Iris-3B by speridlabs

Iris-3B by speridlabs

Generates 1024-pixel images from text prompts by painting every pixel directly, with no VAE or latent space. Fine-tuned versions estimate depth and restore or upscale images. For vision researchers and image makers.

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

About

What it is: Iris-3B is a 3-billion-parameter text-to-image model that works in pixel space. Instead of drawing in a compressed latent space and decoding it, the network outputs every pixel itself. With fine-tuning and no change to the architecture, the same model also estimates depth and restores and upscales images.

Who it suits: Researchers and image makers who want to try a pixel-space generator and its depth and restoration variants.

Industry | Who uses it | How they use it | Example
Research | Vision researchers | Study pixel-space priors as an alternative to vision foundation models | Compare it with a latent-space generator
Design | Artists | Generate images from detailed text prompts | Make a studio-style portrait or a 3D render look
Imaging | Developers | Use the depth and restoration variants | Upscale a low-resolution photo

If you need a tiny model that runs on a phone, you can skip it.

Where to use it:

• Your own computer with Python and PyTorch, on Windows, macOS or Linux
• The demo pages the card links, for a quick look
• The Turbo version in the turbo folder, for faster images

Use case | Who | What they get back
Text-to-image | Artists | A 1024-pixel image from your prompt
Depth estimation | Developers | A depth map from a fine-tuned variant
Restoration and upscaling | Developers | A cleaned and enlarged image

How to set it up:

• Python with pip
• A GPU with room for the weights, about 12 GB in F32, or about 1.5 GB at 4-bit (an estimate)
• The project's code from its GitHub page, which the card tells you to install first

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Install the project's code as its model card describes, then run it.

pip install -U huggingface_hub
hf download speridlabs/iris-3b

How to use it: You type a prompt and get a 1024-pixel image. The card's examples use the default settings, a guidance value of 3 with 100 steps. The part of the card shown here cuts off before the run commands, so check the model card for them.

Example: You enter a prompt describing a bronze horse sculpture under a museum spotlight, at the default settings. The model returns one image of about one megapixel at a native aspect ratio.

For newcomers and experts: If you are new, try the hosted demo pages first to see what it makes, then run it locally. Use the default settings to start.

For experts, the idea is that a pixel-space prior avoids the lossy, texture-biased latent that decoders add. The project explores pixel-space generators as a substitute for backbones like DINOv2. The Turbo variant is a 4-step distilled version that makes a 1024 by 1024 image in about 0.7 seconds on one H100.

You type a text prompt
The network paints every pixel directly
No VAE or latent decoder is involved
You get a 1024-pixel image

Good to know: The card lists both a standard model and a faster Turbo version.

Title: Sampling steps from the card
Iris-3B default: 100 steps
Iris-3B Turbo: 4 steps

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key or account is named. You need a GPU for the 3B weights.
• Data it reaches: downloading reaches Hugging Face and GitHub. The card does not describe other data access.
• Not for: the card names no banned uses. The Turbo timing refers to one H100 GPU.

On Hugging Face, 13 November 2026

• Task: text to image
• Library: pytorch
• Parameters: 3B
• Weights: about 12 GB in F32; about 1.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 997
• Likes: 350
• License: apache-2.0
• Quantized versions: 19
• Fine-tunes: 4
• Tags: pytorch, safetensors, text-to-image, diffusion, pixel-space, flow-matching, general-vision-learner, vision-foundation-model, depth-estimation, image-restoration, super-resolution

Links

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

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download speridlabs/iris-3b

Then follow the usage in its model card. Nothing in this kit runs by itself.