On the problem that the same resolution can not be generated stably

[RokiRokiPoi]

When I use stable diffusion webui, I can generate 1600 896 images, but this is unstable, and black images usually appear. Besides, if I restart webui-user.bat, I may not be able to generate 1600 896 images.
I want to know why, if I can generate such a picture, why I can't generate it again because I restart it.If I really need to generate such a picture, what should I do to help me realize it?

• OS:window10 i711800h 3060 6g laptop
• Browser Microsoft edge
• Commit revision :maybe it's Commit hash: 6a9ea5b
  
  This is a 1600 960 resolution picture that I once generated.
  I am waiting for your reply.Please solve my problem.

[RokiRokiPoi]

there is the error report

[aleksusklim]

Probably something in your system uses GPU resources too. Try this:

1. Close all other applications (especially all other browser tabs).
2. Disable hardware acceleration in your current browser (actually, you can use for this GUI one browser you like less – for example Edge or Firefox if you prefer Chrome for everyday tasks) and in other applications (notable example is Telegram Desktop).

[RokiRokiPoi]

Actually, I started webui when I turned on my computer, At that time, there was no software that occupied memory or video memory. but it didn't work. Still can't generate pictures.

As for the choice of browser, I'm using edge now, and I can't generate such a picture without enable the hardware acceleration, that is, nothing has changed, which is why I feel very puzzled.

[ClashSAN]

@RokiRokiPoi
make use of --lowvram and --xformers. put both flags in commandline args in webui-user.bat. If you can't install the xformers, then --lowvram only.

..What you said above is strange, you cannot turn off hardware acceleration in the browser? Try a different one, I can disable this on Brave browser.

[kryztoval]

You can check how much vram you have available by using nvidia-smi command and reading the output
I see you are using waifu diffusion, that model is bigger unless you have pruned it.

You may have better luck using the low ram setting mentioned by others earlier.

In my setup it uses a lot more than 4gb of ram with no optimizations: Torch active/reserved: 8835/11526 MiB, Sys VRAM: 11458/24576 MiB (46.62%)

I can try your prompt if you share the txt generator for it.

[RokiRokiPoi]

@ClashSAN
I'm sorry for my mistake. I mean, I still can't generate such a picture even though I have turned off hardware acceleration.
The last time I generated the resolution of 1600896, I used---medvram --no-half, but now I can't generate the image of 1600 896 by starting webui-user.bat even though I haven't changed anything.
Now I want to continue to use --medvram and --nohalf, I tried once --lowvram, but it was too slow, and it took nearly ten minutes to generate a picture.
Is xformer useful? If it can help me, I will try to install it.

[RokiRokiPoi]

OK, I found the problem:--no-half, delete--no-half. I can generate 1600 900 images (although many of them are black image).
Can someone tell me what role --no-half has? I feel it can help me improve the image quality.I'm not sure.

[aleksusklim]

Can someone tell me what role --no-half has?

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Troubleshooting#green-or-black-screen

But you stated you have GTX 3060 6Gb, it should not matter. You don't need no-half here.

I see you are using waifu diffusion, that model is bigger unless you have pruned it.

Personally, I merged WD with itself, saving as float 16 with that checkbox in merging tab – and it gave more lightweight file.

[RokiRokiPoi]

@aleksusklim
-- no-half really affects the quality of drawing for me, and I don't know how it is work. Even if I am a 3060, I still have black pictures, but there is no green picture.

But you stated you have GTX 3060 6Gb, it should not matter. You don't need no-half here.

I will upload those two pictures later, with the same tag and the same parameters but different results.

[kryztoval]

--no-half tells the model to use full precision numbers in memory instead of half precision which require less memory. The issue is that some cards do not support half precision like the GTX1650 IIRC.

Did you disable censorship? black images normally result if you create a NSFW image and did not tell the generator that NSFW was ok for you.

[RokiRokiPoi]

disable the --no-half

enable the--no-half

[RokiRokiPoi]

--no-half tells the model to use full precision numbers in memory instead of half precision which require less memory. The issue is that some cards do not support half precision like the GTX1650 IIRC.

Did you disable censorship? black images normally result if you create a NSFW image and did not tell the generator that NSFW was ok for you.

no,i never generate the nsfw image
Maybe using nohalf is a negative optimization for me?

Turn off nohalf, my black image generation rate is much higher.

[kryztoval]

No half does disable an optimization. using half is faster. You only want to disable half if the gpu does not support it.

[kryztoval]

the images I see from your exact prompt are not even close in style to what you posted - how did you manage to get that with just 20 steps?

[ClashSAN]

@RokiRokiPoi xformers will nearly double your previous batch count, inferring it doubles available vram. so yes, very helpful it would auto-install and enable for you automatically. add to commandline args:

--xformers

also turn off the censorship and see if there's a difference

[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