Okay lads, this thing tries to remove blur, especially the light one given by Flux models, DOES NOT work on hypermegablurried images, huge artifacts/noise or stuff with a strong strong bokeh effect.
You prolly want to rebuild the image from zero, and it can't understand what it's looking at if it's just a splat of colors, rite?
TESTED ON:
flux kontext dev Q8/ flux kontext nunchaku
1-2 lora strenght, but feel free to experiment
2.5-3.5 CFG, but again didn't test much
20 steps euler simple seem decent, also tested with nunchaku plus turbo lora at 8 steps, worked unexpectedly well!
-Sometimes- describing what you're deblurring helps, meaning you can prompt:
" deblur, hand "
When the blur is quite strong it might give weird results, but a bit crispier, so maybe with a later upscale you can get something decent out of it, ye?
A hint: if your first deblur worked but it's still a bit blurry, don't just crank up the strenght, copy the new image, and rerun with the new image as a starting point!
BUZZZZZZZZZZZ if you use it and like it, thank you!
Description
Credits to lym0 for providing the fixed model and akedia for the patch about Nunchaku's Lora issue!
FAQ
Comments (26)
My nunchaku still can't use thisLora——'lora_unet_double_blocks_0_img_attn_proj.alpha'。I am using the int4 model
are you using the dedicated nunchaku lora loader?
update nunchaku nodes and make sure you downloaded the v1.1
- i use nunchaku and it work fine with me
I am using the dedicated nunchaku lora loader,It doesn't work
but using the lora loader it's ok@heyheyheyheyhii
@kaikaikid it's unfortunately really hard to check without having the PC under my eyes, the only fixes I can think of are:
- Update comfyui (through update all)
- Go to the manager and search for nunchaku nodes, then update
- Make sure you're using the nunchaku version 1.1, the file is called "FLUX_kontext_deblur_nunchaku.safetensors" and NOT "FLUX_kontext_deblur.safetensors", because that's the older version, non compatible with nunchaku.
Other than that it's out of my hands...I have feedbacks from other people using it and it works correctly, so I'm not sure what could be your specific problem.
u r the BEST <3
thank you! @lym0 found the solution, I just updated the lora, and the patch comes from Akedia :) guess that's the beauty of open source right?
You are so hardworking, bro. I just downloaded the FLUX version yesterday and the Nunchaku was released today.
thanks mate, but check the details of the model on the right, it was fixed by another person, I just reuploaded it :)
@heyheyheyheyhii OK
you need to try to fix that artifacts specially for low res outputs , wait your next improved version.
I'm personally not having many artifacts, but when I do they can easily be fixed with a simple hires fix workflow and an extremely low denoise, it will keep most of the details.
the lora is trained with hires images, and a higher than standard amount of steps, so there's not much that can be done on my side except for increasing the number of training data images (which unfortunately costs a lot and doesn't guarantee results at all).
Kontext, the base model, works well on resolutions revolving around 1024x1024 or similar.
if you try it with much lower or much higher resolutions you can't expect it to work well, especially with Loras that are quality related (sharpening, blurring, exposure change, color shift etc..).
That is out of my control, the best solution is to create a workflow that upscales or downscales your images to a correct resolution, use Deblur, and add a last step with hires fix.
I'm using a similar one and the results are really good.
@heyheyheyheyhii stacking beside other loras can reduce the deblur result quality?
@amazingbeauty I'd use deblur by itself, afterwards stack a new ksampler with hires fix.
If you haven't enough vram just pull the result from deblur once it's finished and pass it through a new hires workflow.
I tried deblur with nunchaku + the turbo lora at 8 steps and it's still a good result to be passed on to a hires fix workflow later.
The examples in the post were made like this on a 3090 in less than 30 seconds each!
I'll try to post the workflow when I have some time.
This is amazing and works great - thanks!
thank you!
Do u have a workflow for this. New to this. Not sure how to use
for the standard version ( not nunchaku ) it's just the standard flux kontext workflow with the Lora between model and ksampler, and in the clip text just write the word "deblur"
does this work with nunchaku ?
yes, the latest 1.1 version does!
heyheyheyheyhii what changes have been made compared to the first version?
heyheyheyheyhii what changes have been made compared to the first version?
Deniaud some training methods give a result that's not compatible with nunchaku, a person here found a quick fix for it and I uploaded it, it's really just a few lines of code (which I don't fully understand tbh)
@heyheyheyheyhii I wonder where I can read them? I didn't find them in your profile
Good job, works very well
thank you!