hidream: 8b parameters, no vae, #1 open-weight image model

hidream-o1-image-dev-2604 is ranked #1 across all open weight text-to-image models on the artificial analysis text-to-image arena

what makes this model interesting is how it's built. many other image models (for example gpt image 2) use a vae, a separate component that compresses images into a smaller latent space before the main model works on them. hidream skips that entirely. it works directly on raw pixels using something called a pixel-level unified transformer. text, images, layout instructions, everything goes into one shared token space

it's also only 8 billion parameters. to put that in perspective, flux.2 dev uses 24 billion parameters plus a 32 billion parameter text encoder, 56 billion total. qwen image uses 7 billion plus 20 billion. hidream achieves better or comparable results with just 8 billion total, no extra encoder, no vae. that's unusually efficient

it handles multiple tasks in one model. text-to-image is the main one, but it also does instruction-based image editing, like "remove the earphones" from a photo, and subject-driven personalization where you give it several photos of a person or object and it places them in a completely new scene. it can render long passages of text accurately in images, in multiple languages, which most models struggle with. it generates natively at up to 2048 by 2048 pixels

the model and code are released under the mit license, so commercial use, modification, and redistribution are all permitted