SDXL Turbo: A Real-Time Text-to-Image Generation Model

[drax-xard]

Today, we are releasing SDXL Turbo, a new text-to-image mode. SDXL Turbo is based on a novel distillation technique called Adversarial Diffusion Distillation (ADD), which enables the model to synthesize image outputs in a single step and generate real-time text-to-image outputs while maintaining high sampling fidelity. For researchers and enthusiasts interested in technical details, our research paper is available here. It's important to note that SDXL Turbo is not yet intended for commercial use.

Blog post
Paper
Model weights

You can try the themo at Clipdrop

[drax-xard]

From my very limited testing with clipdrop it is, indeed, fast AF.
The quality is meh, mainly on faces and photographic details, but hopefully fine-tuning can improve that.

[lxkaka]

+1

[ilikenwf]

Seems that the comfyui guys have a turbo workflow for sd 1.5:

https://old.reddit.com/r/StableDiffusion/comments/187ps59/got_realtime_turbo_workflow_working_sd_15_lcm_and/

[vladmandic]

read the thread - that same stuff works here. quality is also improved in dev branch already.

[ilikenwf]

Well, I tried, assuming the 1.5 lora here is the one to use...

With LCM do I need models other than normal 1.5 models or is the lora sufficient, because using the guide int he comments my outputs suck.

https://huggingface.co/latent-consistency/lcm-lora-sdv1-5/tree/main

[vladmandic]

use dev branch or wait until tomorrow, planned merge dev to master. like i said, quality is improved in dev branch.