Perceptual Distortion for Overfitted Video Coding reports 9.2% BD-rate reduction in DISTS and 3.0% in LPIPS on UVG and HEVC Class B. The paper introduces a local sliced Wasserstein distortion (SWD) for spatial matching and a flow residual for temporal error alignment, evaluated with the COOL-CHIC codec.
Key points
BD-rate reductions: 9.2% in DISTS, 3.0% in LPIPS, averaged over test sequences.
Temporal consistency metric (tOF) lowered by 10.2% at matched rate.
Sources
A Proxy-Based Method for Measuring Racial Disparities in Text-to-Image Generators
A Proxy-Based Method for Measuring Racial Disparities in Text-to-Image Generators introduces an automated audit that classifies generated images by race and compares distributions to demographic benchmarks. Experiments on Imagen 4 Fast, Stable Diffusion 3.5 Large Turbo, and Flux Schnell show model-specific over- and under-representation patterns and skin-tone variations.
Key points
Imagen 4 Fast overrepresents white classified images and produces lighter, more homogeneous skin tones.
Stable Diffusion 3.5 Large Turbo diverges most from census data but yields the most diverse range of skin tones.
Sources
MuJoCo 4.0 release – fast objects no longer pass, soft bodies deform
MuJoCo 4.0 is out with fast or thin objects no longer passing through each other and soft bodies that can deform permanently. It also adds MuJoCo Studio with physically based rendering via Filament.
Key points
Fast or thin objects no longer pass through each other.
Soft bodies can now deform permanently and rods can twist.
MuJoCo Studio replaces simulate and uses Filament for physically based rendering.
Sources
- Original post
- mujoco.org - Linked in the post
SplatAnything launches 3D Gaussian conversion tool
SplatAnything converts everyday photos and aerial captures into 3D Gaussian scenes in a few seconds. A project page and a Huggingface space are provided for the tool.
Key points
Transforms simple captures into 3D Gaussian scenes in a few seconds.
Huggingface space for SplatAnything is available.
Sources
- Original post
- zju3dv.github.io - Linked in the post
- huggingface.co - Linked in the post
DAWN foundation model introduced by PXLD with 360° control
DAWN is a new foundation model announced by PXLD. It claims exact performance as shot, 360 degree world control without text prompts, and unlimited clip length for film and VFX.
Key points
Exact performance as shot, not regenerated
Unlimited clip length and scenes for film and high-end VFX
Sources
DAWN foundation model for world harmonisation launch
DAWN is a foundation model that places a video of a person into any photo, video, 360 or 3D scene, rebuilding light, shadows and reflections to match. It keeps the original performance and identity unchanged, with no regeneration of the person.
Key points
One-shot videos: every shown video is single-take, no re-running with different seeds.
Full light control: lights can be placed anywhere in the world, including behind the camera.
Sources
- Original post
- pxld.ai - Linked in the post
- pxld.ai - Linked in the post
- pxld.ai - Linked in the post
- pxld.ai - Linked in the post
NEO-VLA replaces fixed encoder with patch conv layers
NEO-VLA swaps a conventional fixed visual encoder for lightweight patch convolutional layers that evolve jointly with language. The flow-matching action expert can use instruction-conditioned visual representations from multiple depths. Results show native architectures outperform encoder-based counterparts and retain advantages under distribution shifts.
Key points
Replaces conventional fixed visual encoder with lightweight patch convolutional layers.
Flow-matching action expert draws on instruction-conditioned visual representations from different depths.
Sources
Wide View Synthesis from Overlapping Perspective Cameras – WV-GS announced
Wide View Synthesis from Overlapping Perspective Cameras in Autonomous Driving introduces Wide-View Gaussian Splatting (WV-GS), a feed-forward Gaussian-splatting framework that synthesizes wide views without per-scene optimization. The paper also releases the synthetic benchmark WideDrive and a LiDAR-projected sparse-reference protocol for real-world datasets.
Key points
Method: Wide-View Gaussian Splatting (WV-GS) uses Piecewise Perspective Rendering to reduce peripheral stretching.
Benchmark: WideDrive provides dense ground-truth wide views for evaluation.
