Skip to content
← Models
Open weightsSperidLabs/10 trustMetrics-checked

Iris-3B

A 3B pixel-space image generator that paints every pixel directly with no VAE — and doubles as a depth-estimation and super-resolution vision learner.

Visit official page →

About

Iris-3B is a 3-billion-parameter text-to-image model from SperidLabs that skips the compressed latent space most generators use: the network outputs every pixel directly, using flow-matching diffusion to avoid anything being lost to a lossy, texture-biased decoder. It produces 1024-pixel images from a text prompt, and the same pixel-space prior — fine-tuned with no architectural change — also serves as a general vision learner for monocular depth estimation, image restoration, and super-resolution, with those variants shipped in the repo. Iris-3B Turbo, a 4-step distilled version in the turbo/ folder, renders a 1024x1024 image in about 0.7 seconds on a single H100 (vendor-reported). The work is backed by a paper (arXiv 2610.09450), code, and live HF Spaces demos. Released 2026-10-05 under Apache 2.0, it has around 1.3k downloads and 299 likes.

Specs

Context window

Not applicable (image model)

Pricing

Free — open weights, self-hosted

Free quota

Weights free under Apache 2.0; free HF Spaces demo and Turbo demo

API access

No managed API — self-host the weights (PyTorch/safetensors) or use the HF Spaces demos

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Iris-3B Turbo: 1024x1024 in ~0.7s on one H100 — provider-reported

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

image-generationvision

Trust score — 5/10

/10 trustMetrics-checked

Scored 7/10 (METRICS tier, metrics-v1). Evidence: Apache 2.0 with paper, code, and live demos — strong openness; novel pixel-space approach documented in detail. Docked for: (1) modest traction so far (299 likes, ~1.3k downloads); (2) only speed claim published, no quality benchmarks against other generators yet. License: Apache-2.0.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
6.0/10
Maintenance · recency, releases, issue load
7.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · HF: 299 likes, ~1.3k downloads, created 2026-10-05
  • · manual review: paper (arXiv 2610.09450), code, and live demos all present
  • · docked: no quality benchmarks against other generators; speed claim only, provider-reported

Sources