Iris 3B

Iris-3B is a 3B-parameter flow-matching transformer from Sperid Labs that generates raw RGB pixels directly, with no VAE in the loop. Images are split into 16×16 patches and processed by 8 dual-stream blocks (separate text and image weights) followed by 16 single-stream blocks. A small pixel-transformer head decodes patch tokens to pixels. Text conditioning comes from a frozen Qwen3-VL-4B encoder, and 2D RoPE lets one set of weights handle any resolution and aspect ratio up to one megapixel.

The project also includes a Turbo variant distilled to 4 steps with Trajectory Distribution Matching (data-free, prompts only), and downstream fine-tunes for dense tasks such as depth estimation and image restoration/super-resolution. The authors frame it as a general vision learner, arguing that image generators can replace vision backbones such as DINOv2.

On the authors' reported benchmarks (official evaluators, 1024², no prompt rewriting), Iris-3B scores 0.798 on GenEval, 86.52 on DPG, 0.857 on LongText and 0.540 on OneIG; Turbo scores 0.820, 86.68, 0.804 and 0.516. This is behind larger models such as Qwen-Image 20B and Z-Image 6B on several metrics. A YouTube reviewer describes it as Apache 2.0 licensed. The provided pages did not list a GitHub or Hugging Face link, so the exact download locations and license text are unverified here.