ABOUT convert multiple views to a single photo
I hope it can enhance the ability to convert multiple views to a single photo. please.
Currently, V1.2 can only try to avoid generating multiple people in the same frame (when multiple views are passed in) by reducing grounding_px and ref_boost. My current solution is to use two sampling methods: first, use relatively low values to obtain a suitable composition in the first sampling, and then use a second sampling to obtain a better similarity.
AND...My English is not very good, so this part was mainly translated using translation software. I hope you can understand it.
Thanks, and that is a smart two-pass workaround, low values first for composition then tightening up. Fusing multiple views into one coherent photo is not something the current model was trained for, so what you are doing (lower grounding_px and ref_boost to avoid it spawning multiple people) is genuinely the best lever available right now. It is on my list of things I would like to support properly in a future version. Really useful to see your approach, thank you for sharing it.