The new DPM Adaptive sampler is fantastic

[bmaltais]

Not sure if others have tried the new DPM adaptive sampler but boy does it produce nice results. Was trying a lexica prompt and was not getting good results. Switched to DPM Adaptive and 4 fold quality improvement:

euler:

DPM Adaptive:

GIve it a try. Note that I am using a custom Dreambooth CKPT trained on Midjourney style... not the default SD14 ckpt.

Lexica prompt: https://lexica.art/?prompt=11a88b5f-833b-4304-a321-1cb5bf4d870c

[bmaltais]

You can also use it now in img2img to further tweak your images. Very impressive work by the automatic webui dev team!

[rpillala]

I've been trying this lately after seeing your post, and it's just a ton of fused arms and extra arms and legs etc. Am I doing something wrong? I've put terms like extra limb, deformed, fused limbs, etc in the negative prompt without much effect. When I put too much in the negative prompt, the results suffer in different ways. DPM Adaptive seems to decide its own number of iterations so I can't play with that like I usually would.

[bmaltais]

Indeed. It can be hit and miss. For the most part it has played nice for me. I always try to stick to coherent resolutions in SD. Perhaps you are going beyond resolutions that produce coherent images. If you multiple the width by the height and it is greater than 290000 that the resolution is going in incoherent territory and will tend to result in difformites. Maybe this is the source of your issues? The better method is to create images with total pixel count close to 290000 and then to scale the result in img2img the keep the coherency with good quality

…

On Sun, Oct 16, 2022, 4:56 p.m. rpillala ***@***.***> wrote: I've been trying this lately after seeing your post, and it's just a ton of fused arms and extra arms and legs etc. Am I doing something wrong? I've put terms like extra limb, deformed, fused limbs, etc in the negative prompt without much effect. When I put too much in the negative prompt, the results suffer in different ways. DPM Adaptive seems to decide its own number of iterations so I can't play with that like I usually would. — Reply to this email directly, view it on GitHub <#2274 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABZA34USBGK56HDF76QPBIDWDRTW5ANCNFSM6AAAAAARCJK6QA> . You are receiving this because you authored the thread.Message ID: <AUTOMATIC1111/stable-diffusion-webui/repo-discussions/2274/comments/3891299 @github.com>

[rpillala]

I've been trying with 512x640 maybe that's the problem. It's not much larger, but it is larger.

[noobamuffin]

using resolutions other than 512 can break a lot of prompts and models.