Iris-3B Goes Open, Skips VAE

// 1h agoMODEL RELEASE

Iris-3B Goes Open, Skips VAE

Sperid Labs released Iris-3B, a fully open 3B-parameter pixel-space diffusion transformer that generates images without a VAE and fine-tunes for monocular depth and 4× restoration. The model matches Qwen-Image on OneIG while shipping weights, training code, and a demo under Apache 2.0. [GitHub](https://github.com/speridlabs/iris-3b) [Paper](https://huggingface.co/papers/2610.09450)

// ANALYSIS

Iris-3B is a meaningful scaling result for pixel-space generation, but its own evaluations temper the VAE-free hype: the representation is viable, not yet clearly superior.

–The 256→512→1024 training curriculum shows pixel-space diffusion can scale to 3B parameters and competitive text-to-image quality.
–The same backbone supports generation, depth estimation, and restoration, offering an unusually broad open research platform.
–Depth performance roughly matches latent FLUX.2 Klein, while 4× restoration does not beat the latent baseline.
–Developers should expect research-grade tradeoffs, including patch-grid artifacts and potentially slower inference at high resolution. [AI Weekly](https://aiweekly.co/alerts/speridlabs-scales-iris-3b-to-3b-finds-no-downstream-edge) [Community discussion](https://www.reddit.com/r/StableDiffusion/comments/1x17iqb/iris3b_speridlabs_research/)

// TAGS

iris-3bimage-genvisionopen-weightsopen-sourceresearchmultimodal

DISCOVERED

1h ago

2026-10-11

PUBLISHED

1h ago

2026-10-11

RELEVANCE

9/ 10

AUTHOR

AI Search

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

LAUNCH33m ago

Tracekit 1.0 Makes Agent Actions Verifiable

Tracekit 1.0 places a separate signer between AI agents and their tools, enforcing policy before execution while creating signed, hash-chained records that can be verified offline. The open-source project supports coding agents, SDKs, LangChain, MCP, and OpenTelemetry. [GitHub repository](https://github.com/Cygnux-Labs/Tracekit)

RESEARCH1h ago

GenIA grounds SAM3D in observed views.

GenIA is an inference-time framework from Meta Reality Labs and the University of Tübingen that aligns SAM3D’s generative 3D prior with geometric and photometric observations. It reconstructs input-aligned objects from monocular images, multi-view captures, and videos without retraining the foundation model.

OPEN SOURCE1h ago

WorldPlay2 Pushes Interactive Worlds Beyond Short Memory

WorldPlay2 is an open-source interactive video world model combining frame-aligned actions with semantic controls for scene, character, and event consistency. Its project includes model weights, inference code, and a Reactor browser demo, though local deployment remains GPU-intensive.