[IROS 2026] SILICA: Repurposing Diffusion Priors for Joint Glass Segmentation and Depth Estimation

SILICA is a unified, single-step latent regression model that repurposes text-to-image diffusion priors for joint glass segmentation and glass-aware monocular depth estimation.

Using task-specific CLIP text prompts injected into cross-attention layers, SILICA establishes a spatial hierarchy that resolves foreground-background visual ambiguities inherent to transparent surfaces.

This work has been accepted for publication at IROS 2026.

Project Links

Project Page: silica-mirage.github.io
Code & Deployment: GitHub Repository (rtarun1/Silica)
Dataset: Mirage18k Dataset

Usage & ROS2 Deployment

For setup instructions, Python standalone inference, environment setup via uv/pip, and ROS2 (Humble) integration, please refer directly to the Official GitHub Repository.

Downloads last month: 172

Safetensors

Model size

0.9B params

Tensor type

F32

Inference Providers NEW

Depth Estimation

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for rtarun1/silica-v1-0

Base model

CompVis/stable-diffusion-v1-4

Adapter

(632)

this model

Dataset used to train rtarun1/silica-v1-0

Evaluation results

mIoU on Mirage 18k
test set self-reported

81.210
Delta 1 on Mirage 18k
test set self-reported

95.340