Is there a way to make InitImage equivalent Img2Img in A1111/Forge?

[10061996]

When I first tried SwarmUI as an alternative to A1111, I was very satisfied with what Swarm offers that A1111 has not, starting with a much more comprehensive model support. I would gladly ditch A1111 completely if there wasn't one special thing that I am missing in SwarmUI.

My main subject are animal pictures and I love to generate them with Flux or Flux finetunes because those follow prompts much better than SDXL. However, Flux and its finetunes are bad at animal fur, so I apply Img2Img to the Flux outputs using a SDXL model like Juggernaut which turns the artificial texture of Flux-generated animal fur into something almost photorealistic.

However, I still use A1111 for that because the SwarmUI variant of Img2Img (InitImage) produces far worse results than Img2Img, lacking depth and texture. The algorithms just seems to work differently. That applies also to the Swarm version of inpainting.

Are there any plans to improve InitImage in a way that it matches Img2Img qualitywise? Or is there a chance that the Img2Img algorithm be implemented in Swarm as an alternative to InitImage? By now, I still (have to) use A1111 for everything that needs SDXL I2I or inpainting. I'd love to get it all done in Swarm and finally get rid of A1111.

[mcmonkey4eva]

The primary relevant difference is some UIs miscalculate steps intentionally. So for example if you have Steps=20, Creativity=0.5, Swarm will skip 10 steps and run 10 steps. Some UIs will skip 20 and run 20. Or if you have Steps=20, Creativity=0.25, Swarm will skip 15 and run 5, whereas some UIs will skip 60 and run 20. So if you're used to that behavior and want to replicate it, hyperboost your steps value weirdly high.

The other relevant difference is just, when inpainting, the advanced parameter Mask Behavior is set to Differential by default in SwarmUI, while it in auto would be equivalent to 'Simple Latent'. (This is defaulted that way because differential is almost always better.)

There are a few other minor differences that apply generally (not specific to image editing), such as: ModelSpecificEnhancements parameter, Auto I'm told has a CFG Scale that's off by 1, Auto also parses prompt weighting slightly differently.

[10061996]

Thanks for the comprehensive reply. I have tried boosting the number of steps for InitImage but to no avail. Makes sense because A1111 also adapts the number of steps to the denoising setting and still yields better results. I have tried to look deeper into why I like the I2I results of A1111/Img2Img better than those of Swarm/InitImage and found out that the results of the latter are a lot less sharp than those of A1111, which I wasn't fully aware of before. A direct comparison is difficult because the creativity setting of Swarm does not exactly match the denoising parameter of A1111 but if I strive for a visually similar degree of image alteration with both UIs, the results of A1111 are always sharper. It seems as A1111 has more success in adding detail to the I2I output, but this is just a perception from the visual appearance of the outputs compared to each other.

[mcmonkey4eva]

To be clear, there's no inherent difference between the UIs for the most part, other than supported feature set (which Swarm has a wider range there anyway). The differences are all down to (1) parameters and parameter interpretation, and (2) random luck / seeds.

I've seen a lot of cases of "wow UI 1 is way better than UI 2"... and then go generate 10 images with same params on both, mix them up, and try to guess which UI made which image and oops actually you can't -- the difference was just random seeds and bias.

The cases where there is actually a consistent and clear difference, it's down to parameters. I've listed above the 3 main parameter differences that occur with a relatively default setup.
You do of course also have to align other parameters (sampler/scheduler selection, resolution, etc.)