Exactly same prompt no more the same ?

[slavoisard]

Hi !
I'm trying to "rework" on an image generated a very long time ago (AI time speaking, so just 7 days ago). I'm using the .PNG info to transfer all the details I used at that time, even reinstalled a fresh Automatic1111 setup so it's not corrupted by all the new parameters I have added recently... Anyway, the image is not at all the same. Any idea what could go wrong ?

[Pernjakovic]

Have you tried ticking the "Use old emphasis implementation. Can be useful to reproduce old seeds." in the settings menu?

[MasterZap]

Also check if you used highres fix or not, and of course, what model file you were using.

[Orisen]

I have the exact same problem and even with "Use old emphasis implementation. Can be useful to reproduce old seeds." enabled i'm not getting the same image anymore with the exact same settings since I cloned the repo to the newest WebUI version.

[philvitthotmail]

Same here. I stopped being able to reproduce old images more or less at the time of the hypernetworks implementation. But honestly at the speed they are merging new untested features into the main branch, it could be anything else.

Can somebody suggest what it the best way to weekly duplicate/archive old versions of the repo without duplicating all the models within (note: i'm a total gitnoob) ?

[slavoisard]

Tried it, still getting the same new result :(

[slavoisard]

Also check if you used highres fix or not, and of course, what model file you were using.

I did use Highres Fix so I'm using it again, same Denoising strength... But before doing the Highres work, you can see from the start of the Init Image generation that i's not at all the same look. :(

[horribleCodes]

Do you use xformers? Using it introduces rounding errors so images with identical parameters will have slight differences.

[horribleCodes]

And just a little tip on the side: the samplers with an a at the end, like Euler a, don't really need more than 3-40 steps, because they constantly change the image, as opposed to the ones that approximate towards a result.

[Kosinkadink]

@Orisen try "git checkout -b beforefix f347ddf", this will create a local branch called "beforefix" thst uses the commit right before the Unlimited Token Works fix.

[slavoisard]

OK... For my problem, I did find the solution. At that time I was not on the regular .ckpt, I was using another model I just tested through Dream Booth. I thought PNG Info was keeping the info about the model used... but not. So using this .ckpt with the old prompt, I now get the same result as initially ! Good !
The strange thing is that I understand that using an alternate .ckpt file does change the seeds results... but it's like all the styles are much more better with this homemade .ckpt than the original one.

[philvitthotmail]

So I unticked "Enable quantization in K samplers for sharper and cleaner results." and I was able to reproduce old images.

I think that any options that have an impact on the final output (old emphasis, quantization, hypernetworks, etc.) ought to be directly in the main interface and not hidden in the settings . Also it should be in the png info.

[Orisen]

@Kosinkadink ok i switched. should i test it? I did. still not the same. :( I mean it's not important, i'm currently more interested in why it is happening and what was causing it.

[slavoisard]

So to demonstrate the impact of xformers, here's the same prompt, without xformers (left), and with xformers (right). Of course, the result is slightly different, but the one on the right as a blurry effect. In fact, since I installed xformers I found that most of my results are a bit blurry... So back to No Xformers ;)

[a3nima]

more like a "furry" effect ^^

[C43H66N12O12S2]

You should compare xformers against --disable-opt-split-attention as split attention itself modified results from the default attention.

No xformers

xformers

[a3nima]

what shall we conclude then?

[0xdevalias]

A friend mentioned that using xformers could make things non-deterministic, and that there were a lot of references to it on the repo issues here. Wanting to understand a bit more about it, and link a bunch of potentially related issues together, I tried to find as many issues as I could that seemed to be related to xformers and the potential for it to be causing non-deterministic / unstable / inconsistent results:

The following may potentially be related (ordered by issue number):

Issues:

• Xformers optimization causes inconsistent generations with the same parameters #1999
• No token limit for prompts broken? Different results with the same prompts and seed #2005
• The same image can no longer be output under the same conditions. #2197
• On the problem that the same resolution can not be generated stably #2293
• Different results after swapping graphics card #2423
• Different images for same seed with xformers and "Highres. fix" enabled when ancestral samplers are used #2528
• Different outputs since WebUI update #2943
• xformers and euler_a variant results for same generation settings #3054
• [Bug]: The prompts do not produce the same results as before the latest updates on 10/22/22 #3476
• [Bug]: xformers always generate different results, seed is the same! #3967
• [Bug]: In img2img, setting denoising strength to 0 does not return original image #4011

Discussions:

• Different results after swapping graphics card #2418
• --xformers CMD line option causes slightly poorer quality images on GTX 1650 (Turing), possibly others? #2592
• Highres.fix composition #2705
• Exactly same prompt no more the same ? #2849
• xformers option has good and bad aspects #2958
• What has changed recently? Different and Worse Results from Same Settings #3461

I also came across this thread in the xformers repo, which while I can't guarantee is related, am wondering if it might be:

• xformers ViT-B ImageNet MAE + Deepnorm training instability facebookresearch/xformers#219

And a question I raised on a PR in the diffusers repo:

• Up to 2x speedup on GPUs using memory efficient attention huggingface/diffusers#532 (comment)

  • A friend just mentioned that apparently using xformers tends to lower the quality of generated images, and apparently even causes StableDiffusion to generate different images for the same seed/settings. I haven't looked too deeply into things to try and validate that, but I was wondering if a) that is something that is currently known/documented, and b) if that's likely to be fixed at any point, or if it's just 'the price we pay' for the benefits it adds. Would any of that be a concern with integrating things here? I did stumble upon this issue on the xformers repo, but not sure if it would be the same/similar root cause as my friend described: xformers ViT-B ImageNet MAE + Deepnorm training instability facebookresearch/xformers#219