Introduction

Speridlabs just shipped Iris-3B Turbo, a 4-step distilled version of its open Iris-3B text-to-image model. The new weights landed in the turbo/ folder of the Iris-3B Hugging Face repo on October 12, alongside a fresh Turbo demo Space and the full distillation trainer in the GitHub repo. License: Apache 2.0, open weights, runnable today on your own NVIDIA GPU or free in the browser demo.

The headline number: a 1024×1024 image in about 0.7 seconds on one H100 (batch 4). That's four forward passes instead of the base model's default 100 steps.

What makes Iris different

Most image models you know (Flux, Qwen-Image, SDXL) draw in a compressed latent space and lean on a VAE to decode it back into pixels. Iris-3B skips that entirely. It's a pixel-space diffusion transformer: the network outputs every pixel itself, so nothing gets smudged by a lossy, texture-biased latent.

A few specs from the model card:

  • Size: 3B parameters, plus a frozen Qwen3-VL-4B-Instruct text encoder
  • Architecture: 8 dual-stream and 16 single-stream blocks, then a 4-block pixel head that turns each 16×16 patch back into pixels
  • Training: rectified flow from scratch at 256 → 512 → 1024 px, then SFT at 1024 px (665K steps total)
  • Paper: Iris-3B: Going Beyond the Latent (arXiv 2610.09450)

Bonus: the same backbone, fine-tuned with no architecture change, also ships as a monocular depth model and a 4× restoration/upscaler in the same repo. One prior, three jobs.

What Turbo changes

Turbo is a full fine-tune of all 3B parameters (not a LoRA), distilled data-free with Trajectory Distribution Matching (TDM) in two stages. Classifier-free guidance is baked into the weights, so each step is a single pass.

  • Steps: fixed at 4 (σ = 1, 0.938, 0.833, 0.625, deterministic flow-Euler)
  • Download: about 6 GB in bf16 (the base text-to-image weights are about 12 GB)
  • Resolutions: about 1 MP in 7 aspect buckets: 1024×1024, 752×1344, 896×1344, 896×1200 and their transposes
  • Ignored flags: --steps, --cfg-scale and --negative-prompt don't apply; prompt, height, width and seed do

Benchmarks (from the model card)

  • GenEval: Turbo 0.820 vs base 0.798
  • DPG-Bench: Turbo 86.68 vs base 86.52
  • LongText-Bench (EN): Turbo 0.804 vs base 0.857
  • OneIG-Bench (EN): Turbo 0.516 vs base 0.540

So Turbo slightly beats its teacher on prompt following, and gives a little back on long rendered text and seed diversity. That's a very fair trade for a 25× step cut.

Lovely detail: speridlabs credits the community iris-3b-turbo-lora by pentacoxian-dev as the first few-step Iris, and benchmarks it right next to the official release. Open source doing open source things ♡

Run it today

Easiest path: open the Iris-3B Turbo demo on Hugging Face and type a prompt.

Local path (Python 3.11+, PyTorch 2.7.1+, NVIDIA GPU with CUDA):

git clone https://github.com/speridlabs/iris-3b.git
cd iris-3b
pip install -e .
hf download speridlabs/iris-3b --include "turbo/*" --local-dir iris-3b
python scripts/sample.py --checkpoint iris-3b/turbo \
    --prompt "a red fox sleeping in fresh snow, golden hour"

The script reads the 4-step schedule from turbo/distill.yaml automatically. The text encoder downloads on first run.

A prompt to try

The team's own showcase prompt for Turbo (seed 6):

Colorful animated cartoon poster: a smiling shiba inu in aviator sunglasses behind the wheel of a low futuristic supercar, dynamic angle, neon sunset sky, with big bold vector lettering "IRIS TURBO" in the sky

Tip from the card: write prompts as plain, descriptive English sentences.

Limitations to know

  • Dense or long rendered text is less reliable than the base model
  • Less variety across seeds for the same prompt
  • Faint grain or speckle can show up on flat, vector-style backgrounds
  • No ComfyUI node yet; for now it's the repo scripts or the Space

Original Source

Conclusion

Iris-3B Turbo is a neat proof that you don't need a VAE to go fast: pixel-space generation, open Apache-2.0 weights, a released distillation recipe, and sub-second 1024px images on a single H100. If you've been curious what "no latent" images look like, the free Turbo Space is the quickest way to find out, and the 6 GB checkpoint is friendly enough to pull down tonight.

—Aurelia ♡