Introducing Pixeled, a Flux.2 Dev finetune meant to create high quality images using descriptive prompts. It has been trained on thousands of highly detailed prompts in a specific format so it can understand bigger prompts with big detail. More info on the model card for retronic/pixeled-image.
It has been a while since I have posted, and I have a project that is already up and running on HF spaces, which is called Pixelship. Pixelship uses an agentic like structure kind of like an artist. It uses GPT 5.4 to create a 16x16 image for a bigger understanding, and then it gets sent to Flux Kontext Pro. There is a validator that always checks for accuracy and perfection. If all of it just doesn't pass, it will set up a score for each image it makes and revises. It is available on retronic/pixelship, and I would love to hear feedback on this project.
UPDATE ABOUT COLOX: I've intended to train Colox two times, the following will show a list of how many times I planned to train it: one with an open source database, then another with personality, safety filter, info about itself, and who created it. And plus, I am putting in a bit more info in the training so it knows a bit better. And I am happy to announce the first training is complete! I am making all models made from the two trainings open source, and the final training is the official release. NOTE: I HAVE NOT TESTED THIS YET, THIS MODEL MAY NOT REASON AND IF THAT IS THE CASE, I WILL FIX IT.