Train a custom anime LoRA on the ANIMA base model

Citron Anima LoRA Trainer

Overview

Citron's Anima LoRA Trainer (app.py, titled "Citron's Anima LoRA Trainer" in the UI) is a local Gradio UI for training LoRA adapters on the Anima diffusion model using kohya-ss/sd-scripts. It trains on ~6GB VRAM with the default settings, the same low-VRAM profile as Anima generation.

• Created by Citron Legacy; UI repo: https://github.com/citronlegacy/citron-anima-lora-trainer-ui. The Aitrepreneur adaptive installers clone the fork https://github.com/aitrepreneur/citron-anima-lora-trainer-ui.
• Training backend: kohya-ss/sd-scripts (https://github.com/kohya-ss/sd-scripts), launched via accelerate launch.
• Trains LoRAs for Anima DiT (Cosmos-2B). Uses Anima's own components: DiT weights + Qwen3-0.6B text encoder + Qwen-Image VAE.
• Output: a standard .safetensors LoRA usable directly in the anima-base ComfyUI workflow.

The network module is networks.lora_anima and the training script is sd-scripts/anima_train_network.py (an Anima-specific kohya script the installer expects). Confirm these exist after the installer's git clone of sd-scripts. app.py references them, but they are pulled from the upstream repo at install time.

Setup

Windows

Run CITRON_ANIMA_LORA_TRAINER-V2.bat. It:

1. Ensures Git and Python 3.10 are present (via winget if missing).
2. Detects the NVIDIA GPU/driver and picks a matching PyTorch CUDA wheel automatically:
   • Blackwell (RTX 50xx) → cu128, bf16
   • Modern (RTX 20/30/40, etc.) → cu128/cu126/cu118 by driver, bf16 (fp16 on Turing)
   • Pascal/Maxwell (GTX 10/9xx) → cu126/cu118, fp16
   • Kepler/older → unsupported
3. Clones the UI repo, patches app.py defaults (base_model → anima-preview3-base, mixed_precision → detected value), writes app_configs/accelerate_gpu.yaml.
4. Creates .venv, installs PyTorch, clones and installs sd-scripts, installs app requirements.txt.
5. Downloads models into models/anima/{dit,text_encoder,vae}/ from https://huggingface.co/circlestone-labs/Anima/resolve/main/split_files/...:
   • dit/anima-base-v1.0.safetensors (~4GB)
   • text_encoder/qwen_3_06b_base.safetensors (~1.19GB)
   • vae/qwen_image_vae.safetensors (~254MB)
6. Writes and launches run_anima_base_windows.bat.

RunPod / Linux

Run CITRON_ANIMA_LORA_TRAINER-RUNPOD-V2.sh. Same flow into /workspace/citron-anima-lora-trainer-ui; it patches server_name to 0.0.0.0. Expose HTTP port 7860 and open Connect → HTTP Service 7860 (or https://${RUNPOD_POD_ID}-7860.proxy.runpod.net).

Launch

app.py runs Gradio on 0.0.0.0:7860, so open http://127.0.0.1:7860. Re-launch later with run_anima_base_windows.bat (Win) or ./run_anima_base_runpod.sh (RunPod). The DiT base model auto-downloads on the first "Start Training" if not already present (uses wget).

Dataset preparation

A flat folder of images, each with a matching .txt caption of the same basename (image-side captioning, kohya style):

my_dataset/
  001.png      001.txt
  002.jpg      002.txt
  ...

• Accepted images: .jpg .jpeg .png .webp .bmp .gif.
• Captions are Danbooru-style tags / natural language (same prompt style as Anima generation). The trainer warns about any image missing a .txt.
• caption_extension = .txt; shuffle_caption = false; caption_dropout_rate default 0.1 (set per dataset).

The UI tab "Training" takes Image Directory (the flat folder above) and Output Directory (where the LoRA is saved). "Configure Training" validates the dataset, prints a step estimate (steps_per_epoch = ceil(images × repeats / (batch × grad_accum)), total = spe × epochs), then writes two TOMLs into configs/.

Key training parameters (defaults from app.py)

Basic

Param	Default	Notes
project_name	my_lora	also the output_name of the LoRA
base_model	anima-base-v1.0	dropdown: anima-preview, anima-preview2, anima-preview3-base, anima-base-v1.0 (installer patches default to anima-preview3-base)
network_dim	32	LoRA rank
network_alpha	32
learning_rate	1e-4
max_train_epochs	10
resolution	768	px; dataset bucketing 256–4096, step 64
repeats	10	per-image repeats
caption_dropout	0.1

Advanced

Param	Default	Notes
optimizer_type	AdamW8bit	choices: AdamW8bit, AdamW, Lion, SGD, Prodigy; optimizer_args = ["weight_decay=0.1", "betas=[0.9, 0.99]"]
lr_scheduler	cosine_with_restarts	+ cosine, linear, constant, constant_with_warmup, polynomial
lr_scheduler_num_cycles	1
lr_warmup_steps	100
train_batch_size	1
gradient_accumulation_steps	1
max_grad_norm	1.0
save_every_n_epochs	1
save_last_n_epochs	4	keep last N checkpoints
mixed_precision	bf16	installer overrides to fp16 on older GPUs
gradient_checkpointing	true	memory saver
seed	42
noise_offset	0.03
multires_noise_discount	0.3
timestep_sampling	sigmoid	+ uniform, logit_normal
discrete_flow_shift	1.0	flow-matching shift
cache_latents	true
cache_text_encoder_outputs	true
vae_chunk_size	64
vae_disable_cache	true
num_cpu_threads_per_process	1

Fixed in the generated training TOML (not exposed): network_module = networks.lora_anima, network_train_unet_only = true, qwen3_max_token_length = 512, t5_max_token_length = 512, save_model_as = safetensors, save_precision = bf16 (fp16 on older GPUs).

Generated config files

configs/<project>_training_<timestamp>.toml references the DiT (pretrained_model_name_or_path), qwen3 text encoder, and vae paths from models/anima/, plus all params above.

configs/<project>_dataset_<timestamp>.toml:

[general]
resolution = 768
enable_bucket = true
bucket_no_upscale = false
bucket_reso_steps = 64
min_bucket_reso = 256
max_bucket_reso = 4096

[[datasets]]
resolution = 768
[[datasets.subsets]]
num_repeats = 10
image_dir = "/path/to/my_dataset"
caption_extension = ".txt"
caption_dropout_rate = 0.1

The sd-scripts command

"Start Training" runs the following and streams logs live to the UI and to logs/<project>_<timestamp>.log:

accelerate launch \
  --config_file app_configs/accelerate_gpu.yaml \
  --num_cpu_threads_per_process 1 \
  --gpu_ids 0 \
  sd-scripts/anima_train_network.py \
  --config_file  configs/<project>_training_<timestamp>.toml \
  --dataset_config configs/<project>_dataset_<timestamp>.toml

accelerate_gpu.yaml pins use_cpu: false, mixed_precision: <bf16|fp16>, single process/machine. CUDA_VISIBLE_DEVICES is set to the selected GPU index.

Output & using the LoRA

• The trained LoRA is saved to your Output Directory as <project_name>.safetensors, plus per-epoch checkpoints (the last save_last_n_epochs are kept).
• Copy it into ComfyUI models/loras/ and load it in the anima-base workflow via LoraLoaderModelOnly (or rgthree Power Lora Loader):

  { "class_type": "LoraLoaderModelOnly",
    "inputs": { "model": ["<unet>", 0], "lora_name": "<project_name>.safetensors", "strength_model": 1.0 } }
  

• Use the same prompt style you captioned with. Typical strength 0.7 to 1.0; stack with the turbo LoRA for fast 12-step generation.

VRAM & tips

• Defaults train on ~6GB VRAM (network_dim 32, res 768, batch 1, gradient checkpointing + latent/TE caching).
• On OOM, the trainer suggests network_dim=8 and/or resolution=512. Also keep batch 1 and use AdamW8bit.
• A GTX 1060 6GB works but is slow; 3GB cards are not realistic. GPUs older than Pascal are unsupported.
• Step count rule of thumb: images × repeats × epochs / (batch × grad_accum). The UI prints the exact estimate before you train.
• Logs stream to the UI and logs/. Training config and last paths persist in config.json so you can re-run.

Unverified / verify before relying

• sd-scripts/anima_train_network.py and networks.lora_anima come from the kohya fork pulled at install time. app.py expects them, but they are not in the local downloaded files here.
• The exact LoRA output filename is <project_name>.safetensors per output_name; confirm in your Output Directory after a run.

Sources

• Official: https://github.com/citronlegacy/citron-anima-lora-trainer-ui and https://github.com/kohya-ss/sd-scripts
• Empirical: installer/launch notes from the pack scripts; training defaults from the UI's app.py as observed.