black-forest-labs / flux

Official inference repo for FLUX.1 models
Apache License 2.0
13.48k stars 949 forks source link

Endlessly generating blurry images #65

Open QL-boy opened 1 month ago

QL-boy commented 1 month ago

It's a great model but it's also screwed up, it throws me a lot of accurate but very blurry images,There seems to be some tokens that have been mis-learned. ComfyUI_00113_

cibernicola commented 1 month ago

Same here, not sure why/when/wich config(s), but it happens.

dgstewart commented 1 month ago

I was experiencing a very similar issue. After some testing I found that modifying the prompt by removing the phrase in front of a solid white background my images are no longer out of focus/blurry. Based on the image you posted it looks like you may have been trying for a similar effect.

CeterPushing commented 1 month ago

very similar issue here as well using use a white background to enhance the contrast and make the elements stand out makes the images blurry to varying degrees.

cibernicola commented 1 month ago

I was experiencing a very similar issue. After some testing I found that modifying the prompt by removing the phrase in front of a solid white background my images are no longer out of focus/blurry. Based on the image you posted it looks like you may have been trying for a similar effect.

reproduction of the "error"(?)

image

agustincaniglia commented 1 month ago

I am having the same issue. Are you running on forge? I am also now. I don't know what's triggering but its giving me blurry images even if I remove things like "white background" now.

cibernicola commented 4 weeks ago

Here I'm using comfyUI, its not only with white background, I can't remember exactly other "prompts"(?), but it seems related to prompt input(?)

HarshVardhan-star commented 2 weeks ago

just use a negative prompt "blur, defocus"

jeroenrijckaert commented 2 weeks ago

It would be nice if that would be possible, but Flux does not work with negative prompts. You need to prompt carefully. It seems flux is trying to divide the text into foreground and background, from which either of the two - but not both - get the focus in the resulting image. A larger portion of the text describing background stuff, will result in focus on the background. If the background is solid, you won't even see it and only see the blurred foreground.