FlameF0X/NanoDiffusion-46M-1k

Mini Latent Diffusion Model

A ~46M parameter UNet trained from scratch for text-to-image generation in latent space (SD VAE + CLIP text encoder, both frozen).

Dataset: BitTranslate/Bittensor_subnet_19_06_04_24
Trained steps: 1000
Image size: 256px → 32×32 latents

Architecture

• UNet base channels: 128
• Channel multipliers: (1, 2, 3, 4)
• Res blocks per level: 2
• Cross-attention (text conditioning): latent res (8,)
• VAE (frozen): runwayml/stable-diffusion-v1-5 vae subfolder
• Text encoder (frozen): CLIP ViT-L/14 from runwayml/stable-diffusion-v1-5

Inference

from train_mini_ldm import generate
imgs = generate("a sunset over the ocean", "./mini_ldm_output/final")