Scaled as in lowered?

(March 7, 2024, 8:45 pm)Mothman125 Scaled as in lowered?

Ok bad wording, I mean rescaled cfg, it's an option that starts the cfg scale at one value then increases it to a final value over the course of the step count.

For example if I set the main cfg value to 7.5 and the starting value as 1.0 (sometimes called the 'min cfg"), the first step will adhere more to the original image with the value incrementing at each step such that the final step will be 7.5. So it gradually adheres more to the prompt at the later steps which can add greater levels of detail without adding extra weirdness.

Whole thing is really still experimental.

Finally getting some decent results on a Comfyui setup... Sometimes. How many steps are you using? Just want to make sure I'm not wasting time.

Just the defaults in the comfy node, didn't find much benefit going over 45 steps when I tested.

I also took their base positive and negative prompts used in their demos.
Postive:

Quote:Cinematic, High Contrast, highly detailed, taken using a Canon EOS R camera, hyper detailed photo - realistic maximum detail, 32k, Color Grading, ultra HD, extreme meticulous detailing, skin pore detailing, hyper sharpness, perfect without deformations.

Negative:

Quote:grain, painting, oil painting, illustration, drawing, art, sketch, oil painting, cartoon, CG Style, 3D render, unreal engine, blurring, dirty, messy, worst quality, low quality, frames, watermark, signature, jpeg artifacts, deformed, lowres, over-smooth

Then added my descriptions of the image at the end, which I sort of jump between writing manually, or just running through the LLaVA hugging face space demo.

I mean sometimes all it takes is a seed reroll, other times a tweak of the cfg depending on the image.

Problem is what works for one image, might not work for another.

The amount of detail it adds to blurry noisy images is pretty incredible (left original -> right upscale):

       

       

            (You can see the difference here between cfg 4 & 7.5, more detail at higher level but at the cost to some weirdness)

A lunch time generation (seen a low res bit from a video I needed enhanced)

Original:
   

SUPIR:
               

It's extremely good at detailing belly buttons

I've noticed simple pixelation is often not enough. Grain and compression certainly gives the thing more to work with.

I believe the SUPIR model (which is a controlnet in a sense) was trained on super resolution images with various degradations applied, so it's likely better at some more than others.

Technically reducing the control scale slightly can get around that if it's retaining too much of the image, at the cost to it not completely following it. Grain itself can sometimes cause issues which makes the image look either blurred at lower cfg, or dirty/blotchy at higher cfg.

Changing the resize method might also net some different results.

I tried to look up what the control scale does exactly and didn't easily find an answer.