[Guide] Significally increase speed with sd 2.1 turbo

[light-and-ray]

In webui was added support of sd_turbo. It's generally turbo version of sd 2.1 512. I've tested, and it works nice with StableSR. You need to:

1. update webui to 1.7.0 version
2. download sd_turbo model
3. (upd webui 1.8 has native lcm sampler support) install lcm sampler or use it from animate-diff extension, if you have it installed
4. install StableSR 512 model
5. decrease sampling steps to 4
6. use Stable SR stablesr_webui_sd-v2-1-512-ema-000117 with sd_turbo

[IceClear]

The 768 model should also work? I guess.

[light-and-ray]

The 768 model should also work? I guess.

There's no sd turbo for the 768 model. Stablesr768 + sd 2.1 512 doesn't work

[light-and-ray]

Also I've noticed higher number of steps produces more blurred image. So 4 steps not only the fastest, it also has the best quality. 3 and less produce a lot of artifacts

[IceClear]

Also I've noticed higher number of steps produces more blurred image. So 4 steps not only the fastest, it also has the best quality. 3 and less produce a lot of artifacts

I guess it is because sd turbo directly distilled on 4 steps. I will consider finetuning current model on the sd turbo if I have time later :）Thx for your help.

[CCpt5]

Also I've noticed higher number of steps produces more blurred image. So 4 steps not only the fastest, it also has the best quality. 3 and less produce a lot of artifacts

I guess it is because sd turbo directly distilled on 4 steps. I will consider finetuning current model on the sd turbo if I have time later :）Thx for your help.

First of all, I love all the work you've done on this. The results are brilliant and I've used StableSR successfully for months now.

I was curious if there's still a plan to eventually release an SDXL model? Or did that idea get abandoned? Perhaps due to time or feasibility?

[IceClear]

Hi. Thx for your interest and glad to hear that.
We are interested in developing a XL version. But I am currently struggling with my PhD projects and may not be available for that.

[CCpt5]

Heads up that an SD2.1 model (LoRA) with similar low step inference was recently released here: https://huggingface.co/h1t/TCD-SD21-base-LoRA

They use a method called Trajectory Consistency Distillation (TCD) as explained in their paper here: https://arxiv.org/abs/2402.19159

Any thoughts on whether or not this would work with StableSR similar to how Turbo has? If so it's worth checking to see if their model was trained to also work w/ the 768x768 resolution.

If anyone gives it a go before I get a chance please report your findings!