README.md

---
license: mit
tags:
- text-to-image
- diffusion
- latent-diffusion
- pytorch
- coco
library_name: pytorch
pipeline_tag: text-to-image
---

# anilegin/lightweight-diffusion-ldm

Custom lightweight latent diffusion text-to-image model.

This repository contains inference-only files:

- VAE config and stripped VAE weights
- LDM/UNet config and stripped LDM weights
- diffusion/sampler code needed for DDPM and DDIM
- a simple `inference.py` script
- generation defaults in `generation_config.yaml`

The checkpoints are stripped to contain model weights only; optimizer state, scheduler state, and training logs are not included.

## Install

```bash
git clone https://huggingface.co/anilegin/lightweight-diffusion-ldm
cd lightweight-diffusion-ldm
pip install -r requirements.txt
```

## Generate images

```bash
python inference.py \
  --prompt "a small dog sitting on a red couch" \
  --sampler ddim \
  --num-steps 50 \
  --guidance-scale 3.0 \
  --precision bf16 \
  --output-dir outputs/example
```

For offline/local-only CLIP loading, make sure `openai/clip-vit-large-patch14` is cached locally and add:

```bash
--local-files-only
```

## Notes

This is a custom PyTorch implementation, not a native Diffusers pipeline. The included source code is required for inference.

## Training data

Trained/evaluated with COCO-style image-caption data. Add more precise dataset, metrics, and limitations here before making the repo public.

## Citation

If you use this model, please cite the project/repository.