Hacker News Reader

Comment by Severian | Hacker News Reader

Works with Automatic111. Generated 20 512x512 on a lowly RTS 2070S with 8GB RAM.

Prompt: a man

Steps: 1, Sampler: Euler a, CFG scale: 1, Seed: -1, Size: 512x512, Model hash: e869ac7d69, Model: sd_xl_turbo_1.0_fp16, Clip skip: 2, RNG: NV, Version: v1.6.0

Examples: https://imgur.com/a/UuuT9qu

View on HN

I've done a bit of fiddling around with it and definitely holding back judgement for now, seems like the 1 and 2 step images are WAY more coherent than LCM, but the images are kinda trash for any kind of prompt complexity so you start to have to use more steps, and since the individual steps take the same amount of time (I think there's a specific sampler for this which may be faster & better?) by the time you start prompting details you end up using 4 steps and the perf is about the same as LCM, and that breaks down the same way as you start going for more complexity (text, coherent bg details etc) because you end up needing 10-15 steps and at that point you're going to get a much better result from full-fat SDXL x dpmpp3msdee (lol)

Curious to see the bigbrain people tackle this over the next few days and wring all the perf out of it, maybe samplers tailored to this model will give a notable boost.

Are you finding dpm++ 3M SDE better than dpm++ 2M SDE in sdxl?

Afaik the second order (2M) version is the recommended one to use for guided sampling vs the 3rd order one.

From here: https://huggingface.co/docs/diffusers/v0.23.1/en/api/schedul...

> It is recommended to set solver_order to 2 for guide sampling, and solver_order=3 for unconditional sampling.

It might be placebo, but I find 3M better for upscaling, when I usually set CFG quite low and use a generic prompt that doesn't describe any localised element of the picture.

Which is what it's meant for, I suppose.

In my tests it's basically been 50/50, i probably did ~40 or so comparisons when i was testing samplers and i felt like there were a couple that seemed really good on the 3rd order one, but idfk, it was very very close, I don't know if I saw a single gen where one of the two was bad but the other wasn't.

I just want to point out that I’ve noticed sdxl isn’t good at producing images that are 512x512 for some reason. It works much better with at least 768x768 resolution.

Normal SDXL requires 1024x1024 output or the quality degrades significantly.

Hi Max! Thanks for all the tuts! Small correction - SDXL wants ~1 megapixel resolutions at a variety of aspect ratios.

https://github.com/lllyasviel/Fooocus/issues/24

Fair point, although in my experience SDXL still isn't great at non-square ratios. I end up just cropping them to the ratio I want.

What have you seen to be the issue? Composition? Realism? Prompt adherence? I’m just finishing a project having generated tons of images at a mix of ~6 aspect ratios and I haven’t noticed any difference.

Indeed, however one of the listed limitations is: The generated images are of a fixed resolution (512x512 pix), and the model does not achieve perfect photorealism.

I did with Automatic1111 as well. With an RTX 4090 with 24G VRAM, steps 1, cfg scale 1, size 512x512, model sd_xl_turbo_1.0. I was generating more than four images a second.

Reply on news.ycombinator.com