Comment by vunderba
7 months ago
Kontext's ability to make InstructPix2Pix [1] level changes to isolated sections of an image without affecting the rest of the image is a game changer. Saves a ton of time without needing to go through the effort of masking/inpainting.
About a month ago I put together a quick before/after set of images that I used Kontext to edit. It even works on old grainy film footage.
https://specularrealms.com/ai-transcripts/experiments-with-f...
> My goal is to share what this model is capable of beyond the basics.
You might be interested to know that it looks like it has limited support for being able to upload/composite multiple images together.
This is a great demo! Kontext's precision with localized edits is impressive – especially how it handles grainy footage without artifacts.
Your multi-image compositing experiments reminded me of how we built https://flux-kontext.io/ to solve a similar problem: enabling real-time collaborative AI edits where multiple users can tweak different image sections simultaneously while seeing live previews. The context preservation feels almost like magic when you see it in action.
Would love to compare notes on your masking-free approach – we've found that combining InstructPix2Pix-style changes with layer-aware diffusion (like in your film example) reduces hallucination by ~40% in our tests. Any plans to open-source the training pipeline?
I had a project for a big brand a couple years ago where we experimented with genai and inpainting and it was a huge hassle to get it working right, required a big comfy pipeline with masking, then inpanting, then doing relighting to make it all look natural, etc.
It's crazy how fast genai moves, now you can do all that with just flux and the end result looks extremely high quality