MStricklin/data665-text-to-image

Text to Image (ZeroGPU)

A Gradio demo that generates images from text prompts using a locally-loaded diffusers SDXL-Turbo pipeline, running on a ZeroGPU-shared GPU that's attached only for the duration of each generation call.

Setup

1. 1.In the Space's Settings → Hardware, select ZeroGPU.
2. 2.No secrets or license acceptance needed — stabilityai/sdxl-turbo is ungated.
3. 3.duration in the @spaces.GPU(duration=60) decorator is the GPU time budget requested per call — raise it if generations are timing out, lower it to reduce queue wait for other users.

Notes

• —SDXL-Turbo is distilled for 1-4 step, guidance_scale=0 sampling (already set in app.py) — that's what makes it fast, but it also means the negative-prompt field has no effect (no classifier-free guidance at guidance_scale=0).
• —It was trained at 512x512; larger sizes work but quality can degrade.