See original GNRI repository.

Editing Images using Text

Observe the similar pose of the foreground animal and background of two images.

"a lion is sitting in the grass at sunset"	"a dog with a funny red hat is sitting in the grass at sunset"	"a raccoon sitting in a field of cherry blossoms, photorealistic"

My Changes

• Fixed various errors to get RealTimeEditingNotebook.ipynb working (import, device errors)
• Updated GNRI algorithm to latest version of paper
• Fixed CFG error, allowing editing with guidance scale > 1.0
• Fixed inversion not actually inverting to the final timestep
  • (this may be an SDXL-specific thing, not sure if this is actually a bug)
  • Appears subjectively to give better results on a small number of test runs
• Additionally separated inversion/ inference to two GPU's because mine aren't large enough to fit SDXL on one GPU
• Implemented GNRI with DDIM Scheduler in addition to Euler Scheduler