Iris-3B through EmbeddingGemma 2

Iris-3B (pixel-space, no VAE) with its Qwen3-VL-4B text encoder replaced by EmbeddingGemma 2 (0.74B) and a 105M adapter: oakoliver/eg2-iris-3b-adapter, Apache-2.0, trained on CC0 / CC BY data only. Text alignment is statistically indistinguishable from Iris's native encoder on our 30-prompt benchmark; rendered text and spatial relations are the weak spots. How it was built.

Turbo = Iris-3B Turbo, 4 fixed steps, no CFG. Base = Iris-3B base with the steps and CFG scale you choose (Iris's defaults are 100 steps, CFG 3). Free ZeroGPU time is limited per visitor, so Turbo is the default.
Model
Size (width × height)
Examples
Prompt Model Denoising steps (Base model) CFG scale (Base model) Size (width × height) Seed

Engine: PyTorch with Iris's own iris3b code; the production path is the Rust/Metal iris-rs engine described in the write-up. The adapter also drives our Qwen-Image 2.1 Space, where sounds and videos can be the prompt.