I'm so sorry you ran into that, and thank you for reporting it. This is exactly the kind of feedback I need at this early stage.

You're right, my backend logs show that most requests are succeeding, which means there must be an error happening somewhere between the front-end and the server that I'm not catching properly yet.

Based on this, implementing a more robust error logging system is now my top priority. I'll get on it right away so I can find and fix these issues for everyone. Thanks again for giving it a try.

Thanks! Not yet. To get this launched quickly and validate the idea, I'm currently using a cloud API to handle the inference.

However, my plan is to eventually deploy the model on my own server. I'll be sure to document the entire process—from setup to optimization—and share it as a detailed guide on the site for anyone interested!

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

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?

That's a great question. Right now, the inference costs are coming out of my own pocket. My main goal was to let people experience the model's potential without any barriers.

To keep the project sustainable in the long run, I'm exploring some options, like potentially offering a paid tier for heavy users or more advanced features. For now, I'm focused on improving the core experience and will do my best to keep costs low so it remains accessible to as many people as possible.

Glad you liked the first result. And ugh, sorry you hit that error on the second go. You're right, that "unexpected error" message is pretty useless. Better error handling is at the top of my to-do list now.

> These models look fantastic, we've finally got something solid in the public sphere that goes beyond stable diffusion style word vomit for prompting.

Stable diffusion models later than 1.x do that, even, e.g., SDXL finetunes that are heavily trained on supporting controlled vocabulary tags for precision support (and benefit from) natural language prompting; and many of the newer “open” models (many of which are a better approximation of open than Flux) even use the same text encoder as Flux (and some use LLMs like Llama).

BFL is really good at promotion, though; it would be nice if open models with sinilar functionality like Omnigen2 got a fraction of the attention non-open Kontext gets.