Open 1ewig opened 1 year ago
More often than not --precision full --no-half
arguments are required to prevent the black squares.
also try --medvram
instead of --lowvram
, may as well be enough, and also get a sizable speed boost.
Also --opt-split-attention-v1
won't hurt.
More often than not
--precision full --no-half
arguments are required to prevent the black squares. also try--medvram
instead of--lowvram
, may as well be enough, and also get a sizable speed boost. Also--opt-split-attention-v1
won't hurt.
All right,,, output is not black image now ty, but I cant generate more than once without restarting the web ui, like 1 picture in 1 run, help me twice eh
I have the same problem as you, even after adding parameters --precision full --no-half
, it still can't be solved, is there any other solution?
Most of black images are due to --opt-sub-quad-attention
option.
--opt-sub-quad-attention
makes speed faster and lowers vram occupation (then you can generate high-resolution images).
But when you add it to commandline arguments, black images can be generated.
The only known solution is to keep generating black images until they are not generated.
(I have kept generating images to debug it, and at some point, I checked that the images were generated properly)
On RX580 (8Gb) black images do appear, but very rarely (1 per 100 generations or even less). I use the whole set of arguments including --opt-sub-quad-attention
.
I just remove --opt-sub-quad-attention
and it works now. I use --medvram
for faster generation on my RX6600XT
@Miraihi I have RX580 too, can you share all arguments you use to launch app?
@Miraihi I have RX580 too, can you share all arguments you use to launch app?
Ones that do matter are --medvram --precision full --no-half --no-half-vae --opt-split-attention-v1 --opt-sub-quad-attention --disable-nan-check
The maximum picture size you can generate is 800x600. When generating 800x600 I'm getting 2-3 seconds per iteration.
@Miraihi I have RX580 too, can you share all arguments you use to launch app?
Ones that do matter are
--medvram --precision full --no-half --no-half-vae --opt-split-attention-v1 --opt-sub-quad-attention --disable-nan-check
The maximum picture size you can generate is 800x600. When generating 800x600 I'm getting 2-3 seconds per iteration.
Sometimes it crushes when I try to do more than 100 iterations. Did you face it?
Sometimes it crushes when I try to do more than 100 iterations. Did you face it?
To be honest I never do more than 30 iterations. 100 seems to me extremely excessive, at least for the samplers I use (Eueler A and DPM++ 2M Karras). I recommend to rather look for a better prompts and more appropriate models than cranking up that many iterations.
Sometimes it crushes when I try to do more than 100 iterations. Did you face it?
To be honest I never do more than 30 iterations. 100 seems to me extremely excessive, at least for the samplers I use (Eueler A and DPM++ 2M Karras).
Previously I was using Google Colab, so I set batch count to 4 and iteration to 50. How are you work with these settings? Maybe you're making single samples?
Previously I was using Google Colab, so I set batch count to 4 and iteration to 50. How are you work with these settings? Maybe you're making single samples?
Making really big batches is not possibe on DirectML at current state. Also the large number of iterations isn't required to get a decent picture (Go to civitai and look for examples.). So what I do currently is using ControlNet extension for a decent txt2image base and then inpainting until I'm satisfied.
Hello, maybe too late to this topic, but i had same issue with black screen outputs. Possible FIX: What helped me, was changing checkpoint to different one, and then back to the one I wanted to use. Now, everytime i get black output, i just quickly change checkpoint (model) from dropdown, Then it generate pictures correctly.
I'm using RX 6700 XT with following arguments:
--use-directml --medvram --precision full --no-half --no-half-vae --disable-nan-check --opt-split-attention-v1 --upcast-sampling
Is there an existing issue for this?
What happened?
UI boots up without error and after generating an image it comes out as black image
Steps to reproduce the problem
What should have happened?
output should have been a bit colorful
Commit where the problem happens
commit: 3e855524
What platforms do you use to access the UI ?
Windows
What browsers do you use to access the UI ?
Google Chrome
Command Line Arguments
List of extensions
none
Console logs
Additional information
model im using is rpg v4 and its a safetensor file