FIX CRITICAL SD 2.1 ERROR

[miguelgargallo]

FIX CRITICAL SD 2.1 ERROR
fix explication here
https://www.reddit.com/r/StableDiffusion/comments/zf21db/comment/izai3di/?utm_source=share&utm_medium=web2x&context=3

[miguelgargallo]

Download Stable diffusion 2.0 files

After installation, you will need to download two files to use Stable Diffusion 2.0.

Download the model file ([768-v-ema.ckpt](https://huggingface.co/stabilityai/stable-diffusion-2/resolve/main/768-v-ema.ckpt))

Download the [config file](https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/v2-inference-v.yaml), rename it to 768-v-ema.yaml

Put both of them in the model directory:

stable-diffusion-webui/models/Stable-diffusion

Source: https://stable-diffusion-art.com/how-to-run-stable-diffusion-2-0/

It will began to download

My folder content

SOLUTION CONFIRMED

Now I need to deal with the black images...

You just need to add --no-half to the file

done!

[AnonymousCervine]

Casually dropping a "--no-half" on everyone like that seems rather irresponsible. This may be fine for those who can soldier on at full-precision, but...

(And even if one were switching to that as the default, this would be the wrong way. It's a command-line switch for a reason; one does not simply hardcode user options in the launch BAT, unless one is the end-user I suppose. Also applying only to the BAT and not the .sh would only help users on Windows. Basically: The "solution" in that reddit fix is a quick convenience hack for users on Windows who want to use 2.1 ASAP; it's not an appropriate approach for everyone.)

[nousr]

we can export the ATNN_PRECISION environment variable to enable half-precision for 2.1, check the commits in the stablediffusion repo for more info

[miguelgargallo]

@nousr has a great constructive answer, so, @AnonymousCervine relax please,this is a temporal fix to let new enthusiast try 2.1 solves the problem and its well documented, lets see, I am in good faith to help the community, so, you just talk one day and got a contributor, let devs and engineers PR the project and see what happen, keep calm bro

[drax-xard]

The user.bat/.sh is for the user to configure, AFAIU. Having a random commit add/change commands to a user defined config file feels off. A disclaimer on the readme or just a post on reddit/discord informing users of this quickfix seems enough imo.

[Orthestral]

This forces everyone on full precision, no matter if they do or don't wanna use the bleeding edge 2.1 before a more proper fix is put into place.

[rubenlg]

@miguelgargallo I'm also excited about 2.1, but if the cost is VRAM, I'd rather stick to 2.0 until a different fix is found. My GPU doesn't have a ton of VRAM, and I'm sure many people are on the same boat.

If the PR is forcing --no-half on everyone (as opposed to just those running 2.1), I'd argue it's not fixing the problem, but rather trading a problem for another. And in my opinion, the problem you are creating (breaking existing users with low VRAM) is more severe than the problem you are solving (helping people trying the new model).

[ashgray0]

What would be the equivalent fix for webui-user.sh?

[drax-xard]

Also, just using xformers fixes it apparently

…

On Wed, Dec 7, 2022, 19:48 Orthestral ***@***.***> wrote: This forces *everyone* on full precision, no matter if they do or don't wanna use the bleeding edge 2.1 before a more proper fix is put into place. — Reply to this email directly, view it on GitHub <#5506 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A3R4EPUUQDMV6VE2BK7VTSLWMEH4BANCNFSM6AAAAAASXELRM4> . You are receiving this because you commented.Message ID: ***@***.***>

[clockworkwhale]

Also, just using xformers fixes it apparently

This is correct, the being-forced-to-full-precision thing can be avoided entirely simply by enabling xformers.

[DrakenZA]

Could we please get some basic testing going before pushing PRs that end up breaking everything for most users.

[DrakenZA]

tshi

@miguelgargallo I'm also excited about 2.1, but if the cost is VRAM, I'd rather stick to 2.0 until a different fix is found. My GPU doesn't have a ton of VRAM, and I'm sure many people are on the same boat.

If the PR is forcing --no-half on everyone (as opposed to just those running 2.1), I'd argue it's not fixing the problem, but rather trading a problem for another. And in my opinion, the problem you are creating (breaking existing users with low VRAM) is more severe than the problem you are solving (helping people trying the new model).

This isnt really SD2.0 . Someone changed something in the code, that break 2.0 support at half-perc in general.
You will get the same issues with SD2.0 sadly.

[ClashSAN]

added this as a note to wiki:

If 2.0 or 2.1 is generating black images, enable full precision with --no-half or try using the --xformers optimization.

we don't edit webui-user.bat, as it breaks user's current settings, we let user decide

[aliencaocao]

@ClashSAN maybe in the wiki you should warn that SD 2.0 is more sensitive to FP16 numerical instability (as noted by themselves in Stability-AI/stablediffusion@c12d960)
So users will know that the potential cost might be larger than using SD1.5 and below.

[ClashSAN]

@aliencaocao Is it with both 2.0 and 2.1, or SD 2.1 specifically?

[aliencaocao]

Both. It is due to their new cross attention module