Stable Diffusion Demo 🤖 🖼️

This repo contains info, configs, and notes for running stable diffusion locally.

Enviroment

Hardware

• GPU: Nvidia GeForce RTX3070 (Laptop) (8GB GDDR6 VRAM)
• RAM: 32GB DDR4

Software

• OS: Win 11
• Python 3.10.0
• Model Checkpoint: v1-4

Installation Steps

1. Download weights ¹
2. Download any fork, I used the web-ui fork ²
3. Change to name: ldo in environment.yaml
4. Rename weights .ckpt file to model.ckpt and place in /models/ldm/stable-diffusion-v1 directory
5. Run conda env create -f environment.yaml in (base) miniconda3
6. Activate env by conda activate ldo
7. Run webui.cmd in (ldo)

NOTE: The original repo ³ gave me a RuntimeError: CUDA Out Of Memory even with --n_samples 1. The optimized fork ⁴ has all the safety_checks removed and uses less VRAM and has a faster result generation time.

P.S. I didn't try the model.half() tip that some people have suggested to use with the original repo for low VRAM usage.

Results (cherrypicked ofc!)

Prompt#1: "a cat wearing a red hat"

Prompt#2: "a busy indian street market, artstation digital art"

Prompt#3: "a rat in a shiny medieval armour posing for a victorian era portrait, oil painting"

Prompt#4: "a steampunk flying machine flying in a beautiful blue sky, 4k"

Stats

I generated images in batches of 4 at a time.

Generation time ~ 40s (10s per image)

Peak memory usage: 7854 MiB / 8192 MiB / 95.871%

The steampunk one (Prompt#4) was an outlier, it took 330.05s total (82.51s per image).

References

Footnotes

1. https://huggingface.co/CompVis ↩

2. WebUI with all optimizations pre-applied: https://github.com/sd-webui/stable-diffusion-webui ↩

3. https://github.com/CompVis/stable-diffusion ↩

4. Optimized: https://github.com/basujindal/stable-diffusion (Worked well enough) ↩