A research team says it can generate convincing new camera angles from a single photo or video clip - without touching the underlying model weights.
NeoMap is a training-free framework that takes general pre-trained video models and coaxes novel view synthesis out of them without task-specific fine-tuning or camera conditioning. The key idea is that video models already encode the geometry of scenes implicitly in the patterns they learned from real footage. NeoMap does not teach the model anything new; it uses a technique called convergent manifold alternating projection to search the model's existing solution space for outputs that are both photorealistic and geometrically consistent. The team tested it on three standard benchmarks - Tanks-and-Temples, LLFF, and DAVIS - and reports state-of-the-art results on all three.
Most competing approaches bolt on extra machinery - camera conditioning signals, stepwise denoising constraints, or fine-tuning on 3D datasets - and still produce artifacts when scenes get complex. NeoMap sidesteps that by treating view synthesis as a search problem rather than a training problem, which means it can slot into any capable pre-trained video model without retraining pipelines or specialized data. That portability matters as video foundation models keep getting better.
The paper is a preprint and has not been peer-reviewed, so the benchmark numbers deserve some skepticism before they get cited in a pitch deck.