XNews

Multimodal & Spatial Models: EgoX, TRELLIS.2, Mistral OCR 3

Multimodal & Spatial Models: EgoX, TRELLIS.2, Mistral OCR 3

Research and product updates introduce first-person view reconstruction, high-fidelity 3D conversion, and a production-grade OCR optimized for complex documents and pricing.

Key insights

EgoX converts third-person to first-person video

A Korean research team released EgoX, a model that transforms third-person video into a realistic first-person viewpoint and reconstructs camera-obscured regions. Demonstrations show plausible first-person renderings derived from third-person input and inferred occluded content.

TRELLIS.2 produces photorealistic 3D with PBR

Microsoft released TRELLIS.2, a model that converts images into near-photorealistic 3D assets with full PBR (physically based rendering) support. The release includes demo code and online examples showing higher visual fidelity and realism than prior versions.

Mistral OCR 3 targets robust document understanding at low cost

Mistral AI launched Mistral OCR 3, an OCR and document-parsing model emphasizing high-precision extraction for forms, tables and handwriting. The model reports a 74% win over the prior generation on difficult document tasks, outputs Markdown/HTML-structured tables, and prices at $2 per 1,000 pages with batch discounts to $1/1,000 pages.

Source posts