DreamStyle3D reports generating a stylized 3D asset within 10 seconds by separating geometry from visual style inside a native 3D pipeline. This morning also brings arbitrary camera control for existing video, a film-language benchmark, faster long-form avatars, visual-state training for interface agents, tests of agent memory and self-verification, tactile world models, real-robot evaluation, and a new PAX East 2027 calendar entry. Unless noted otherwise, the results are author-reported from new preprints and have not been independently replicated.

What Changed Overnight

  • ArXiv’s July 28 announcement batch added new work on stylized 3D assets, camera-controlled video, cinematic evaluation, avatars, interface agents, memory, code repair, world models, and robot evaluation.
  • PAX confirmed that PAX East 2027 will run April 22–25 in Boston. General exhibitor enquiries are open, while attendee badge sales have not been announced.

Game Assets And Cinematic Control

DreamStyle3D reports native stylized 3D generation within 10 seconds

Kai Wang, Ziheng Ouyang, Xuying Zhang, Ming-Ming Cheng, and Qibin Hou built DreamStyle3D around separate cross-attention paths for geometry and style. The design injects a requested appearance while preserving the object’s structure, reducing reliance on the indirect image-to-3D route used by many stylization systems.

The team assembled about 15,000 content-style-stylized triplets through an automated pipeline. It reports producing high-fidelity, geometrically consistent assets within 10 seconds and has released code and data. The abstract does not specify the hardware used for that timing or report how the outputs behave after rigging, animation, or import into a commercial game engine.

CameraAnything changes viewpoint, lens, and format in one video edit

Yixuan Li and collaborators introduced CameraAnything, an ECCV 2026 system that edits an existing video with joint control over camera position, focal length, and output resolution. The system injects per-pixel camera rays into the target video representation and can produce a new viewpoint, lens change, aspect ratio, or multi-shot transition without first reconstructing a complete 3D scene.

The researchers generated paired training data by recording synthetic dynamic scenes with multiple cameras, focal lengths, and shot structures. The project demonstrates 16:9 source video adapted to 4:3, square, portrait, and 9:16 outputs, but the paper abstract does not report one aggregate camera-accuracy score.

FilmBench tests whether generated video follows cinematic language

FilmBench replaces generic web prompts with 1,169 prompts reverse-engineered from award-winning live-action films across 20 genres. Directors helped select the references, and 1,056 prompts require multiple shots rather than one continuous clip.

The benchmark scores nine text-to-video and seven reference-to-video models across three axes, 12 components, and 35 core submetrics. Its automatic evaluator reproduces human model rankings with Spearman correlations of 0.95 and 0.96, while the authors report a consistent decline from single-shot to multi-shot tasks.

AptAvatar reports a 60-fold speedup for long-form 720p avatars

Hengyuan Zhang, Jingna Sun, Meiguang Jin, and Junfeng Ma built AptAvatar as a 14-billion-parameter audio-driven video system. It distills generation into two denoising steps and reuses earlier model outputs during training so that the avatar learns to continue from its own history.

The authors report 720p long-form generation with a 60-fold inference speedup while preserving identity and motion quality. Code is available, although the abstract does not state the serving hardware, real-time frame rate, or results from viewers evaluating lip synchronization.

Interface Agents, Memory, And Verification

Visual state transitions give GUI agents a second training stage

Xiangyan Liu and collaborators pretrain one multimodal model to infer the action between two interface states and to predict the next state from an action. The same model is then fine-tuned on task instructions and complete trajectories.

The authors report consistent gains over direct trajectory fine-tuning on AgentNetBench, AndroidControl, and GUIOdyssey, covering desktop and mobile interfaces. Performance also rises with more transition data. The abstract does not provide a single pooled improvement, so comparisons remain benchmark-specific.

InMind exposes a retrieval blind spot in long-term agent memory

InMind contains 125 expert-verified tasks across ten life domains. Each task requires a stored fact to affect a later request through background knowledge—for example, connecting a nut allergy to almond flour in a dessert request—even when the memory and query share few words.

With the decisive memory kept in context, the tested backbone answers 84.0% of indirect queries. When the same fact must be fetched, six vector, graph, and agentic memory systems reach at most 14.4%, despite recalling directly requested facts at rates as high as 100%. Increasing embedding dimensionality eightfold improves target recall but leaves the answer gap largely intact.

Repeated code revision can overwrite a correct repair

Xueping Gao, Jianwei Yang, and Qiang Yang study 900 three-revision HumanEval repair trajectories. Under forced revision, programs that are both currently correct and supported by current traces fall from 82.0% after one revision to 67.3% after two, even as the share that had been correct at some point rises to 84.7%.

In a separate 14-billion-parameter replication, stale traces harm 34 of 135 correct starting programs, compared with four of 135 when traces match the current code. The authors provide a reference contract that binds verifier evidence to an exact program state, but they explicitly do not claim that it improves the agent’s repair skill.

A sealed acceptance test catches self-improvement regressions

Diandian Guo and collaborators examine agents that rewrite both a heuristic policy and the tests used to approve later edits. Self-authored scores can remain high while performance on an unseen deployment evaluation falls.

Their Sealed Exogenous Acceptance Loop retains the agent’s own tests but compares each candidate with the current version through a harness-side audit the agent cannot inspect. Across six models and three random seeds, the method outperforms unprotected baselines. The paper argues that at least one acceptance decision should remain outside the agent’s control.

World Models And Physical Evaluation

FeelWorld predicts touch before planning contact

FeelWorld predicts future visual features alongside contact, a three-dimensional force-related tactile representation, and slip. A gate keeps irrelevant touch readings out of visual prediction before contact, then joins the two streams when interaction begins.

Across chip grasping, fruit grasping, and USB insertion, the authors report that ten-step visual error falls from 0.084 to 0.058 LPIPS. After an 80-step rollout, error is 61% below a visual-only comparison. Contact-aware planning reaches 81.7% average zero-shot success.

ArmnetBench runs 2,518 policy trials on a low-cost robot farm

ArmnetBench v0.1 evaluates seven policies across 12 tasks, using both single-arm and bimanual setups on a fleet of low-cost SO-101 robot cells. Every policy receives 50 demonstrations per task under the shared setup.

The released collection contains 2,518 policy rollouts and 600 reference demonstrations. Human raters label each of the 3,118 episodes as successful, suboptimal, or failed, creating a common real-world comparison set and mixed-quality trajectories for later policy or reward-model training.

WorldDiT joins action generation and future-frame prediction

WorldDiT uses one diffusion transformer to generate continuous robot-action chunks and predict normalized image patches from future camera frames. It does so without using a large pretrained vision-language model as the action backbone.

Across four LIBERO simulation suites, the authors place the sub-billion-parameter model on the reported frontier between parameter count and mean task success among methods that disclose all four results. The initial report offers a compact baseline, not evidence of performance on physical robots.

SHARE cuts AR latency in a shared human-robot map

SHARE prioritizes the map updates needed by a person wearing an augmented-reality headset while preserving robot tracking in the same space. It also removes redundant visual features before edge processing.

A deployment with commercial AR headsets and a ground robot records 13.22 milliseconds of average AR latency, 43.3% below the comparison, while keeping tracking error below two centimeters. The user study reports statistically significant perceptual improvements, although the abstract does not state the participant count.

Events Calendar

PAX East 2027 opens exhibitor enquiries

PAX East will run April 22–25, 2027, at the Thomas M. Menino Convention & Exhibition Center in Boston. The organizer lists general exhibitor and advertising enquiries as open on an ongoing basis. Applications for PAX Rising and other programs are scheduled for late 2026, and individual attendee badges have not gone on sale.

See the 2027 games and education-technology calendar.

Watch Next

  • DreamStyle3D’s released code and data can show whether the reported timing transfers across GPUs and whether generated meshes are ready for rigging and engine import.
  • CameraAnything and FilmBench still need public results that connect camera-control accuracy and multi-shot scores to finished creator workflows.
  • InMind makes memory routing measurable; the next question is whether systems can keep indirectly relevant facts available without flooding the model’s context.
  • ArmnetBench’s public episodes can test whether policy rankings remain stable as more hardware sites, operators, and tasks join the evaluation.

This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.