If you are reviewing for @cvprconference.bsky.social and going to print the submission, use this command to save ink and crop those black rectangles from the top pdfcrop --margins "0 -60 0 0" input.pdf output.pdf

And yeah, others have also expressed that they don't want filtering for non-research topics. We are multidimensional human beings with hobbies and political thoughts, and it's nice to talk about those things socially! (this reply is brought to you by a different, non-bluesky app!)

So people are delivering text and image content on Youtube just because Youtube is the only convenient way to monetize the content?

My Christmas tale about the data -- Beer or Trdelnik. That is a story of my teaching at Vision and Sports school. I have told this story privately for years, time to go public blog.dmytro.ai/posts/2025-1...

My Christmas tale about the data -- Beer or Trdelnik. That is a story of my teaching at Vision and Sports school. I have told this story privately for years, time to go public blog.dmytro.ai/posts/2025-1...

We can have problems with robustness, speed/ cost, etc, but the scene understanding is essentially solved with VLMs. You can give the code of the renderer to Opus, screenshot of the render with a comment: "You see the issue, right?" -- AND IT PROCEEDS WITH FIXING THE ALGORITHM.

I'm working on a piece about peer review and conference publication in AI/ML/NLP/CV, and I need YOUR help to understand how the community uses conferences and peer review in paper discovery. Your four short multiple choice answers will help a lot :)

Speedrunning ImageNet Diffusion Swayam Bhanded tl;dr: in title, list of tricks to speed-up DiT training (yes, based on ready representation REPA). A lot of ablations, W&B runs available. arxiv.org/abs/2512.12386

Grab-3D: Detecting AI-Generated Videos from 3D Geometric Temporal Consistency Wenhan Chen, Sezer Karaoglu Theo Gevers tl;dr: vanishing point are still good way to detect AI-generated stuff. arxiv.org/abs/2512.13665

Patch-wise Retrieval: A Bag of Practical Techniques for Instance-level Matching Wonseok Choi, Sohwi Lim, Nam Hyeon-Woo, Moon Ye-Bin, Dong-Ju Jeong, Jinyoung Hwang, Tae-Hyun Oh tl;dr: CLIP/DINO -> 4x4 grid -> PQ. Old tricks make come back? No BoW or ASMK arxiv.org/abs/2512.12610

"Everyone knows we should write fewer papers, and yet it feels like career suicide to do so. "

An excellent article describing how LLMs fit into the bureaucratization of science: open.substack.com/pub/artifici...

OK, M1 Pro GPU is trash ( or the Metal drivers for PyTorch, which is the same) - I run 10 images segmentation and matching and now laptop screen is all pink/purplish. Force restart helped, thankfully

Sharp Monocular View Synthesis in Less Than a Second Lars Mescheder, Wei Dong, Shiwei Li, Xuyang Bai, Marcel Santos, Peiyun Hu, Bruno Lecouat, Mingmin Zhen, Amaël Delaunoy, Tian Fang, Yanghai Tsin, Stephan R. Richter, Vladlen Koltun tl;dr: feedforward 3DGS arxiv.org/abs/2512.10685

Evaluating Foundation Models 3D Understanding Through Multi-View Correspondence Analysis Valentina Lilova, Toyesh Chakravorty, Julian I. Bibo, Emma Boccaletti, Brandon Li, Lívia Baxová, Cees G. M. Snoek, Mohammadreza Salehi tl;dr: do NOT use VGGT for segmentation arxiv.org/pdf/2512.11574

On Geometric Understanding and Learned Data Priors in VGGT Jelena Bratulić, Sudhanshu Mittal, Thomas Brox, Christian Rupprecht tl;dr: all magic happens in the middle layers (as tested on the ShapeNet) arxiv.org/abs/2512.11508

Are you tired of context-switching between coding models in @pytorch.org and paper writing on @overleaf.com? Well, I’ve got the fix for you, Neuralatex! An ML library written in pure Latex! neuralatex.com To appear in Sigbovik (subject to rigorous review process)

If we will have EurCVPR (separate from ECCV, and held in Europe), I am willing to review 10 papers for it.

In machine learning, it is impossible to disentangle technical innovation from technical debt.

Check the whole thread for sure, but this is the most surprising result to me

Emergent Outlier View Rejection in Visual Geometry Grounded Transformers Jisang Han, Sunghwan Hong, Jaewoo Jung, Wooseok Jang, Honggyu An, Qianqian Wang, Seungryong Kim, Chen Feng tl;dr: distractor images have low attention scores in VGGT, exploitable. arxiv.org/abs/2512.04012

ControlVP: Interactive Geometric Refinement of AI-Generated Images with Consistent Vanishing Points Ryota Okumura, Kaede Shiohara, Toshihiko Yamasaki tl;dr: ControlNet from (user-provided) VP. Also: CFG hurts when ControlNet signal is rich arxiv.org/abs/2512.07504

You know what season it is! Right. Internship application season. Niantic Spatial is offering research internships on a multitude of 3D vision topics: relocalization, reconstruction, 3D VLMs... Top tier papers regularly come out of our internships 🚀 nianticspatial.careers.hibob.com/jobs/0fc4871...

Emergent Outlier View Rejection in Visual Geometry Grounded Transformers Jisang Han, Sunghwan Hong, Jaewoo Jung, Wooseok Jang, Honggyu An, @qianqianwang.bsky.social, Seungryong Kim, Chen Feng tl;dr: threshold internal attention and feature signals arxiv.org/abs/2512.04012

Nothing unusual. It's just that for russians, murder is a holiday. #UkrainianView

I'm gonna miss the entire #NeurIPS2025... 🥲 If you are attending NeurIPS'25 and find peer-review related topics, let me know Specifically, I want to know - any interesting ideas exchanged for improving the system - reactions from workshops targeted by a single author who submitted about 80 papers

Scholar Inbox for NeurIPS is live now.

Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation Weining Ren, Hongjun Wang, Xiao Tan, Kai Han tl;dr: free decoder & re-normalization LoRA adapter->encoder & monocular teacher+unlabeled dataset->distill model arxiv.org/abs/2511.22429

Emergent Extreme-View Geometry in 3D Foundation Models Yiwen Zhang, Joseph Tung, Ruojin Cai, David Fouhey, Hadar Averbuch-Elor tl;dr: how to finetune VGGT or Pi3 on non-visually overlapping data: unfreeze bias in backbone, freeze heads. arxiv.org/abs/2511.22686

Seeing without Pixels: Perception from Camera Trajectories Zihui Xue, Kristen Grauman @dimadamen.bsky.social Andrew Zisserman, Tengda Han tl;dr: in title. I love such "blind baseline" papers. arxiv.org/abs/2511.21681

WTF is happening with PDF reader in safari? Not for all figures. Safari - broken, Chrome - fine.

Captain Safari: A World Engine Yu-Cheng Chou, Xingrui Wang, Yitong Li, Jiahao Wang, Hanting Liu, Cihang Xie, Alan Yuille, Junfei Xiao tl;dr: dataset and method for consistent video-from-drone generation. Code & data "coming soon". arxiv.org/abs/2511.22815

Match-and-Fuse: Consistent Generation from Unstructured Image Sets Kate Feingold, Omri Kaduri, Tali Dekel tl;dr: Input image set -> RoMa matches -> joint denoising, enforcing the correspondences -> multiview consistent output. Based on Flux. arxiv.org/abs/2511.22287

JMLR addresses a different community though. I doubt you can publish in JMLR by submitting AwesomeModel-V3 with an improved architecture, training recipe, etc. This is mostly what we see at confs, which some will say is incremental, but also the core of what made the field advance in the last 10y.

Ok this is the best counter against non-blind reviews I've seen so far. It would be possible to detect collusion. But who is going to do that?

Regarding fully-non-blind reviews. I am not afraid of putting my name on my review.You are not afraid So far,so good. But, imagine some amount of bad actors, who want to exploit the system - it will be super easy for them to operate in fully open.Collusion ring - you can create new every time. 1/

Do NOT believe the country analysis based on the THE FILE. It is either name-based guessing, or something, not related to current or previous institution.

Update from #ICLR2026 Program Chairs

I hope that the entire community got the memo that ICLR ACs will now be changed and papers re-assigned. Else, there could be additional fallout. The new AC of a paper of my current batch takes a controversial decision and I will take the blame among the people who have access to "THE FILE".

I wonder if the OpenReview dataset gets released, would it help to train LLMs to do better reviews? Also given that we have quality assessments from the follow-up papers for the past conferences.

she shared excellent insights on what this implies (your brain runs probably faster on those questions, when you a future CVPR PC 😅): - Nothing will ever be blind anymore because there's enough data in the wild to train a reviewer de-anonymizer - What about fine-tuning a seq2seq model conditioned

… and where some feel pressured enough to lose sight of basic ethics (eg reports of authors contacting reviewers to change their score or much worse if social media reports are to be believed). Perhaps it is time to rethink our conference culture. Share your thoughts in thread. 2/2

TIL you can cook wheat in a rice cooker and it is amazing

OK, my arXiv debt is paid off for now.

Illustrator's Depth: Monocular Layer Index Prediction for Image Decomposition Nissim Maruani et 8 al. tl;dr: in title. new dataset -> finetuned DepthPro model. arxiv.org/abs/2511.17454

Understanding Multi-View Transformers Michal Stary, Julien Gaubil, Ayush Tewari, Vincent Sitzmann tl;dr: DUSt3R self-attention is it secretly a diffusion model, and cross-attention is matching. arxiv.org/abs/2510.24907

Room Envelopes: A Synthetic Dataset for Indoor Layout Reconstruction from Images Sam Bahrami, Dylan Campbell tl;dr: in title. Not available yet, but looks cool sambahrami.com/room_envelopes arxiv.org/abs/2511.03970

Toward A Better Understanding of Monocular Depth Evaluation Siyang Wu, Jack Nugent, Willow Yang, Jia Deng tl;dr: how different metrics sensitive to perturbations (intrinsics, alignment, normals, metric-vs-relative), new metric aligned to humans. arxiv.org/abs/2510.19814

Unlocking Zero-shot Potential of Semi-dense Image Matching via Gaussian Splatting Juncheng Chen, Chao Xu, Yanjun Cao tl;dr: train LoFTR purely on the 3DGS renderings -> better than original LoFTR. arxiv.org/abs/2511.21265

PLANA3R: Zero-shot Metric Planar 3D Reconstruction via Feed-Forward Planar Splatting Changkun Liu, Bin Tan, Zeran Ke, Shangzhan Zhang, Jiachen Liu, Ming Qian, Nan Xue, Yujun Shen, Tristan Braud tl;dr: MAST3r, but predicts planar patches instead of pointmap arxiv.org/abs/2510.18714