Is it me or has this gotten a lot looser with the reference image lately?

#472
by deleted - opened
deleted

Prompt:

fantasy art, woman wearing a green strapless dress

Sketch:

Generations:

same here. Since yesterday the quality of the output just got really bad a lot of times. It seems there has been a downgrade of the model... :-(

hmm, not that we know of, no..

I was sometimes using the included rest api with a much lower strength value. It worked since yesterday flawlessly. Now when I'm using strength=0.1 for example, it still will almost completely ignore my input image and come up with something very different

deleted

This probably demonstrates what's going on better. Sketch and prompt taken from an earlier thread. Back then, all of the results were pretty close to the original. Now, colors and placement are pretty much ignored.

Prompt:

cupcake, realistic, hd

Sketch:

Generations:

Huggingface Projects org

Oh. We have switched the backend from "vanilla stable diffusion" to controlnet. Looking into it

Huggingface Projects org
โ€ข
edited Mar 8, 2023

Unfortunately, the ControlNet sketch model was only trained on black & white sketches. Therefore, I've removed the color picker option from the demo. Please let me know wdyt
Screenshot 2023-03-08 at 09.26.51.png

deleted

Maybe it just takes some getting used to, but this seems kind of disappointing

image.png

@mishig kinda destroys what Diffuse-the-rest has become famous for in the first place ๐Ÿ˜‚
@bugmenot "Stable Diffusion, please give woman" Stable-Diffusion: "Fuck it, I'll do a cocoon" ๐Ÿ˜‚

This is painful. Somebody Puh-leeze bring back the old version!!! Or something like it.
Best art tool ever, because you could nudge each iteration towards your goal with settings around 0.2 or 0.3, and not throw the baby out with the bathwater. You could slowly coax out a great image. Now it's like everything's a 0.99 setting.

This comment has been hidden

something happened, the old version (a month ago maybe?) was much better, what happened?

Praying for a brand new space called "diffuse-the-rest classic" like the original "diffuse-the-rest" by dbirks. There's no good substitute right now.
Also, it would be nice if you could go up to 1024 x 1024px at least. That was the only drawback with the old one: the resolution limit.

sign... missing creation tools and accuracy very badly :-(

I'm going to join in and complain too.
I had great time proomptin' fantasy and horror portraits with consistent good results:
new IWD3 style.png
japanese horror.png

It was such an amazing tool!
But now everything look like... secret embarrassing DeviantArt days of Boris Vallejo from alternate universe. And results seem to take no inspiration from the provided picture. And it somewhat hates heads.
canvfffas.png

Sign up or log in to comment