Inversion

[filipstrand]

Testing out some inversion techniques. WIP at the moment and will have to be properly integrated later. Also have a WIP PR up for FlowEdit. Still evaluating them this stage...

Been looking at two techniques so far mentioned in https://github.com/logtd/ComfyUI-Fluxtapoz?tab=readme-ov-file:

• RF Inversion (simple, does not need any architecture change, have tested this briefly in local branch, a bit mixed results so far, more testing required...)
• RF Edit - currently in this branch (adds higher order term to the noise schedule (+1 extra transformer call for each step - and if I have not made any mistake the Algorithm 1 in their paper simplifies a bit to simply the "noise2" term in this branch), need to save the inverted attention V-values in a cache for a later forward generation). Pretty good results on some images so far