From Images to 3D and Flat to Layered: Two AI Technologies Set to Reshape Design Workflows

T-RELLIS .2: Image‑to‑3D Model

Microsoft Research released the open‑source T‑RELLIS .2 model, a 4‑billion‑parameter image‑to‑3D system that generates high‑fidelity PBR‑textured 3D assets from a single sketch within seconds to a minute.

Core concept : a native, compact structured latent variable based on a 3D VAE enables 16× spatial compression and supports resolutions up to 1536³.

Main Features

Speed & resolution : 3 s (2 s shape + 1 s texture) → 512³; 17 s (10 s + 7 s) → 1024³; 60 s (35 s + 25 s) → 1536³ (tested on NVIDIA H100 GPU).

Arbitrary topology handling : robust on open surfaces, non‑manifold geometry, and internal structures, breaking the limits of traditional isosurface methods.

Rich material modeling : generates PBR attributes – base color, roughness, metallic, opacity (alpha channel) – ready for game engines or renderers.

Minimal asset pipeline : encoding (mesh → O‑voxel) under 10 s on a single CPU; decoding (O‑voxel → mesh) under 100 ms with CUDA acceleration.

Workflow

The process starts with an “instant bidirectional conversion” that transforms a 3D mesh into an “O‑voxel” representation, then a sparse‑compression VAE encodes the voxels into a compact latent space for generation or reconstruction.

Qwen‑Image‑Layered: Native Editability via Layer Decomposition

Qwen‑Image‑Layered, from the Tongyi Qianwen team, addresses the “pixel entanglement” problem of raster editors by decomposing a single RGB image into multiple semantically decoupled RGBA layers, enabling independent manipulation.

Technical Highlights

RGBA‑VAE : unifies latent representations for RGB and RGBA images, eliminating distribution gaps of previous methods.

VLD‑MMDiT architecture : a novel design that supports variable‑length layer decomposition.

Multi‑stage training strategy : adapts a pretrained image‑generation model into a multi‑layer decomposer.

To overcome scarce high‑quality multi‑layer data, the team built an automated pipeline that extracts and annotates layers from Photoshop PSD files.

Results

Quantitative comparison on the Crello dataset shows Qwen‑Image‑Layered achieving the lowest RGB reconstruction error and highest Alpha‑soft IoU among competing methods.

Qualitative examples demonstrate accurate separation of foreground subjects, backgrounds, decorative elements, and precise alpha channels.

Project open‑source : code and models are available on GitHub at https://github.com/QwenLM/QwenImage-Layered

Conclusion

Both technologies point to a future where AI moves from simple assistance to structural understanding: T‑RELLIS .2 lowers the barrier and time cost for 3D content creation, while Qwen‑Image‑Layered brings native, editable layer representations to AI‑generated images, opening doors for fine‑grained editing, style transfer, and content recombination.