Transformers News
The latest Transformers coverage curated for game developers.
arXiv cs.GR details PART, a transformer-based system that can retrieve 3D parts from a library and assemble them into a target shape. The approach tackles variable part counts and 6-DoF …
ACM Transactions on Graphics has published an erratum for OmniHands, the hand-motion capture paper aimed at interactive hand tracking. The correction lands in Volume 45, Issue 5, dated October 2026. …
ARDY is a new streaming motion-generation system that aims to make real-time humanoid animation both controllable and high quality. It combines autoregressive diffusion with a hybrid motion representation, letting interactive …
RenderFormer++ tackles feed-forward neural global illumination with two practical upgrades: physics-informed transport guidance and hierarchical object-centric tokenization. The result is a model aimed at better cross-scene generalization while cutting the …
A SIGGRAPH 2026 paper reports a 99.98% motion-reproduction success rate using a tokenized, GPT-style controller trained on large motion datasets. For animation teams, the interesting part is the shift from …
A new micro-expression recognition paper proposes a dual-branch model that combines graph attention with a transformer, instead of leaning on apex-onset frame pairs. The interesting part for developers is the …
MeshPad turns rough sketches into editable 3D meshes, and the practical win is interactive iteration: each edit can complete in a few seconds. For graphics teams, the interesting part is …
A new 3D light-transport embedding aims to predict global illumination directly from scene geometry and materials, instead of leaning on screen-space cues or per-scene training. For graphics teams, the interesting …
A new research tool lets artists poke at video diffusion transformers in real time, not just via prompts. AttentionBender exposes self-attention, cross-attention, feed-forward blocks, layers, tokens, and even hidden neurons …
Cloth simulation is getting a Transformer treatment that aims to scale beyond one-off setups. ClothTransformer models garments, robotic manipulation, and free-fall collisions in a shared latent space, with reported error …
A new token-selection scheme cuts the cost of visual geometry transformers by over 85% on 500-image scenes, while keeping accuracy flat or better. The paper splits pruning into frame-level and …
For animation teams, the interesting bit is a retargeting system that keeps self-contact and near-body interactions intact even when characters have very different proportions. The paper replaces fixed correspondences with …
A transformer-based method for N-sided hole filling reports a 12% gain in tolerance-satisfaction over industrial and academic baselines. For graphics and technical artists, the interesting part is that it predicts …
RigidFormer introduces a novel approach to simulating rigid-body dynamics using Transformers, significantly enhancing efficiency and accuracy. This model allows developers to work with mesh-free inputs, which is crucial for modern …
A new diffusion-transformer sampling trick aims to fix the “samey output” problem without the usual quality hit. By pushing repulsion into the model’s contextual attention space during the forward pass, …
A new BIM-native layout model turns room design into token sequences, aiming to generate constraint-aware spaces from building data instead of generic scene descriptions. The system uses wall-referenced coordinates and …