What is Iris-3B?
Iris-3B is an open-weight text-to-image model from Speridlabs, released on Hugging Face in October 2026 with an arXiv paper published on 7 October. Unlike most image generators, it does not work in a compressed latent space decoded by a VAE: a 3-billion-parameter diffusion transformer (8 dual-stream and 16 single-stream blocks) outputs raw RGB pixels through a small pixel head. Text comes from a frozen Qwen3-VL-4B encoder. It was trained from scratch at 256, 512 and then 1,024 pixels, 665,000 steps in total, and generates images of about one megapixel.
The repository also holds Iris-3B Turbo, a 4-step distilled version that makes a 1,024 by 1,024 image in about 0.7 seconds on one H100, and two fine-tunes of the same backbone for monocular depth estimation and for 4x image restoration and upscaling.
Everything is under Apache 2.0, including the training code; the Qwen3-VL-4B text encoder is also Apache 2.0.
Is Iris-3B worth it?
Iris-3B's case rests on doing more with less. On the project's own table it scores 0.540 on OneIG-Bench, level with the 20B Qwen-Image's 0.539 and just behind Z-Image 6B's 0.546, and 0.798 on GenEval against 0.87 for Qwen-Image. It falls further behind on long rendered text: 0.857 on LongText-Bench against 0.943 for Qwen-Image and 0.935 for Z-Image. These are Speridlabs' runs with official evaluators and no prompt rewriting; the comparison rows come from the other models' reports, and no independent arena has rated it yet.
The paper is unusually candid. It set out to show that skipping the VAE gives a better base for detail-heavy vision tasks, and reports "no significant improvement": fine-tuned for depth, Iris-3B only draws level with the latent FLUX.2 Klein, and on 4x restoration it does not beat a latent FLUX.2 Klein fine-tune. The project page's line that image generators can replace vision models such as DINOv2 goes further than the paper does.
Turbo is the practical download. Four steps with guidance baked in give about 0.7 seconds per image on one H100, and it scores 0.820 on GenEval, slightly above the base model, at the cost of less variety across seeds and less reliable long text (0.804). The full model defaults to 100 steps.
It asks for more than its parameter count suggests. You need an NVIDIA GPU with CUDA and the project's own code, since no diffusers or ComfyUI support is documented, and the text-to-image weights alone are about 12 GB in float32, with Turbo 6 GB more. For a researcher, the openly released recipe and training code are the main attraction.
Where it earns its place
A computer-vision researcher at a university lab
Study pixel-space diffusion without training a model from scratch
The weights, training recipe and distillation configs are all public under Apache 2.0, and the paper documents its failure modes. Starting from Iris-3B saves the 665,000-step pretraining run; fine-tunes for depth used 10,000 steps.
A two-person studio building a product-mockup app
Generate draft images on its own GPU without licence fees
Turbo produces a 1,024-pixel image in about 0.7 seconds on an H100 with no per-image charge and an Apache 2.0 licence. Check text-heavy layouts by eye: rendered text is its weakest area.
Before you commit
- Setup
- git clone the speridlabs/iris-3b repository, pip install -e . (Python 3.11 or later, PyTorch 2.7.1 or later), download the weights with hf download and run scripts/sample.py; the text encoder downloads on first run. An NVIDIA GPU with CUDA is required.
- Integrations
- Command-line scripts for sampling, depth (scripts/depth.py) and upscaling (scripts/upscale.py), plus two Hugging Face Spaces demos. A community 460 MB LoRA offers a lighter 4-step variant.
- Data and privacy
- Runs locally. Apache 2.0 allows commercial use; the card warns that outputs can be inaccurate, biased or unsafe and should be reviewed.
- Cost at scale
- No API fees. Downloads are about 12 GB for text-to-image, 12 GB each for the depth and upscaler models and 6 GB for Turbo; Turbo's 0.7-second figure is for an H100 at batch size 4.
- Limits
- About one megapixel per image; exact object counts and long text in images are unreliable. Turbo is fixed at 4 steps with no CFG or negative prompt and gives more similar images across seeds.
What’s good
- 0.540 on OneIG-Bench, level with 20B Qwen-Image (self-reported)
- Turbo: 1,024 x 1,024 in about 0.7 s on one H100
- Apache 2.0 weights and full training code
- Depth and 4x upscaling fine-tunes in the same repository
What’s not
- Long rendered text trails Qwen-Image and Z-Image
- NVIDIA CUDA only; no documented diffusers or ComfyUI path
- About one megapixel maximum output
- Paper finds no clear gain from the pixel-space prior
Best for
Researchers working on pixel-space diffusion and developers who want a permissively licensed, from-scratch image model they can retrain.
Iris-3B alternatives
Iris-3B: common questions
- Is Iris-3B worth it?
- A small, fully open pixel-space generator that matches Qwen-Image on one benchmark, from a lab whose own paper finds the pixel-space advantage it set out to prove did not appear. We rate it 3.7 out of 5.
- How much does Iris-3B cost?
- Free weights and code under Apache 2.0; free demos on Hugging Face Spaces. No paid API.
- What are the downsides of Iris-3B?
- Long rendered text trails Qwen-Image and Z-Image; nVIDIA CUDA only; no documented diffusers or ComfyUI path; about one megapixel maximum output; paper finds no clear gain from the pixel-space prior.
- Who is Iris-3B best for?
- Researchers working on pixel-space diffusion and developers who want a permissively licensed, from-scratch image model they can retrain.
- How much does Iris-3B cost for a team or in production?
- No API fees. Downloads are about 12 GB for text-to-image, 12 GB each for the depth and upscaler models and 6 GB for Turbo; Turbo's 0.7-second figure is for an H100 at batch size 4.
- Is Iris-3B safe for company data?
- Runs locally. Apache 2.0 allows commercial use; the card warns that outputs can be inaccurate, biased or unsafe and should be reviewed.
- What are the best Iris-3B alternatives?
- The closest alternatives we have reviewed are Z-Image-Turbo, FLUX 3 Image Klein and Qwen-Image-2.1.
Answers checked Oct 31, 2026.