Can't reproduce results? Try this first

[freecoderwaifu]

When people say they aren't able to generate the same results, they (sometimes vaguely) mostly tend to mean two things:
-Generating an exact same image as before as a benchmark or sanity check
-Generating the same overall pixel quality and/or prompt accuracy as before

Some things to check or remember if you're facing either of those cases:

Check your basic settings: sampler, steps, CFG, Clip Skip, resolution, checkpoint, etc.

Check if you were using options in the Extra checkbox section

Check if you were using highres fix and its specific settings: resolution, upscale by or upscale to, etc.

Remember the new method for enabling highres fix: if dropdown/accordion expanded = enabled, if contracted = disabled.

Check if Upscale didn't revert to Latent with low denoising in highres fix.

Check your prompt, prompts like: word, word:1.9, meme prompt engineering (1:999:6969), etc. might have parsed in a good way by miracle before an update, but now they might be overbaking. Try with lower weights or none at all.

Check Use old emphasis implementation. Can be useful to reproduce old seeds. in Settings > Compatibility to reproduce old prompting styles, it might still help with those types of prompts.

Also check other compatibility settings there.

Check your negative prompt, same as above.

Check if your prompts used embeddings and check that they're still in the embeddings folder.

Check if your prompts used hypernetworks and check that they're still in the models/hypernetworks folder.

Check if your prompts used Loras and that they're still in the models/Lora folder, or models/Lycoris.

Check and try the built in support for Loras/Lycoris/Loha, etc. instead of extensions. With built in support you can place all Lora types in the same Lora folder and add them with <lora:name:strength> regardless of type.

Try either the built in Lora support or any of the extensions you might have been using. Disable either of them in the Extensions tab and test which gives you the very specific results you might want.

If using built in support for Loras, try checking this option Lora/Networks: use old method that takes longer when you have multiple Loras active and produces same results as kohya-ss/sd-webui-additional-networks extension in Settings, Compatibility

Check your VAE and that it's applying correctly. Place your VAE in models/VAE and manually select them from the dropdown.

Check the options in Settings > VAE and see if the TAESD options didn't get enabled if you didn't want them.

Add any of these options to the main UI for faster access in Settings > User Interface > Options in main UI - txt2img. and img2img, note you can have separate options for either tab.

You can also add Restore Faces and Tiling here. Restart your UI to apply changes.

Any options you add there will show up below seed in the main txt2img or img2img settings frame.

Check your launch parameters in your launch.bat and see if they're still the ones you want.

Check your optimization method in case it changed. Note that some optimizations options are non deterministic, which means you'll see minor changes regardless. This depends on their version, newer versions of xformers for example can prevent this.

Check additional optimizations there in case they were used before.

Check your Sampler Parameters to see if they were using non default values like some (possibly outdated) guides, like the NAI guide, etc. might suggest.

Check the options in Settings > Stable Diffusion.

Specifically, these ones

Check if you weren't using additional extensions: Adetailer, ControlNet, CFG Scheduling, etc.

Check if you weren't using extensions that inject additional prompts: wildcards extensions, UmiAI, Stylepile, CloneCleaner, Model Keyword etc.

Check if your extensions didn't get enabled or disabled. Check if extensions you didn't want to use didn't get enabled. Delete them from the extensions folder to uninstall them.

Check your upscaler options

Check your img2img options

Check this page:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Seed-breaking-changes

Last option, delete your venv folder and let the launch.bat remake it, in case it's a dependency issue.

Additionally remember FP16 vs FP32 models, and upcasting to FP32 can also change an image. Other optimizations here might (?) also have an effect.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations

Still having issues? Feel free to post about it, but please provide your prompt and settings, as it is very difficult to guess the results you're looking to get, by just vaguely stating you have an issue.

[w-e-w]

thanks, I put a link to this on the wiki https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Guides-and-Tutorials

[wbzy00]

Gosh, that 1.6 change in seed breaking changes is so confusing.

[Ermanov76]

Hi, i've got a serious difficulty to reproduce a result. i,ve make a post on reddit, but no one fixed it: https://www.reddit.com/r/StableDiffusion/comments/16z6709/strange_behavior_of_adetailer_thanks_for_the_help/ if u got time to take a look thx a lot

[transferb3d]

CLIP SKIP has ruined everything. Where do we change that? I just want A1111 webui default settings back.

[gongzhang]

After installing webui on my new MacBook Pro, I found that there is a certain probability of generating incorrect images every time, even if the seed does not change. After messing around for an afternoon, I finally found that everything was normal when I switched from Safari to Chrome.