Enlighten is an AI research lab exploring neuro-symbolic world models, aiming to develop superhuman creative assistants and exploratory agents in physical world.
Meet the team
CO-FOUNDER AND CEO
Bell previously founded and built ST World, a sandbox simulation and LLM-driven game engine. His AI journey began at the age of 17 when he trained a tiny Copilot with next-word prediction LSTM before Transformer era and co-published work on early image generation models when he was in high school.
CO-FOUNDER and Head of research
Zike, a PhD student at University of British Columbia, has been pioneered in world model research. With his authorship in Consistent3D (CVPR 2024) and MV-Gamba (NeurIPS 2024), Zike demonstrates his vision in world models to spearhead the research of the company.
November 2024
Enlighten-SDS 1.5
Text-to-3D (dense Tri Mesh)
Text-to-Physical-Material (Standard BRDF)
Overview
Enlighten-SDS 1.5 is a cutting-edge 3D synthesis system optimized for realism-focused projects like filmmaking, CG rendering, and virtual production. Using text prompts, image guidance, or mesh input, it generates photorealistic, high-resolution 3D models with unmatched high-frequency detail, closely resembling real-world visuals. While artifacts may occur, the added detail enhances post-processing versatility and flexibility for human editing.
Performance and Limitation
As of February 2025, Enlighten-SDS maintains the highest fidelity in texture generation. However, its optimization techniques result in a one-hour generation time, 50–100 times higher computational cost, reduced geometric detail compared to inference-only methods, and frequent failures. Efforts are ongoing to integrate pre-trained Flow Matching models and post-processing methods to improve pipeline compatibility especially for texture creation.
Approach
Enlighten-SDS is a state-of-the-art score distillation sampling (SDS) variant with interpreting its process as trajectory sampling of a stochastic differential equation (SDE), enabling the transition of 2D knowledge from a pre-trained latent diffusion model into 3D space. Additionally, it incorporates advanced differential ray tracing and inverse shading techniques, integrating topological and semantic guidances, including Vision-Language Models.
Data Transparency and Safety
Enlighten-SDS transforms visual knowledge from image generation models like Stable Diffusion and PixArt into 3D, prioritizing knowledge from sensor-captured photos over handcrafted art. It enhances geometric consistency using techniques trained with the G-buffer Objaverse dataset containing web-collected 3D assets. A built-in security protocol prevents the generation of harmful or copyrighted 3D content, including nudity, violence, child abuse, and protected characters.