Iris-3B

Iris-3B

A demo of Iris-3B, one 3B model that makes images from text, estimates depth from a single image and restores or upscales pictures. For vision researchers and developers trying a general image model.

Made by speridlabs on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space is a demo of Iris-3B, a general vision model that works directly in pixel space. One 3B model handles text-to-image, depth estimation from a single image, image restoration and upscaling. It runs on ZeroGPU at Hugging Face, and a research paper backs it.

Who it suits: Computer vision researchers, developers and curious makers who want to try one model across several image tasks.

Industry | Who uses it | How they use it | Example
Research | Vision researchers | Test one model on generation, depth and restoration | Run the preset examples and compare outputs
Software | Imaging developers | Study how a pixel space model is wrapped in a demo | Read app.py, viewers.py and the iris3b folder
Creative | Designers | Make an image from a prompt | Type a scene and pick a size
Photography | Photo editors | Restore or upscale a picture | Add an image and run the restoration task

If you want a finished consumer editor, a research demo has fewer polished controls.

Where to use it: It runs in two places.

• A web browser, on the Space's page at Hugging Face with nothing to install
• A local copy on Windows, macOS or Linux, which needs Python 3.12

Use case | Who | What they get back
Text to image | Creators | A generated picture from your prompt
Depth from one image | Researchers | An estimated depth map for a single image
Restore or upscale | Photo editors | A cleaner or larger version of your picture

How to set it up: Using it online needs nothing. A local copy needs more.

• Python 3.12 and pip
• The packages listed in requirements.txt
• A Hugging Face token set as a secret named HF_TOKEN, only while the model repository is private
• A GPU that suits a 3B model, since the page was built for ZeroGPU

1. Install the Hugging Face tools.
2. Download the Space into a folder.
3. Move into that folder.
4. Install Gradio and the packages from requirements.txt.
5. Start the app.

pip install -U huggingface_hub
hf download speridlabs/iris-3b --repo-type space --local-dir iris-3b
cd iris-3b
pip install gradio==5.49.1
pip install -r requirements.txt
python app.py

How to use it: Open the Space page and pick a task. You can start from one of the curated example presets, each of which records its prompt, size, seed, steps and guidance. Change those values to explore. The app loads exported weights from a model repository named by the IRIS_MODEL_REPO setting.

Tool | What it does
Text-to-image | Generates a picture from a written prompt
Monocular depth | Estimates depth from a single image
Image restoration | Cleans up a damaged or noisy picture
Upscaling | Makes a picture larger

Example: You open a preset, read the prompt and the settings it was sampled with, then change the seed and run it again. The page shows a new image from the same prompt.

For newcomers and experts: Start with a preset to see the quality you can expect, then change one setting at a time. The seed picks the random start, and the steps and guidance shape the result.

Pixel space means the model works on pixels directly instead of a compressed latent form. One 3B prior serves several tasks. The weights load from a safetensors file plus a config file, and the iris3b folder copies the source package from the code repository. The Space also names a Qwen3-VL 4B Instruct model.

You pick a task and settings
The app loads the exported Iris-3B weights
The 3B pixel space model runs on ZeroGPU
The result shows in the page

Good to know: The README is short, so the supported sizes and limits of each task are not listed.

• License: Apache 2.0
• Cost to run: ZeroGPU time at Hugging Face, plus an HF_TOKEN secret only while the model repository is private
• Data it reaches: the prompts and images you add, and model files from the Hub

On Hugging Face, 13 November 2026

• Likes: 65
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/speridlabs/iris-3b
• Files: https://huggingface.co/spaces/speridlabs/iris-3b/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.