Skip to main content
GameDev.net gamedev.net

Diffusion Models News

The latest Diffusion Models coverage curated for game developers.

DSD: Learning Diverse and Reusable Motor Skills via Diffusion Skill Discovery

arXiv cs.GR details DSD, a diffusion-based skill discovery method that learns a wider set of reusable humanoid motor skills. For game teams building animation or AI-driven characters, the pitch is …

Research arXiv cs.GR · 2 days, 2 hours ago
Learning Interaction between Image and Layout Priors for Joint Image-Layout Generation in Design Templates

A new template-generation model, InterIL, jointly creates a background image and foreground layout from text instead of building them in sequence. For UI, marketing, and tools teams, the practical hook …

Research arXiv cs.GR · 1 week, 1 day ago
SceneHI: High-Resolution 3D-Consistent Scene Texturing with Controllable Illumination

SceneHI brings high-resolution, 3D-consistent scene texturing with controllable illumination into a single generative pipeline. The system lifts priors from 2D diffusion models onto 3D objects, then bakes geometry-aware shadows directly …

Research arXiv cs.GR · 1 week, 2 days ago
UniMate: One Unified Model to Animate Diverse Skeletons

UniMate aims to remove a major bottleneck in character animation: motion generation for arbitrary rigs. The model takes a rigged 3D asset and a text prompt, then produces articulated motion …

Research arXiv cs.GR · 1 week, 5 days ago
LightBridge: Feed-Forward Generative Relighting for 3D Gaussian Splatting

LightBridge aims to make relighting 3D Gaussian Splatting assets a one-pass process instead of a per-scene optimization job. The system uses a feed-forward generative pipeline plus a new relighting dataset …

Research arXiv cs.GR · 2 weeks, 2 days ago
DReSG: Diffusion Residuals for Stylized Gaussian Splatting

DReSG tackles a familiar 3D Gaussian Splatting problem: how to add strong reference-driven style without wrecking view consistency. The method turns diffusion outputs into residual targets and feeds them back …

Research arXiv cs.GR · 2 weeks, 4 days ago
A Framework for Low-Effort Training Data Generation for Urban Semantic Segmentation

A new framework cuts the cost of building urban segmentation training data by turning rough synthetic scenes into target-aligned images. It adapts an off-the-shelf diffusion model with only imperfect pseudo-labels, …

Research arXiv cs.GR · 3 weeks, 1 day ago
EditStream: A Unified Autoregressive Framework for Interactive Video Generation and Editing

EditStream folds text-to-video, image-to-video, video-to-video, editing propagation, reference-guided edits, and camera pose changes into one DiT-based system. The big shift for developers is its few-step autoregressive streaming setup, aimed at …

Research arXiv cs.GR · 3 weeks, 4 days ago
Erratum: Loops2Roofs: Diffusion-based 3D Roof Generation using a Loop Representation

ACM Transactions on Graphics has published an erratum for Loops2Roofs, the diffusion-based roof generation system built around a loop representation. The correction lands in Volume 45, Issue 5, dated October …

Research ACM Graphics · 3 weeks, 4 days ago
Generalized Audio-Driven Synthesis of Precise Drummer Motion

Disney Research Studios has built a diffusion-based system that can generate drummer motion from audio with centimeter-level stick accuracy while keeping the body movement natural. The model separates skeletal motion …

Research Disney Research Studios · 1 month ago
RGBX-Next: Towards Realistic Generative Rendering from G-Buffers

RGBX-Next turns G-buffers into a controllable input for generative rendering, aiming to bridge the gap between diffusion models and traditional 3D pipelines. The system can synthesize realistic images, video, and …

Research arXiv cs.GR · 1 month ago
RealMat: Realistic Materials with Diffusion and Reinforcement Learning

RealMat combines Stable Diffusion XL with reinforcement learning to generate more believable material maps for 3D authoring. The pipeline starts from synthetic 2×2 material grids, then pushes the model toward …

Research arXiv cs.GR · 1 month ago
NuclearDiffusion: Text-to-Image Foundation Models for Learning Nuclear Energy Concepts

A new study shows that fine-tuning open-source diffusion models on 1,000 captioned nuclear-energy images can materially improve text-to-image accuracy for specialized technical prompts. SDXL benefited the most, while SD-v3.5-Medium saw …

Research arXiv cs.GR · 1 month, 1 week ago
Fourier-Latent Diffusion for Constrained Generation of Triply Periodic Minimal Surfaces

A new diffusion pipeline can generate triply periodic minimal surfaces with far tighter geometric control than earlier TPMS tools. The system combines a 18K-surface dataset, a Fourier latent space that …

Research arXiv cs.GR · 1 month, 2 weeks ago
S-Avatar: Diffusion-Guided Gaussian Head Avatars from a Single Image

A new single-image avatar pipeline can generate photorealistic 3D head models with real-time expression control. S-Avatar combines diffusion-guided 3D Gaussian splatting with FLAME alignment to improve view consistency under motion …

Research arXiv cs.GR · 1 month, 2 weeks ago
Two2Four: Generative Quadruped Puppeteering from Human Motion

Disney Research Studios has unveiled Two2Four, a human-to-quadruped puppeteering system that turns ordinary human motion into plausible animal animation. The pipeline uses a two-stage diffusion model trained on quadruped motion …

Research Disney Research Studios · 1 month, 2 weeks ago
Two2Four: Generative Quadruped Puppeteering from Human Motion

Two2Four is a new human-to-quadruped puppeteering system that turns ordinary human motion into plausible animal animation. Built on a two-stage diffusion model trained only on quadruped motion, it aims to …

Research arXiv cs.GR · 1 month, 2 weeks ago
MMOE: Modernizing Diffusion Transformers with Efficient Expert Design

MMOE brings sparse-expert routing to diffusion transformers with a stronger focus on efficiency, not just raw parameter growth. Trained on a single 8×H100 node for 400k steps, it hit lower …

Research arXiv cs.GR · 1 month, 3 weeks ago
Feature-Guided Diffusion for Non-Differentiable Inverse Rendering

A new black-box inverse rendering pipeline, Feature-Informed Diffusion Evolution, skips gradients and hand-tuned initialization entirely. It uses ViT features to steer a diffusion model, then tightens candidates with CMA-ES. The …

Research arXiv cs.GR · 1 month, 4 weeks ago
Neural Motion Blending Across Arbitrary Character Topologies

A new neural motion-blending system can interpolate animation across characters with different skeleton topologies, not just near-identical rigs. It pairs a semantic motion encoder with a diffusion decoder to reconstruct …

Research arXiv cs.GR · 2 months ago
Page 1 of 5 Next