Anima Lora Trainer

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