EDGAR: Efficient Diffusion-Driven Gaussian Avatar Refinement via Synthetic Failure
EDGAR uses reference-guided one-step diffusion and synthetic avatar failures to repair Gaussian head avatars, improving detail, identity, and cross-view quality.
ls ./papers
Trending AI research papers from Hugging Face, summarized in plain language: what each paper does, why it matters and where to read more.
140 papers
A hybrid decision model learns whether extra reasoning will help, routing easy requests through a zero-token path and reserving autoregressive thinking for harder cases.
TURTLES tests neural PDE surrogates on turbulent 3D wakes, millions of irregular mesh points, and long rollouts—revealing where accuracy, stability, and speed still fail.
A new UAV stabilization benchmark shows how moving vehicle groups can hijack homography estimation—and why normal inlier statistics may fail to reveal it.
DeepStress tests search agents against unreliable, irrelevant, and false evidence, revealing why accuracy alone misses overconfident failures and wasted computation.
MIXUA gives multi-user, multi-agent systems four memory scopes, aiming to transfer useful lessons while preserving preferences and limiting private-data leakage.
UniSpark combines full and linear attention to unify image understanding, generation, editing, and interleaved context. Its strongest gains appear in multimodal understanding.
FOAM simplifies SOAP with input-only curvature and factored moments, aiming to deliver second-order optimization quality with near-AdamW cost.
A study of embodied agents finds that exposing benchmark criteria can raise scores by teaching the verifier, not the task—and that the advantage disappears under faithful evaluation.
A controlled study by Brown and colleagues finds that multimodal models need dedicated visual capacity, but not a separately pretrained or pre-fusion vision encoder.
WorldDraft tests whether video models can turn executable 4D simulations into realistic videos without breaking motion, structure, or target appearance.
TTB shifts cross-view attention into fast-weight MLPs, delivering constant-complexity novel-view rendering while retaining much of decoder-only synthesis quality.
Ropedia Xperience 2 links LiDAR scene scans with egocentric human activity, creating reusable assets for robot simulation, perception benchmarks, and long-horizon behavior generation.