Model Card: Iris-3B

🌏 中文版

Model Info

Capability Highlights

• Skips the VAE and latent space entirely — the diffusion transformer outputs every pixel directly (each 16×16 patch is decoded back to pixels through a 4-block PiT pixel head); training runs through a 256→512→1024-pixel curriculum, 665K steps in total
• Text-to-image hits GenEval 0.798 and DPG-Bench 86.52 (both at 1024px, official evaluators), and matches Qwen-Image's overall OneIG-EN score (the paper doesn't name a specific competing version)
• Without any architecture change, the same backbone fine-tunes into a monocular depth estimator (AbsRel 0.071, δ1 0.946, zero-shot mean over NYUv2/KITTI/ETH3D/ScanNet/DIODE, Marigold V2 protocol) and a 4x image restoration/upscaling model (LPIPS 0.292, PSNR 20.68, DIV2K validation set)
• Weights, training code, and the paper are all public (Apache 2.0); it also ships fine-tuned depth and upscaling checkpoints plus a HuggingFace Space demo you can try without a GPU

Benchmark Results

⚠️ All numbers above are self-reported by Sperid Labs; no independent third-party reproduction has surfaced yet. The paper's own headline finding is itself a negative result: dropping the VAE showed no significant advantage on either depth estimation or image restoration.

Comparison with Predecessor/Competitors

This is Sperid Labs's first public release, so there's no predecessor to compare against — the comparisons come from the paper's own two control groups: a native latent model, FLUX.2 Klein base 4B, and that same FLUX.2 Klein converted into pixel space. Starting with text-to-image: tying Qwen-Image on OneIG-EN and posting respectable GenEval/DPG numbers shows that the first half of the hypothesis holds — dropping the VAE to output pixels directly doesn't cost you quality.

But the paper's real point — that a pixel-space model, free of the VAE's lossy compression, should have an edge on downstream tasks where fine detail matters — doesn't hold up. After fine-tuning, depth estimation only ties the latent (VAE-based) FLUX.2 Klein, and 4x image restoration doesn't even beat it. Converting an existing FLUX.2 Klein into pixel space (rather than pretraining from scratch like Iris-3B) performs even worse — neither route validated the original hypothesis.

Both sides are open-weight, self-hosted, with no official API, so the real comparison is hardware cost: Iris-3B needs an NVIDIA GPU with CUDA, about 12GB of weights just for text-to-image, with another 12GB each for the depth and upscaling checkpoints. Running 100 diffusion steps through a 3B model plus a 4B text encoder isn't light, and the authors don't give any inference-time or VRAM figures of their own.

Implications for Agent Development

The biggest takeaway for agent developers here isn't "one more image model" — it's that this paper saves you a round of redundant validation. If you're weighing whether to swap your agent's vision-perception module for a VAE-free, pixel-space backbone — say, using a generative prior for depth estimation or image restoration — Sperid Labs has already run the controlled experiment and gotten a negative result, so you don't have to re-run the same test yourself.

• If you're building: a small, self-hosted image-generation agent that doesn't want closed-API billing (asset generation, prototype visualization) — Iris-3B is an Apache 2.0, 3B-parameter option that runs on a single data-center or high-end consumer GPU, and its OneIG-EN parity with Qwen-Image means the quality is usable
• If you're building: a pipeline that needs generation, depth, and restoration from one backbone without maintaining three separate architectures — Iris-3B demonstrates that "same architecture, separately fine-tuned" is a viable path
• Not a fit: any scenario betting that pixel space means better downstream vision performance — if your depth-estimation or image-restoration agent already runs a mature latent model (the FLUX family, DINOv2-style encoders), the paper gives you no reason to switch; also not a fit for low-latency interactive scenarios, since 100 diffusion steps plus downloading extra sub-models isn't a lightweight path

Today's Takeaway

"A VAE is lossy compression, so removing it should buy you finer downstream detail" sounds like a reasonable assumption that barely needs testing — but Sperid Labs ran the actual controlled experiment (same data scale, same training steps, two different routes) and got "it works, but there's no edge." That's a reminder that architecturally "reasonable-sounding" assumptions are still just assumptions until someone runs the control. Choosing to publish a negative result honestly — instead of packaging "ties Qwen-Image on text-to-image" as a sweeping breakthrough — is itself worth noting more than the model's raw numbers.

References

• HuggingFace: speridlabs/iris-3b model card
• arXiv 2610.09450: Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning
• GitHub: speridlabs/iris-3b (training code)
• explainx.ai: Iris-3B deep dive (paper details and caveats)
• AI Weekly: SperidLabs Scales Iris-3B to 3B, Finds No Downstream Edge
• Sperid Labs official research page