The feedforward method improves generation speed and structural fidelity
This is the way.
📣New Paper: Feedforward Latent Triangle Splatting for Geometrically Accurate Scene Generation. 🌐 https://flat-splat.github.io/ 🤗: https://huggingface.co/papers/2606.24876
A feedforward latent scene decoder for generating explicit triangle-based 3D scenes from video diffusion latents.
Users are praising the Feedforward Latent Triangle Splatting paper for its flexible decoding of video-model latents and especially practical test-time refinement that improves 3D scene generation.
No Digg Deeper questions have been answered for this story yet.
FLAT decodes explicit triangle-based 3D scenes directly from video diffusion latents in a single forward pass. Instead of first generating RGB frames and then reconstructing geometry, it learns a feedforward latent scene decoder for explorable 3D scene generation.
Non-volumetric primitives are harder to predict than 3D Gaussians because small orientation errors significantly impact gradient flow. FLAT addresses this with a ray-centered rotation parameterization and a new product window function for differentiable triangle rendering.
📣New Paper: Feedforward Latent Triangle Splatting for Geometrically Accurate Scene Generation. 🌐 https://flat-splat.github.io/ 🤗: https://huggingface.co/papers/2606.24876
A feedforward latent scene decoder for generating explicit triangle-based 3D scenes from video diffusion latents.
@OKupyn Cool work @OKupyn 🔥
A key point is the representation. Many feedforward scene decoders predict volumetric primitives such as 3D Gaussians. FLAT instead predicts non-volumetric, surface-aligned triangle splats, giving geometry with a better-defined surface for downstream use.
Another nice property is flexibility. FLAT is trained to decode video-model latents directly, so the same scene decoder can plug into Wan-2.1 style pipelines beyond a single mode, enabling text to 3D, image to 3D, real-time generation and others.
@OKupyn @janheld14 @rvandeghen another triangle splatting.
@OKupyn That looks really cool ! Looking forward to read it. @janheld14 👀
@janheld14 Thanks, wouldn’t be possible w/o Triangle and Mesh Splatting, thanks for your great work too!
The final step is what makes the representation especially practical. A lightweight test-time refinement turns the predicted triangle soup into a fully opaque, game-engine-ready asset that supports real-time rendering and direct physical interaction.
