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.
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
One of the issues with staggered releases is they are easy to forget about. BFL have apparently made the weights to the Kontext Dev model available, but because it didn't happen when they announced Kontext I'd compeltely forgotten about it.
These models look fantastic, we've finally got something solid in the public sphere that goes beyond stable diffusion style word vomit for prompting. It was obviously coming sooner or later, but happily it seems to be here. It is unfortunate for the public that, as far as I can see, they didn't actually open the weights up since they aren't free for commercial use.
I was able to try it with one image, I liked the results and wanted to try something more complex. However, I started getting error messages: "Processing Failed An unexpected error occurred. Please try again.", retrying didn't help.
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.
So, are you like a multi-millionaire's kid or something, how are you paying for / able to afford giving away public access to this? This is written like it's an individual doing this alone, not the case?
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.
> Restore and colorize old photos with AI precision
I’ve got to admit, I chuckled to myself at the absurdity of the phrase “AI precision”, given how badly these things are known to go off the rails. Sure, sure, things have improved a lot in the last few years, and Kontext’s limitations make such problems far less likely to occur, but still, permit me to be amused. :-)
… but then too, do compare https://fluxkontextlab.com/pages/home/showcase/2/1.jpg and https://fluxkontextlab.com/pages/home/showcase/2/0.webp closely, there are material differences. A few of the most notable ones: the picture is reframed, with a significant amount invented at the bottom (which has realism concerns that you can see when you actually examine it); fog effects have been reduced (perhaps implied by “restore … its clear texture”, which seems a weird instruction to me); and something’s gone wrong with the right wing of the pigeon at the bottom that’s facing the camera.
I think it would be nice to, in each case, align the two as well as possible (even the Product Display example) and present them in such a way that you can rigorously compare the beginning and end points, and see what modifications have been made, intended and unintended.
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.
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!
If you click on the time next to the username, you get a dedicated page for that comment with additional options: | parent | context | flag | vouch | favorite | ... but you need some minimum karma for the "flag" and "vouch" options. The unofficial FAQ says 31: https://github.com/minimaxir/hacker-news-undocumented?tab=re...
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.
https://fal.ai/models/fal-ai/flux-pro/kontext/max/multi
[1] https://github.com/timothybrooks/instruct-pix2pix
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
One of the issues with staggered releases is they are easy to forget about. BFL have apparently made the weights to the Kontext Dev model available, but because it didn't happen when they announced Kontext I'd compeltely forgotten about it.
These models look fantastic, we've finally got something solid in the public sphere that goes beyond stable diffusion style word vomit for prompting. It was obviously coming sooner or later, but happily it seems to be here. It is unfortunate for the public that, as far as I can see, they didn't actually open the weights up since they aren't free for commercial use.
I was able to try it with one image, I liked the results and wanted to try something more complex. However, I started getting error messages: "Processing Failed An unexpected error occurred. Please try again.", retrying didn't help.
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.
Keeps saying "something went wrong" for a variety of images / requests. Can't say I'm hugely impressed. MidJourney does it flawlessly.
So, are you like a multi-millionaire's kid or something, how are you paying for / able to afford giving away public access to this? This is written like it's an individual doing this alone, not the case?
I tried a picture with instructions and it says "something went wrong". I would love to try and see how well it works for my use case.
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.
> Restore and colorize old photos with AI precision
I’ve got to admit, I chuckled to myself at the absurdity of the phrase “AI precision”, given how badly these things are known to go off the rails. Sure, sure, things have improved a lot in the last few years, and Kontext’s limitations make such problems far less likely to occur, but still, permit me to be amused. :-)
… but then too, do compare https://fluxkontextlab.com/pages/home/showcase/2/1.jpg and https://fluxkontextlab.com/pages/home/showcase/2/0.webp closely, there are material differences. A few of the most notable ones: the picture is reframed, with a significant amount invented at the bottom (which has realism concerns that you can see when you actually examine it); fog effects have been reduced (perhaps implied by “restore … its clear texture”, which seems a weird instruction to me); and something’s gone wrong with the right wing of the pigeon at the bottom that’s facing the camera.
I think it would be nice to, in each case, align the two as well as possible (even the Product Display example) and present them in such a way that you can rigorously compare the beginning and end points, and see what modifications have been made, intended and unintended.
Cool, but how are you paying for the inference if it's free?
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.
Cool. Are you running the model on your own server?
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!
[dead]
[flagged]
What is this? How does one report such a comment?
If you click on the time next to the username, you get a dedicated page for that comment with additional options: | parent | context | flag | vouch | favorite | ... but you need some minimum karma for the "flag" and "vouch" options. The unofficial FAQ says 31: https://github.com/minimaxir/hacker-news-undocumented?tab=re...