All 8 papers dropped on arXiv last week. None of them are clickbait demo papers. Every one of them solves a real, painful problem that anyone building production CV systems has run into.

This is not a summary. I will tell you what works, what doesn't, what you can copy into your codebase this week, and which results you should ignore.

Stop fixing 3D reconstruction. Fix the attention layers instead.

MVTrack4Gen is the most important video generation paper released this quarter.

Everyone building novel view video generation has been stuck on the same tradeoff for two years: you either use explicit 3D and get garbage motion, or you use pure camera conditioned diffusion and get beautiful video that drifts geometrically after 12 frames.

Nobody had correctly diagnosed the failure mode until now. The authors found that motion drift is not caused by missing geometry. It is caused by misaligned correspondences inside the diffusion model's own attention layers.

Query features already attend to the correct corresponding points across views and time. They just are not trained to stay aligned. No external reconstruction module is required.

The fix is trivial. Route the output of those specific attention layers to an auxiliary point tracking head, add a standard tracking loss alongside the diffusion loss, and train end to end. That is it. No changes to the base diffusion architecture. No NeRF. No depth estimation.

This works. Geometric consistency error dropped 41% across all tested benchmarks. Camera trajectory following error dropped 28%.

You can add this head to any existing video diffusion model today. You do not need to retrain from scratch. Fine tuning for 2000 steps is sufficient.

Heterogeneous stereo deblurring: the problem nobody talked about

Every modern smartphone has two very different cameras. One is sharp. One is fast. One has f/1.5 aperture. The other has f/2.8.

When you capture stereo video, you get asymmetric blur. Every existing stereo algorithm assumes both cameras produce identical quality. All of them break completely in this case.

This paper fixes that. First they released a real dataset of 12,000 stereo pairs captured from 7 different production phones. This dataset alone is worth the read.

Then they built PECA: a 12 line cross attention modification that restricts cross view matching to a physically bounded epipolar window.

PECA adds 0.8% parameters to any baseline model. It improves PSNR by 2.1dB on the new dataset. It works unchanged on CNN, Transformer and NAFNet backbones.

This is not a state of the art chase. This is fixing a fundamental unstated assumption that broke every stereo method for the last 5 years.

Region drag editing fixes the biggest flaw in DragGAN

Everyone used DragGAN once. Nobody uses it for real work.

Point based drag is ambiguous. You drag two points. The model has 1000 different ways to satisfy that constraint. Almost all of them look wrong.

ICRDrag solves this properly. You do not drag points. You draw a mask of the region you want to move, draw a mask of where you want it to go. That is the full interface.

They added two extremely simple attention regularizers. First: the image and mask attention maps for the target region must match. Second: attention mass between source and target region must be symmetric.

That is the entire innovation. No new architecture. No fine tuning per image. It runs at standard diffusion speed.

User preference testing put ICRDrag 76% ahead of all existing drag editing methods. Code and weights are already released.

Medical VLM hallucinations are solved with one simple loop

Brain MRI diagnostic models hallucinate. A lot.

The standard approach is to feed the entire scan into the model once and ask for a diagnosis. They will happily report non existent tumors on perfectly normal scans 30% of the time.

BrReMark does something extremely obvious in hindsight.

1. Generate a list of candidate abnormalities
2. Draw a bounding box around each candidate
3. Re-run inference only on that bounding box
4. Only report the finding if it is confirmed on the second pass

That is it. They trained this loop with RL that rewards correct localization and penalizes hallucinations.

False positive rate dropped 45.7% on out of distribution pathologies. Diagnostic accuracy went up 19 percentage points.

No one had tried this before. Everyone was trying to build bigger single pass models.

Hand pose estimation finally gets absolute scale

All existing hand pose estimators output relative pose. They can tell you the angle of every finger. They cannot tell you how far away the hand is. They cannot tell you how big it is.

This makes them useless for AR and robotics.

ScaleHP fixes this without any depth sensor, without any stereo, from a single monocular RGB image.

The trick is extremely simple. Human hand bones have fixed proportional ratios. The ratio between the length of your proximal and middle phalanx does not change. It does not vary between people more than 4%.

The model solves for absolute scale by enforcing these anthropometric constraints inside the least squares pose solver.

It gets 35.8mm CS-MPJPE on FreiHand. That is 22% better than the previous state of the art. It runs at 90fps on an A16.

This is production ready right now.

Video segmentation now answers questions about the future

All existing referring video segmentation answers: "what is the thing doing right now?".

FeVOS asks: "what thing will be used next?".

This is not an incremental improvement. This is an entirely new task.

If you are building any kind of assistive AR system, this is exactly the capability you need. You do not want to segment the thing the user is holding. You want to segment the thing they are about to pick up.

They released a dataset with 14,524 annotated queries and chain of thought reasoning annotations. The baseline model they provide gets 62% mIoU on this task.

Human performance is 79%. There is a lot of headroom here.

Invertible image hiding stops looking terrible

Invertible image hiding lets you hide a full resolution image inside another full resolution image, such that you can perfectly recover the hidden image later.

All existing methods produce extremely obvious artifacts. You can always tell an image has something hidden inside it.

CrosInv fixes this by adding a single non invertible dense block inside the otherwise invertible network.

This is a beautiful architectural hack. The network remains almost perfectly invertible. But the small amount of non linearity added by that one block is enough to eliminate almost all visible artifacts.

PSNR of the cover image went from 36.2dB to 43.1dB. For reference, 42dB is indistinguishable to human observers.

DTI reconstruction gets proper diffusion models

TensorLDM is the first diffusion model that produces physically valid diffusion tensors for brain MRI.

All previous models would regularly produce tensors that are not positive semi definite. Which means they are physically impossible.

TensorLDM splits the 6 tensor components into two separate groups, runs them through separate encoders, and couples them with cross component attention.

SPD violation rate is 1.54%. That is almost identical to ground truth scans. Tractography produced from these reconstructions is usable for clinical work.

Benchmark summary

Paper	Core Metric	Prior SOTA	New Result	Relative Improvement	Production Ready?
MVTrack4Gen	Geometric Consistency Error	12.7	7.5	-41%	Yes
PECA	Stereo PSNR	27.2	29.3	+7.7%	Yes
ICRDrag	User Preference Share	24%	76%	+217%	Yes
BrReMark	OOD False Positive Rate	14.2%	7.7%	-45.7%	Yes
ScaleHP	FreiHand CS-MPJPE	45.9	35.8	-22%	Yes
FeVOS-R1	Foresight mIoU	41%	62%	+51%	Partial
CrosInv	Cover Image PSNR	36.2	43.1	+19%	Yes
TensorLDM	SPD Violation Rate	8.9%	1.54%	-83%	Partial

The unifying architectural pattern

None of these papers use bigger models. None of them trained on 10x more data. None of them announced a new foundation model.

Every single one of them follows exactly the same training structure.

Everyone stopped building entirely new architectures 6 months ago. All progress now comes from taking existing working base models, attaching one tiny auxiliary head, and training with an additional simple loss that enforces a known physical constraint.

This is the dominant architectural pattern for 2026.

Closing observations

None of these papers got any viral twitter threads. None of them have demo videos that got 10 million views.

Every single one of them will be running inside production systems 12 months from now.

That is the pattern. The important research never makes the front page. It just quietly makes everything work better.

If you only have time to read one paper this month, read MVTrack4Gen. If you have time for two, add ScaleHP. Both will change how you build systems.