Speridlabs Releases Iris-3B Pixel-Space Generator

AI research firm Speridlabs has launched Iris-3B, a 3-billion-parameter text-to-image diffusion transformer distributed under an Apache 2.0 open-source license. Unlike standard visual generators that rely on variational autoencoders (VAEs) for latent compression, Iris-3B produces raw RGB pixels directly. The architecture uses 8 dual-stream blocks followed by 16 single-stream blocks, passing data through a specialized pixel-transformer (PiT) head. To manage compute costs, the system divides images into 16 by 16 pixel patches, forming a 64 by 64 grid of 4,096 patch tokens at 1024 by 1024 resolution.

In performance evaluations at 1024px, the open-weight model achieved a 0.798 GenEval score and an 86.52 DPG rating. These metrics allow Iris-3B to match Qwen-Image on the OneIG-EN benchmark. Beyond base generation, the same foundational backbone can be fine-tuned into derivative models, including a monocular depth estimator and a 4x image restoration upscaler.

Speridlabs highlighted a notable negative finding regarding pixel-space priors. Under the test's training budget, generating pixels directly without latent compression failed to beat the latent-based FLUX.2 Klein model on depth estimation and image restoration tasks. Nevertheless, the release provides developers with public weights, code, a paper, and a Hugging Face demo, alongside community-built ComfyUI workflows and Turbo variants. For machine learning practitioners, Iris-3B offers a practical testbed for exploring non-latent diffusion architectures without sacrificing compatibility with existing tooling.