TRELLIS

Generates 3D models from natural language text descriptions using a multi-stage diffusion-based architecture that progressively refines geometry and appearance. The system employs a two-phase approach: first generating a coarse 3D representation via latent diffusion, then refining surface details and textures through iterative denoising steps conditioned on the text embedding. This enables conversion of arbitrary text prompts into exportable 3D assets without requiring 3D training data paired with text.

Provides real-time 3D visualization and manipulation of generated models directly in the browser using WebGL-based rendering with orbit controls, lighting adjustment, and material preview. The interface streams the generated 3D asset to a Three.js-based viewer that supports rotation, zoom, pan, and dynamic lighting to inspect geometry quality and texture details without requiring external 3D software.

Exports generated 3D models in standard interchange formats (GLB, GLTF, OBJ) with automatic geometry optimization and texture embedding. The export pipeline applies mesh simplification, vertex quantization, and texture compression to reduce file size while preserving visual quality, enabling seamless integration with game engines, 3D printing software, and other downstream tools.

Processes natural language text prompts through a pre-trained vision-language model (likely CLIP or similar) to extract semantic embeddings that condition the 3D generation diffusion process. The system maps arbitrary text descriptions to a learned embedding space that guides geometry and appearance synthesis, enabling intuitive text-based control over 3D model generation without requiring structured 3D descriptors or parameter tuning.

Implements a multi-step diffusion denoising process that progressively refines 3D geometry and texture quality through repeated denoising iterations, each conditioned on the text embedding and previous refinement state. The pipeline starts with coarse geometry and iteratively adds detail, surface refinement, and texture information across 20-50 denoising steps, with each step reducing noise and improving coherence.

Manages multiple concurrent generation requests through a queue-based system that serializes GPU inference while maintaining responsive user feedback. The system caches generation results keyed by prompt hash, enabling instant retrieval of previously generated models for identical prompts without re-computation. Queue management prevents GPU overload and ensures fair resource allocation across simultaneous users.

Exposes the 3D generation pipeline through a Gradio-based web interface that provides real-time feedback during inference, including progress indicators, intermediate generation visualizations, and streaming status updates. The interface abstracts away infrastructure complexity, enabling users to interact with the model through simple text input and visual output without API knowledge or local setup.