Provide workflow example

[sneccc]

Hello found your repo, looks nice , is there any example workflow?

[blepping]

sorry, i don't have an example workflow published right now because i'm still frequently making changes to the project. it is relatively straightforward to use though.

if you don't connect sigmas or samplers, it'll do 15 steps of Karras at 0.3 denoise with Euler. if you don't connect a reference image, it will generate one for you based on the latent you connect to the sampler. for example:

this will generate an initial 512x512 reference image (via 20 Karras steps), then do two 2x upscale passes afterward - so 512x512 to 1024x1024 then 1024x1024 to 2048x2048.

as for what to try other than the defaults, i have some suggestions in the main README. most of my testing is with SD15, i like something like this:

you can also use iteration overrides (see the README) to do stuff like switch schedules/parameters. i've found using an upscale model seems to make the most difference going from low res to medium and it's less of a benefit when you're already at a decent resolution: it's possible to use the upscale model just for the first iteration (possibly also image guidance) and then switch to simple scaling/latent guidance for subsequent iterations. you also can do more iterations with a lower scale factor.

[bkutasi]

I tried to make it work with my workflow so here is my take, might help others.

I use DMD2 lora with the euler a cfg++ sampler, i tried GITS and AYS scheduler, but there were no differences between them.
I think it's a good implementation, but depending on the model/settings the quality is lower than with HiDiffusion. Here are my observations:

• Speed is not great, due to multiple passes. Depends on the steps/config you use
• Better compatibility. As I mentioned, HiDiffusion is faster but doesn't work with a lot of models.
• Compared to doing highres fix(efficiency nodes) its comparable in speed for single images
• You need a lot of vram, because it cannot do batch images, each image is individually handled
• It gives a great upscaled image, unlike HiDiffusion which completely changes the attention mechanism therefore the composition(if using RAUNet)

[blepping]

@bkutasi Thanks for the interest and feedback!

Compared to doing highres fix(efficiency nodes) its comparable in speed for single images

It basically is an enhanced hires fix so this makes sense. :) Personally I recommend generating the initial reference separately so you don't waste time upscaling/processing a generation with significant flaws. Also, the guidance is a bit of a double edged sword in that case: you get better adherence to the initial image in your high-res generation but that also means there's less chance the high-res pass will fix existing flaws. So it's pretty important that the initial reference image be as close to perfect as you can manage.

You need a lot of vram, because it cannot do batch images, each image is individually handled

I don't think that's correct. It handles batches normally, but a batch of high res images is obviously going to multiply the memory/compute requirements. What you might want is something that doesn't do batch images and instead splits them up and processes them serially. You can do that with ComfyUI's execution inversion stuff (I haven't really used it myself so I can't give you an example) but it isn't really too user-friendly.

It gives a great upscaled image, unlike HiDiffusion which completely changes the attention mechanism therefore the composition(if using RAUNet)

You're right about it changing the composition (and generally, you get a completely different image) but it isn't because of attention. HiDiffusion has the MSW-MSA attention component but at least with my nodes it is something you can enable separately and isn't included with RAUNet. It also doesn't really change the generation dramatically.

HiDiffusion, like DeepShrink AKA Kohya Hires Fix scales down some layers in the model early in the generation and then lets the effect end once major details are established. By default you wouldn't normally be targeting the attention layers (but it is possible by manually setting those blocks).