Adobe's Long-LRM++ Renders Sharp 3D Scenes 8x Faster Than Rivals
Adobe Research's new feed-forward reconstruction model matches implicit-method quality while rendering scenes at 14 FPS on a single A100 GPU.
Takeaways
- Long-LRM++ from Adobe Research reconstructs scenes at 14 FPS on an A100, 8x faster than LaCT.
- Each 3D Gaussian carries a 32-dim feature vector instead of RGB, decoded by a lightweight 5-layer network.
- Uses only 1/4 the Gaussians of Long-LRM while producing sharper text and fine detail.
- Generalizes zero-shot from 32 training views to 16 or 64 views on DL3DV-140.
- Beats Long-LRM on ScanNetv2 by +4.6 dB PSNR and cuts depth Abs Diff by 0.053.
- Details, figures, and video demos on the project page.
Rebuilding a 3D scene from a stack of photos usually forces a choice between two bad options: a fast but blurry model, or a sharp one that renders too slowly to move a camera around. Long-LRM++ from Adobe Research, Tripo AI, Hillbot, and Oregon State University proposes a middle path that keeps the speed of Gaussian splatting while approaching the fidelity of the slower implicit models.
The blur-versus-latency wall
Feed-forward reconstruction has been climbing a scaling curve. The predecessor Long-LRM was the first to push this idea to full-scene coverage, taking 32 input images at 960x540 resolution and producing the Gaussian reconstruction in about a second on a single A100 GPU. The trick was a mixture of Mamba2 blocks and classical transformer blocks, plus token merging and Gaussian pruning to handle the 250K-token input sequence.
The problem, as the new paper spells out, is that directly predicting millions of Gaussian parameters at once remains highly error-sensitive: small inaccuracies in positions or other attributes lead to noticeable blurring, particularly in fine structures such as text. Implicit methods like LVSM and LaCT skip explicit Gaussians and compress scene information into model weights, decoding RGB frames using the full transformer or TTT backbone, which makes real-time rendering infeasible. LaCT looks great, but each frame requires running an entire 24-layer network. That is fine for a benchmark, useless for a walkthrough.