Six unrelated papers landed on arXiv within 48 hours this week. None have been announced on social media. None have press releases. Taken together they remove almost every major technical bottleneck that has held back production 3D reconstruction and physically consistent video generation for the last two years.

This is not one big model launch. This is the quiet moment when an entire field crosses the line from research demo to usable technology.

We stopped fine tuning video generators. We started rewarding them.

Camera controlled video generation worked great on test sets. It failed completely on real world input.

Every existing method used supervised fine tuning on synthetic paired video. You train the model on 10 million clips where you know exactly what a 30 degree right pan looks like. Then you give it a real phone video and ask for the same pan, and it generates an entirely different scene. There is no large dataset of synchronized multi-view real world video to train on, and there never will be.

Geo-Align abandons this approach entirely. It is the first RL framework built for camera controlled video re-rendering. Instead of training against ground truth output, it runs a lightweight metric 3D estimator directly on the generated video. It measures exact deviation in camera rotation, translation and scene scale, and uses that error as reward signal.

No paired data is required at all. No ground truth. No synthetic labels. The model learns to respect geometry by being graded on its own output. Across all tested benchmarks it outperforms every supervised baseline on both camera controllability and visual fidelity.

This is not an incremental improvement. This is a complete replacement for the dominant training model for controllable video.

Geometry transformers just got 7x faster.

Visual geometry transformers are the best available method for multi-view 3D reconstruction. Nobody uses them at scale. Their global attention layers run at O(n²) cost, which meant input sequences were hard capped at roughly 60 images. Run anything larger and inference time becomes unusable.

Good Token Hunting solves this with an approach so obvious it is remarkable nobody implemented it correctly before this paper. 90% of tokens fed into geometry transformers contribute nothing to the final output. They can be discarded entirely before attention runs, with zero loss and often improved accuracy.

The paper uses a two stage selection system. First it selects useful frames using a diversity metric that ensures full scene coverage. Then within selected frames it discards redundant tokens, guided by the entropy of the attention pattern from the previous layer.

The result is an 85% reduction in inference compute. That is 6.6x faster. For scenes with 500 input images it runs in reasonable time on consumer hardware, and produces better reconstruction quality than the unmodified full model.

Streaming 3D reconstruction no longer drifts.

Every online 3D reconstruction system drifts. Always. Run any SLAM or streaming pipeline for 1000 frames and your scale will be wrong, walls will be bent, the entire scene will have silently rotated a few degrees.

For ten years everyone blamed bad feature matching. Everyone was wrong.

HorizonStream traces the entire failure mode to the attention architecture used in every modern system. Sliding windows throw away old geometric evidence entirely. Ungated causal attention forms permanent sinks that stop updating and ignore all new input. Both approaches produce pathological behaviour on long sequences.

This paper formalizes geometric propagation as an evidence influence kernel, then builds a transformer that explicitly factorizes this kernel. Geometric linear attention learns per-channel decay rates for old evidence, allowing stable multi-timescale propagation without unbounded cache growth. Metric readout tokens extract stable global scale and pose directly from the running state.

Trained only on 48 frame clips, HorizonStream generalizes stably to sequences exceeding 10,000 frames. It runs with constant memory, linear time, and produces state of the art reconstruction quality. There is no measurable drift.

This is the exact capability every AR headset manufacturer has been begging for since 2020.

Video generators already know physics. They just weren't using it.

Modern video diffusion models produce beautiful footage. They also make objects float, pass through walls, change size between frames and violate every law of motion.

Until now every proposed fix involved attaching external simulators, training on labelled physics data, or running expensive correction passes. All of them hurt general generation quality.

LaMo demonstrates none of this is necessary. Every video diffusion model has already been trained on billions of hours of real world video. It has already seen every physically valid motion that exists. The correct motion prior is already inside the model. Nobody had bothered to extract it.

LaMo adds two lightweight readout heads totalling 112,000 parameters on top of any existing video diffusion backbone. No retraining of the base model is required. No external data. No simulator. It pulls out the implicit motion knowledge that was already present, and applies it as soft guidance during sampling.

On VideoPhy2 it reduces physical violation errors by 62%. It preserves general generation quality on all standard VBench metrics. You can drop this onto any existing production video generator next week.

Gaussian Splatting grew up.

4D Gaussian Splatting was the most impressive demo of 2025. It was also effectively useless for real world input. It melted moving objects. It jittered. It turned night scenes into unrecognizable glowing smears.

Two papers fixed almost all remaining flaws this week.

RiGS splits Gaussian primitives into three separate classes: static, rigid and transient. Each has different motion modelling behaviour optimized for their timescale. Static Gaussians never move. Rigid Gaussians track whole object motion. Transient Gaussians model high frequency deformation. The system automatically transitions primitives between classes based on observed motion. There is no more melting. There is no more jitter.

GlowGS fixes night scenes. Prior 3DGS methods failed in low light because they relied on visible edges and texture that do not exist under glow. GlowGS uses foundation model semantic features as implicit structural cues. It reconstructs street lights, car headlights and illuminated signs as cleanly as it reconstructs daytime scenes.

What this actually means

This is not incremental progress. This is the point where all the separate pieces click.

Twelve months ago you could not:

• Run streaming 3D reconstruction for more than two minutes without drift
• Generate video that followed an exact requested camera path
• Reconstruct dynamic scenes from monocular video without artifacts
• Get physically consistent motion out of any general purpose video generator

All of these things work now. All six papers have working reference code scheduled for release. None of them require larger models. All of them run on consumer hardware.

Most people will not notice this happened for another six to twelve months. Then suddenly every AR demo, every game engine, every video generator will be using one or more of these methods.

That is how progress works in this field. Nobody announces the turning point. You just look back one week and realize all the hard problems got solved while everyone was watching the big model launches.