hi-res for txt2img and i don't like it

[dominikmau]

I had a brief conversation about the hi-res feature and I've been bothered by it for a while.
First of all, I like the new design and functionality of hi-res - but what do I want with it in txt2img, I need the crap in img2img. (with all its functions - sexy latent antialiased)

So far i generate batch e.g. 100 images 512x768, look through them if there is an image i like and worth to upscale - then i send it back to txt2img and generate the same image with hi-res 1024x1536. (after that i usually scale it again with ultimate sd upscaler in img2img etc.)

In img2img there are so many more possibilities:

• change/customize the prompt
• load another model, another HN etc.
• change values
• change the seed and make several upscale until it 'fits'

Of course img2img is nice even now, but it could be a nicer.
On the other hand i can't see a real advantage of txt2img hi-res over img2img.

a) generate 20 images with hi-res without knowing if the images are without errors/deformations or if you like the image at all, etc.
(you eat what you get)

b) 100 images and you pick the handful you like from the appearance and scale them up yourself with a hi-res function in img2img which has the same functionality as in txt2img.
(you eat what you choose)

Is my thought that far-fetched?
I really celebrated when latent upscaler came for img2img because that's where it belongs for me. Hope the rest comes soon too.
(may be i am just imagining it, but for the first upscale i find the hi-res function in txt2img almost always generates better results than img2img)

[freecoderwaifu]

I mean, for me it does exactly what it says, it gives a higher res image and it fixes it. Some models and styles really need and benefit from a second pass to fix wonky details, and the highres pass does that. That's usually more than enough for me 99% of the time to produce a good image. For more control or harder generations there's always img2img and inpainting.

[mohitsrivastavazzz]

I mean I agree but its not even that far from using 0.3 denoise

[midihex]

Wouldn't be able to make this in txt2img without it

[ukaserge]

Wow, this looks amazing, is it possible to share prompt and model name?

[Centurion-Rome]

See also "Bring back old Hire.Fix": #6725

[masslevel]

@dominikmau You're right that it could be designed differently, but we have at least a couple of more options now.

I use Hires fix. maybe a bit differently. I have been using Hires. fix for almost every image since SD2 launched. I mostly generate images in 1024x768 with DPM++ SDE Karras (30 Steps, 7.5 - 9 CFG).

• My 1st pass image is set to a low resolution. 592x448 in this case, because this is how it behaved in the old highres. fix with a wrong aspect ratio and everything, denoising strength at 0.7.
• This results of course in an unusable very low-res 1st pass image. But this image acts as a guide / init image for the 2nd pass.

The fidelity and quality that I'm getting as a result is something I have not been able to reproduce otherwise.

This means the Hires. fix image is my actual real resulting image and not the 1st pass without it.
This method served me very well so far.

[medledan]

<On the other hand i can't see a real advantage of txt2img hi-res over img2img.>

I'm glad I read this post. If I did a txt2img at 768x768. The image takes around 4-7 seconds to render, depending on sampler. If I turn on highres.fix, and punch that to 1536*1536, the damn thing takes around 1 minute 17 seconds.

If I use the same seed, take the original 768x768 and popping it into img2img, and adjust the res to 1536x1536, use Heun at 30 steps -> I get an image back in around 22 seconds. If you compare the highres.fix image and the img2img.....i get a BETTER quality image with img2img than highres, at almost 3x the speed!

[mohitsrivastavazzz]

Hi-res fix regenerates the base image afaik, and in img2img there's already pixels to work with so its faster and better. img2img is pretty good.

[helluscorus]

img2img does less steps depending on denoising strength value unless you changed the With img2img, do exactly the amount of steps the slider specifies (normally you'd do less with less denoising). setting.