Computer Vision
Research on Computer Vision: explanations, evidence, limitations and connections between papers.
28 results
Match visual compression to computation and cache memory
An image question may seem easy after most visual tokens are removed, yet a small word or an object's location can still be lost. A separate constraint appears during answer generation, when cached visual information occupies memory. The first useful step is to identify the constrained stage. Reducing visual tokens changes the representation the model processes; evicting cached visual information changes what later generated words can draw on. The studies test these choices with different models and questions.
What Compression Changes in Video and Audio-Visual Systems
A video generator can take too long to produce a clip, while an audio-visual classifier may need a large collection of paired examples for training. Both invite compression, but they put different resources under pressure. The studies suggest a bounded way to judge it: name the object made smaller, then read the measured task outcome in its own setting. Video compression has reported latency and generation-quality evidence; a compact synthetic dataset has classification evidence. A language-generation design clarifies the computation tradeoff without establishing a vision result.
TableVision uses printed sums to check scanned table structure
TableVision uses printed totals to check scanned tables, improving figure placement on tested clean election panels.
One signal prunes visual tokens across two stages
TwinPrune reuses spatial density to prune visual tokens and cache entries across two inference stages.
VideoJev scores video decisions directly and reuses context
Across event-location and video-question tests, VideoJev improves mean scores over corresponding instruction-model backbones.
Small samples predict large image retrieval scores
A tail-aware method predicts full-gallery retrieval scores from random samples, matching full-pool results in tested image datasets.
Video scores shift with setup, so Meta-Video records the settings
Meta-Video largely preserves video-model rankings while reducing compute for a stated frame-count sweep across source tests.
A projector can align vision and language without captions
PORTAL builds a vision-language projector from separate feature statistics and stays near trained baselines across seven tested model combinations.
A regression method distills generative models into fewer steps
SMD distills diffusion and flow models into fewer steps, with stronger prompt alignment in a tested image-model comparison.
Bifocal preserves image detail and context while trimming visual input
Bifocal combines question-focused patches, image-wide evidence and pooled context, preserving performance across tested visual tasks.
Synthetic night images change test-set model rankings
A three-stage image method adds missing night conditions and shifts test-set model rankings toward those seen with real night images.
A shared interface lets image generators exchange denoising states
InstantFusion transfers denoising states across image generators, improving image-generation scores for tested Qwen-Image combinations.
LapMerge makes ordered token merging learnable
LapMerge merges ordered tokens to a fixed budget; it improved object-question scores over selection at tighter token limits.
Disaster damage scores drop when a whole event is held out
An audit finds familiar-event scores overstate new-disaster transfer, while source-only feature changes help unevenly.
Two-way attention helps compact audio-visual data stay useful
X-AVDD uses two-way cross-attention to create compact audio-visual training data, with higher accuracy than AVDD in a VGGSound test.
Suppressing image-to-text feedback improves generation
A training-free method suppresses image-to-text feedback, improving tested image and video generation scores in bounded evaluations.
Human videos can improve robot policies after alignment
LAPT-Scale turns human-video changes into robot-policy supervision and raises tested real-robot success as video grows.
Video compression cuts generation time while preserving Wan2.1 scores
In Wan2.1 tests, GRACE uses fewer video tokens while keeping generated-video quality scores near the uncompressed model.
Two adapters let audio–video generation stream in a few steps
A parallel adapter method combines streaming and few-step audio–video generation, matching a bidirectional system’s image quality in held-out tests.
Patch-level visual learning scales with a simpler objective
JEM aligns image patches across views and performs strongly in frozen-feature tests, especially on dense visual tasks.
Evoke uses bounded scene memory for long-video sessions
Evoke combines bounded scene recall with long-horizon training; tests show a specific brightness-stability gain, not universal consistency.
TopoGPU streams image topology calculations through GPUs
TopoGPU speeds image-based persistent homology against CRipser, while CPU memory limits some very large cases.
One model learns faster image generation and understanding
A two-stage method cuts decoding steps while improving image and captioning results for a discrete multimodal model.
A world model checks whether GUI actions lead as expected
LGWM checks GUI transitions by comparing predicted and observed screens, and distinguishes harmful from benign mismatches in tested cases.