Comment by roenxi
7 months ago
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.
> 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.