece, best quality, score_7, safe, " | HuggingFace model card |
| Negative prefix | "worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration" | HuggingFace model card |
| Model files | diffusion model + text encoder + VAE | HuggingFace model card |
Samplers and Schedulers
A beta57 scheduler (available through the ComfyUI RES4LYF custom node pack) can improve texture quality for a realistic or painterly look. The behavior differences between KL Optimal, FlowMatchEulerDiscrete, Beta, and SGM Uniform schedulers are not documented in official sources. For broader sampler guidance, see samplers and schedulers guide.
| Sampler | Style | When to use it | Source |
|---|---|---|---|
er_sde | Neutral, flat colors, sharp lines | Reasonable default | HuggingFace model card |
euler_a | Softer, thinner lines, 2.5D feel | When CFG can be pushed higher | HuggingFace model card |
dpmpp_2m_sde_gpu | More variety, creative | When you want creative freedom | HuggingFace model card |
euler | Basic, slightly creative | For Turbo and Aesthetic versions | HuggingFace model card |
| beta57 (scheduler) | Better textures, painterly | Via RES4LYF custom node pack | HuggingFace model card |
Choosing Between Anima Versions
Anima comes in three releases. Anima-Base is the pretrained, unrefined base model and offers maximum flexibility, diversity, and style adherence. LoRAs should be trained using this version. Anima-Aesthetic is fine-tuned for better consistency and a higher quality default art style, with quality tags stripped from captions — you do not need quality tags in the positive prompt when using it, though masterpiece, best quality, is safe to keep. Anima-Turbo is a distilled version for fast generation at CFG 1 with 8-12 steps; it is slightly worse than Aesthetic on average but much faster and cheaper, and the distillation process also increases stability.
| Version | Best for | Strengths | Weaknesses | Source |
|---|---|---|---|---|
| Anima-Base | LoRA training, maximum flexibility | Largest creative range, unrefined base | Plain default style without artist or quality tags | HuggingFace model card |
| Anima-Aesthetic | High-quality default outputs | Consistent, high quality out of the box | Reduced diversity from aesthetic tuning | HuggingFace model card |
| Anima-Turbo | Fast iteration, lower cost | 8-12 steps at CFG 1, fast and cheap | Slightly lower quality than Aesthetic on average | HuggingFace model card |
Training a LoRA Locally
Training a LoRA adapter for Anima requires a local setup or a cloud notebook. Two main paths exist: a standalone trainer with a graphical interface, or a Google Colab notebook. For a general introduction to LoRA fine-tuning concepts, see our LoRA basics guide. For ComfyUI setup, see ComfyUI installation guide.
Local GUI Trainer
The CitronLegacy Anima LoRA Trainer UI provides a local graphical interface for training. It requires Python 3.10+, PyTorch 2.6+, and CUDA 12.4+ (CUDA 11.8 may work but is not guaranteed). The trainer runs on less than 6 GB VRAM with default settings (dimension 20, resolution 768).
Colab Notebook
The CitronLegacy Colab trainer is a Google Colab notebook that automates the pipeline. It installs sd-scripts, downloads the three Anima model files (~5.6 GB total), generates TOML configuration files, and launches training via accelerate. It includes a step estimator, live streaming output, and Google Drive integration so outputs survive session resets. On a free-tier T4 GPU, 1000 training steps takes over 4 hours.
Training Best Practices
Start with a learning rate of 2e-5 for rank 32 and adjust from there. The Anima base model has no aggressive aesthetic tuning or RLHF to overcome, and it contains a vast amount of visual knowledge, so a light training touch is all that is needed.
Do not train the LLM adapter. The adapter processes text embeddings before they reach the diffusion model and therefore has an outsized influence on generated images. It contains a surprising amount of knowledge and degrades easily when trained. Set llm_adapter_lr to 0 in your training script; most trainers including sd-scripts support this option.
| Hyperparameter | Default | Notes | Source |
|---|---|---|---|
network_dim | 20 | LoRA rank; lower values (e.g., 8) use less VRAM | CitronLegacy Colab README |
network_alpha | 20 | Usually equal to network_dim | CitronLegacy Colab README |
learning_rate | 0.0001 | Reduce if you see NaN loss | CitronLegacy Colab README |
max_train_epochs | 10 | Keep low to stay under 1000 total steps | CitronLegacy Colab README |
resolution | 768 | Reduce to 512 for out-of-memory errors | CitronLegacy Colab README |
repeats | 5 | Images × repeats = steps per epoch | CitronLegacy Colab README |
caption_dropout | 0.1 | Probability of dropping a caption during training | CitronLegacy Colab README |
For the Anima Standalone Trainer, the requirements are Python 3.10+, CUDA 12.1+, and it runs on Windows and Linux.
Common Errors and Fixes
| Symptom | Cause | Fix | Source |
|---|---|---|---|
| CUDA out of memory | Default settings exceed available VRAM | Reduce network_dim to 8 and/or resolution to 512 | CitronLegacy Colab README |
| NaN loss | Learning rate too high | Lower learning_rate; ensure PyTorch ≥ 2.5 | CitronLegacy Colab README |
| "No images found" | Caption file format or filename mismatch | Ensure caption files end in .txt and image filenames do not | CitronLegacy Colab README |
| Colab disconnects mid-run | Session timeout exceeds runtime limit | Keep total steps below 1000; use Google Drive for checkpoints | CitronLegacy Colab README |
| Models missing on re-run | /content/ resets each Colab session | Re-run the Setup cell, or mount Google Drive | CitronLegacy Colab README |
| LoRA has weak artist style effect | Missing @ prefix on artist tag in prompt | Always prefix artist tags with @ | HuggingFace model card |
| LoRA degrades model output | LLM adapter was trained alongside the LoRA | Set llm_adapter_lr to 0 to disable training the adapter | HuggingFace model card |
FAQ
How many images do I need to train an Anima LoRA? There is no single answer. Community trainers have reported usable results with as few as a handful of images for character LoRAs, while larger datasets of 30-100 images provide more variety. Focus on quality over quantity.
What learning rate should I use for Anima LoRA training? The recommended starting point is 2e-5 for a rank 32 LoRA. Adjust from there based on results. Keep learning_rate at 0.0001 (1e-4) for most sd-scripts-based pipelines as a default.
Why shouldn't I train the LLM adapter on Anima? The LLM adapter processes text embeddings before they reach the diffusion model and therefore has an outsized influence on generated images. It degrades easily when trained alongside a LoRA. Set llm_adapter_lr to 0 to disable training it entirely.
What resolution does Anima support? Anima works at resolutions between 512² and 1536² pixels. The default training resolution in the Colab trainer is 768, which can be reduced to 512 if you encounter out-of-memory errors.
How do I format prompts for Anima? Anima accepts Danbooru-style tags, natural language captions, or a combination of both. Use lowercase for tags with spaces instead of underscores (score_* tags use underscores). Prefix artist names with @. Order tags as: quality/meta/security, subject, character, series, artist, general tags.
What is the VRAM requirement for training Anima LoRA locally? The CitronLegacy UI trainer runs on less than 6 GB VRAM with default settings (dim=20, res=768). Reduce to dim=8 and/or res=512 to lower VRAM usage further. The Anima Standalone Trainer requires CUDA 12.1+.