Flux1 Redux+Fill like LayerDiffusion
10.11.2025, updated a json built on Nunchaku SVDQ-FP4.
Once upon a time... though it's not like it was that many years ago,
(though it's scary how in this world it feels like a long-forgotten past)
There was a feature called LayerDiffusion that could only be used in Forge and ComfyUI, but to put it bluntly, it didn't finish with satisfactory image quality, and before I knew it, it faded away (which is also scary), and I feel like it has completely vanished into oblivion.
Regarding the point of cutting out only the subject from an image and performing post-process background compositing, I think IC-Light actually had more usable performance.
...Now, about Flux Fill, I initially intended to use it for things like hand correction or other mainstream applications, and I was planning to test it first... but before that, I learned how to build nodes that can be used like LayerDiffusion by combining it with Flux1 Redux, so I built that first.
As usual, the source material is from OpenArtAI. Although I modified the settings for the LoRA and prompt generation parts to my usual style.
Nouvo.ai... seems to be from Korea, but I learned yet another new way to use Florence2... like, so that's how you use it, sort of.
This background removal process part... this is basically doing the same thing as transparent-background or BiRefNet,
to be honest, you would get the same result using those nodes, but I was like... I see, so you can do the same thing with Florence2.
As expected, since it's a Flux1-based program, it finishes beautifully. It's a significant difference from the old LayerDiffusion....
Even with IC-Light, which runs on an SDXL base, if the original image is Flux1, you get almost the same image.
Since I couldn't reproduce it on my own, I won't be releasing it for free here, but with the IC-Light Ver II created by the ever-familiar ControlAltAI, if you assume the original image is created with Flux1, you can perform almost the same compositing.
As always... What is thissssss!? How could I ever imitate thissssss!!...is the kind of setup they create. Moreover, the image below doesn't even show the whole thing. The full picture is even longer.
Even among the support members who are provided with the json, there are people dropping out because they can't reproduce it (probably because they can't obtain the models or configure the individual nodes).
When you use the ControlAltAI version Ver II, it looks like this, but since IC-Light inevitably has to use SD1.5 in its process, it has to concede in terms of background detail quality compared to this program, which can generate everything in Flux1.
In exchange, IC-Light has the advantage of being able to add lighting effects.
