iris-3b: Pixel-Space Generative Model Skipping VAE Achieves Depth Estimation with 3B Parameters

This is a 3-billion-parameter generative model that outputs all pixels directly from the network, without passing through a compressed latent space or a separate decoder. It generates images at 1024-pixel resolution without losing texture or detail, producing high-resolution results solely from prompt inputs.

1 / 8

Beyond text-to-image generation, a single model performs monocular depth estimation and image restoration and upscaling. Using fine-tuned weights without structural changes, it instantly generates detailed depth maps or 4x upscaled images with a single forward pass using an empty prompt.

As an alternative to vision foundation models like DINOv2, it validates the effectiveness of generative priors in dense vision tasks. It runs in NVIDIA GPU and CUDA environments and is licensed under Apache 2.0, allowing free modification for commercial and research purposes.