It will remove blur (...like even heavy blur and magically figure out what was in original image, wtf).
Can be used also for upscaling but it was trained only on gaussian blur and not resized images (from lanczos) so will probably make separate lora.
It does not remove blurry background/bokeh (for this check this lora) or partially blurred images.
Prompt is Wibble snargle floof - yup, random nonsense.
Workflow in images use this turbo lora Viggle turbo but for better quality do 20 steps no turbo, cfg 2-4
Description
FAQ
Comments (21)
can it sharpen photos with shallow focal length (where part of the image is in focus, and other parts are soft), or does it only deblur guassian blurred images?
Most anticipated CSI enhance lora :D
Very cool, thanks for sharing!
great lora but horrible trigger, somtimes creating fur fluff woal stuff on photos
oof, yea you can try something like "fioof" or no prompt
@Luntrix Wibble snargle floof , this was the prompt i used, it happens more in photos showing back, but still worst triger messes with some images
Were the sample images part of the training set? Kinda hard to believe it'd get facial features this accurate from that level of blurring.
of course not
This trigger is cumbersome and poorly written. The official Qwen Image 2.1 PE prompt-expansion model cannot understand it at all. I hope it can be triggered with natural language in the future. Thank you for your hard work!
"prompt-expansion model cannot understand it at all" - that's the point. Current "remove blur" is doing "something" and I want to avoid this "something".
@Luntrix LuntrixD3B$R
how many seps work best for you and sampler?
Just make a prompt that combines two things and put it at the very front, that's all.
Isn't the trigger word needed for LoRA training supposed to be an unambiguous phrase that barely exists in the model, and it can't conflict with existing concepts..
佬，什么是“呜呜呜，毛茸茸的？”
Great, now we can finally get some clear UFO pics!
Thank you, Sir!
I cant get this to work at all images are output almost identical to the input. I also used your workflow used a blurred image and set the blur to zero and again output looks like the input?
Your image must be too blurry or have non supported blur type. Do some tests with sharp image and comfy blur (like blur radius up to like 16 max).
does it support comfyui?