Computer Vision
Research on Computer Vision: explanations, evidence, limitations and connections between papers.
53 results
A game engine keeps playable worlds coherent as video adds detail
Matrix-Game 4.0 uses an engine to keep game state coherent as video renders detailed scenes in three tested games.
Smaller vision models can outperform larger ones on specific tasks
Tests find recurring category-specific declines in larger vision-language models, though released variants do not isolate model size.
A video renderer follows new scene formats in four steps
D SolarRenderer renders structural videos in four steps and follows tested unseen formats, though results vary by format.
Attractive faces shape image rankings and scores
An audit finds attractiveness linked to face rankings and image scores, and tests an adjustment that weakens the score association.
Point trajectories make multi-view forecasts more coherent
Video-Point Model tests whether tracked 3D point paths improve future video consistency across simulated and real scenes.
Brush strokes remove marks that common edits leave behind
BrushOff redraws paintings as strokes, removing tested VINE marks more effectively than common distortions while preserving closer resemblance.
A video framework applies world rules before rendering actions
Stateful Harness separates rule execution from video rendering, improving required outcomes in tested hidden-state interactions.
A small corrector reduces drift in long generated videos
Clean Forcing reduces measured quality drift in 50-second video generations with a small corrector trained on a frozen base model.
Repeated facts can distort document evidence tests
An audit finds repeated facts can distort long-document retrieval, citation scoring, and masking-based attribution when tests mark one location.
Token choice can change measured hallucination scores
A POPE audit finds that tokenizer-aware yes/no extraction changes measured scores, while fixed token IDs fail on some models.
Early signals can flag risky image prompts
CSTI reads early image-generation signals and had the highest unsafe-prompt detection rate in most tested settings.
A camera system shares edge work and adapts models centrally
YUTELA shares camera processing at the edge and supports central updates, but accuracy gains remain unmeasured.
Match generation shortcuts to the work they actually skip
A shorter generation process may save computation yet leave the user waiting on other work. It may also damage the part of an output that matters most: a molecule's geometry, a caption's coherence, or a newly revealed video region. The studies suggest choosing a shortcut by the work it changes. Training can shorten a shared sequence of refinements; an inference procedure can stop refining selected parts. Their reported counts and quality results belong to different tasks, so a product choice still needs delay and quality measured on its intended workload.
Match the test boundary to the transfer claim
A damage classifier may score well when test images come from familiar disasters, while a new disaster presents a different test. An image question-answering model may follow a scene clue that points to the wrong answer. For a robot, the outcome is whether its visual guidance helps complete a task. The evidence becomes useful when the claimed shift, the test boundary, and the measured outcome match. The studies examine those pieces through held-out disasters, selection for misleading target clues, and success on specified robot tasks.
A model’s cue sensitivity can change which model is best
SRA uses cue sensitivity to select image models, lowering pooled error on new misleading image tests.
Anvil links spatial learning across related tasks
Anvil links spatial examples and specialist guidance; its model scores higher on average across seven public spatial tests, with a task-specific regression.
OpenGCUA lets computer agents coordinate screen and shell actions in code
OpenGCUA combines screen and shell actions in code, raising task-completion scores over batched screen-only actions on computer tests.
Video generators can stop refining tokens that have settled
Speculative Forcing skips settled video tokens early, preserving aggregate quality in tested LongLive and Lingbot generators.
SurgStreamer times alerts for events in surgical video
SurgStreamer separates video perception from event reasoning; sealed surgical tests show better alert matching than exact rules.
Fixed-work FPGA design bounds two occupancy steps, not the full pipeline
FALCON makes two camera occupancy operations cycle-bounded, while full-pipeline timing and vehicle use remain unproven.
Loud and Clear helps generators use retrieved page evidence
A training-free method combines larger pages and selected text, improving answers across three document question sets.
A smaller contrastive temperature helps CLIP retain accuracy
ComCLIP uses a smaller contrastive temperature to improve CLIP post-training while preserving zero-shot accuracy in a tested setup.
ZoomCap routes image evidence into richer captions
ZoomCap routes image evidence into captions; a two-anchor test found more source-supported, reconstructable claims than ScaleCap.
Synthetic game practice improves Minecraft task success
Player-1 adds synthetic game practice after gameplay examples, improving one model's average success on 300 Minecraft tasks.