AI BriefingExtracting Triangle Splats from Video Diffusion Latents
·2026.06.25 09:00
Key point
FLAT directly maps video diffusion model latents to triangle splats, producing accurate 3D scene geometry.
1 / 2
Details
FLAT introduces a technique that directly maps compressed Video Diffusion Latents into explicit, non-volumetric scene parameters. Instead of decoding 3D Gaussians as in existing methods, it predicts Triangle Splats in a single pass, improving geometric accuracy.
This approach has the following key features:
- Direct Triangle Decoding: Unlike existing generate-then-optimize pipelines, it extracts triangle splats directly from the latents.
- Geometry-Specialized Training: To prevent gradient vanishing caused by rotational errors, it uses ray-centric triangle parameterization and a Product Window Rendering function.
- Usable Asset Conversion: The predicted triangle chunks are converted into opaque assets through a lightweight refinement stage, which is compatible with standard renderers and physics engines.
FLAT is designed to leverage the latent space of the Wan-2.1 model. As a result, as long as various Wan-2.1 variants—such as Text-to-Video and Image-to-Video—produce the same latent representation, it offers high flexibility to instantly generate triangle-based geometry without separate model training.
