AnimeGen - Lightweight Latent Diffusion Text-to-Image Model

Compact text-to-image model trained from scratch on 65,675 real anime images with Danbooru tags. Uses latent diffusion + DDIM + classifier-free guidance (CFG).

Architecture

1. Autoencoder - 96x96 -> 64-channel latent. Reconstruction PSNR 27.8 dB.
2. GRU tag encoder - comma-separated Danbooru tags -> conditioning vector.
3. Conditional Latent Diffusion U-Net (base=256) + CFG at sampling.

Key Metrics

• Trained on 65,675 unique anime images (96x96)
• AE reconstruction: PSNR 27.8 dB
• Classifier-free guidance (cfg=4.0) at sampling
• Latent normalization (per-channel mean/std) prevents out-of-distribution latents
• DDIM 30 steps + CFG: fast generation

Usage

python infer.py --ckpt model.pt --prompt "1girl, halo, cowboy shot, long hair" --out results -n 4 --cfg 4.0

Use comma-separated Danbooru-style tags: 1girl, halo, cowboy shot, long hair, 2girls, school uniform, etc.

Files

• model.pt - full trained model (15 MB)
• infer.py - inference script