---
title: Stable Diffusion Lab
emoji: 🎨
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.30.0
python_version: "3.12"
app_file: app.py
pinned: false
license: creativeml-openrail-m
short_description: Explore Stable Diffusion with live denoising previews
---

# Stable Diffusion Lab

A teaching Space for exploring text-to-image diffusion. Students can set:

- positive and negative prompt
- CFG scale (0.0 to 15.0) and denoising steps (4 to 48)
- sampler: Euler, Euler Ancestral, DDIM, UniPC
- checkpoint: Stable Diffusion 1.5 or Realistic Vision V6.0
- seed (or a random seed) and resolution (512x512 by default, 256 to 768 px)

While an image is generated, a live preview of the denoising is streamed to the page.

## Hardware

Designed for a dedicated GPU (fp16, about 6 GB VRAM for both checkpoints). It also runs on CPU, but slowly. Both checkpoints load at startup, so the first build takes a few minutes while the weights download.

## How the live preview works

The app runs its own denoising loop. At every step the U-Net predicts the noise, and from that prediction the app computes the model's current guess of the final image. That guess is decoded with TAESD, a tiny and fast VAE decoder. The final step uses the full VAE of the checkpoint.

## Settings (Space variables)

| Variable | Default | Meaning |
| --- | --- | --- |
| `SAFETY_CHECKER` | `1` | Set to `0` to turn off the NSFW filter (it can flag harmless art) |
| `MAX_CONCURRENT` | `1` | Generations running at the same time |
| `PRELOAD` | `1` | Load both checkpoints at startup instead of on first use |
| `SD15_REPO` | `sd-legacy/stable-diffusion-v1-5` | Base model repo (the original runwayml repo is deprecated) |
| `RV_REPO`, `RV_FILE` | Realistic Vision V6.0 B1 noVAE, fp16 file | Finetuned checkpoint |
| `RV_VAE_REPO` | `stabilityai/sd-vae-ft-mse` | VAE for Realistic Vision, whose repo ships without one |

## Licenses

Both checkpoints use the CreativeML OpenRAIL-M license.