krita-ai-diffusion
krita-ai-diffusion copied to clipboard
Acly
Generated images are just black
Hi, I setup a local server following the installation guide and everything seems to be setup correctly. But after I give a prompt and "Generate" it takes about a minute to create 4 all black alternatives. I have and M1 Macbook with 8 GB of RAM, so I assummed it would take forever to generate, 1-2 minutes is pretty fast but it doesn't create anything. Can it be the lack of RAM? Thanks!
Hi, I setup a local server following the installation guide and everything seems to be setup correctly. But after I give a prompt and "Generate" it takes about a minute to create 4 all black alternatives. I have and M1 Macbook with 8 GB of RAM, so I assummed it would take forever to generate, 1-2 minutes is pretty fast but it doesn't create anything. Can it be the lack of RAM? Thanks!
Definitely not due to lack of RAM. I have an M1 Max with 64 GB of RAM and I'm seeing the same problem. Initial image gen works great, but subsequent expansion, fill, etc, does not
I was able to get images generating again by selecting "add a control layer", setting the strength 0, which is non-optimal, but may lead us to a solution
Seems like it happens when there's transparency for the background (the checkered gray background). When I painted it with white, it worked fine
Seems like it happens when there's transparency for the background (the checkered gray background). When I painted it with white, it worked fine
That sounds like you used transparent/black image as input though. ai generation ignores alpha/transparency entirely, it's same as black.
The original post sounds more like a issue with recent update of macOS/pytorch, but I can't find a lot of information on it.
That sounds like you used transparent/black image as input though
No, I'm not using img2img at all (at least as far as I can tell, I've only used Krita for a few days, but I'm broadly familiar with how SD works and was running my own ComfyUI workflows beforehand.
The workarounds I'm mentioning are only working in a few scenarios, sometimes they're not. Like right now I'm trying to do some inpainting on some sky with a new prompt, and the entire output is black. There isn't any transparency or blackness in the image at all.
Hey jmgpp!
I had the same issue on my M1 even after complete reinstall of krita and the ai-plugin. For some reason my server log spit out an error with the "comfyui-tooling-nodes" which is a ComfyUI custom node.
I went via terminal to the krita local server installation (which is in "/Users/USERNAME/Library/Application Support/krita/ai_diffusion/server/ComfyUI/custom_nodes"), removed the comfyui-tooling-nodes (rm -R) and did a clean git clone (git clone https://github.com/Acly/comfyui-tooling-nodes.git).
After restarting krita my images aren't black anymore. Maybe it helps ya, too.
Henry
Hey, I have no clue. One day it worked after updating the local server nodes, day after I'm getting black images again.
Strange thing is, you can log on the local server via browser because it's just a ComfyUI installation. Prompting an image there works. If you install a complete ComfyUI server and connect it to Krita, it works too. Seems like there is a bug in the pipeline between Krita and the local server.
Gesendet: Mittwoch, 03. Juli 2024 um 21:17 Uhr Von: "iamcavalon" @.> An: "Acly/krita-ai-diffusion" @.> Cc: "Hendrik van Heek" @.>, "Comment" @.> Betreff: Re: [Acly/krita-ai-diffusion] Generated images are just black (Issue #783)
What if there are no "tooling node" errors, any idea what to do?
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>
Can you try and export a workflow which results in black image and run that in ComfyUI directly? I'd expect output to be black there too, but it might be some specific node/combination causing it.
Can you try and export a workflow which results in black image and run that in ComfyUI directly? I'd expect output to be black there too, but it might be some specific node/combination causing it.
I have the exact same problem. I did what you asked for and it resulted in a black image too.
change the sampler to "sd3-dpm++ 2m uniform" can perfectly fix this issue. seems it just being a sampler issue
change the sampler to "sd3-dpm++ 2m uniform" can perfectly fix this issue. seems it just being a sampler issue
Yes, that solves the problem. Thank you very much.
I am "kind" of having the same problem: black images in files that a week ago were rendering fine.
MacOs 14.6.1 - M1 128Gb
I cant figure out what made the difference because almost at the same time, i updated OS (to 14.6.1), I updated Krita (to 5.2.3), I updated the ai-diffusion plugin (to 1.22.0) and i installed the flux models (flux1-schnell-fp8.safetensors flux1-dev-fp8.safetensors)
Everything seem to work OK and i was able to play with the Flux models for a week. Today as i came back to a project i am working on using SDXL (juggernaut), every image came back black!, i read this thread and after some experimentation i found that changing the samplers does work, but only if i stay below 20 steps.
Lets say i use a XL model, with the default sampler (DPM++ 2M -Karras cfg:7 steps:20): i get a black image but if i change that to 19 steps, it works fine. (I did the same in the ComfyUI directly (http://127.0.0.1:8188/) : same thing: black images with anything above 20 steps).
I thought this was a XL issue, but not, its exactly the same with 1.5 models. 19 steps fine, 20 steps black image. And this happens with any combination of samplers (euler ancestral -normal, DPM++ sde-gpu -Karras, etc...) black images if i jump from 19 to 20 steps.
Funny enough if i come back to flux, there i can jump to any number of steps (euler-simple) 60 steps!, and it works fine.
This looks related: https://github.com/comfyanonymous/ComfyUI/pull/4548
If I read it correctly you can try providing --force-upcast-attention to ComfyUI startup
If you are using manage server install: How to pass arguments
If I read it correctly you can try providing
--force-upcast-attentionto ComfyUI startup
Thanks a lot That worked! Ive been working like an hour with 1.5 and xl models and no more black images, with any number of steps.
If I read it correctly you can try providing
--force-upcast-attentionto ComfyUI startup
Works! Thank you
Hey guys,
I tried to expand a picture but outside the original border, it only prompts black. I've tried every solution in this topic but it didn't work.
Any ideas ?
Hey. For the next two weeks I'm on holiday. If the issue is still going on after my break I gonna export a workflow for you. --Gesendet mit der WEB.DE Mail AppAm 05.07.24, 10:08 schrieb Acly @.***>:
Can you try and export a workflow which results in black image and run that in ComfyUI directly? I'd expect output to be black there too, but it might be some specific node/combination causing it. —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you commented.Message ID: @.***>