Skip to main content
GameDev.net gamedev.net

Generative Models News

The latest Generative Models coverage curated for game developers.

Balancing Image Compression and Generation with Bootstrapped Tokenization

SelfBootTok splits image tokens into global and local groups, cutting redundant detail from the generator’s job. The paper reports roughly 40% less generator computation while improving reconstruction and generation quality. …

Research arXiv cs.GR · 3 months, 3 weeks ago
Synthetic Sociality: How Generative Models Privatize the Social Fabric

The emergence of generative models is reshaping social interactions, potentially impacting game developers by altering how social dynamics are integrated into games. This article critiques how these models commodify sociality, …

Research arXiv cs.GR · 4 months, 2 weeks ago
A Few-Step Generative Model on Cumulative Flow Maps

A new generative modeling framework leveraging cumulative flow maps offers game developers a streamlined approach to long-range transport in probability space. This innovation allows for efficient one-step and few-step generation, …

Research arXiv cs.GR · 4 months, 3 weeks ago
BulletGen: Improving 4D Reconstruction with Bullet-Time Generation

BulletGen introduces a novel approach to 4D reconstruction by leveraging generative models to enhance dynamic scene representation. This technique addresses common challenges in monocular video processing, making it particularly relevant …

Research arXiv cs.GR · 5 months, 3 weeks ago
Gen-C: Populating Virtual Worlds with Generative Crowds

Gen-C introduces a groundbreaking framework for simulating crowds in virtual environments, leveraging Large Language Models to generate synthetic datasets. This innovation allows developers to create more realistic agent interactions without …

Research arXiv cs.GR · 6 months ago
Fillerbuster: Unified Generative Scene Completion Model for Casual Captures

Fillerbuster introduces a groundbreaking generative model that enhances 3D scene completion by utilizing a multi-view latent diffusion transformer. This innovation is particularly relevant for graphics programmers and artists, as it …

Research arXiv cs.GR · 6 months, 1 week ago
Improved 3D Scene Stylization via Text-Guided Generative Image Editing with Region-Based Control

Recent advancements in text-driven 3D scene editing are reshaping how artists and designers approach stylization. The introduction of region-controlled style transfer enhances both quality and consistency, addressing long-standing challenges in …

Research arXiv cs.GR · 6 months, 3 weeks ago
Untwisting RoPE: Frequency Control for Shared Attention in DiTs

The new insights into Rotary Positional Embeddings (RoPE) could significantly impact how artists and designers approach style transfer in generative models. By modulating frequency bands, developers can achieve a better …

Research arXiv cs.GR · 7 months, 3 weeks ago
ZipLoRA: Any Subject in Any Style by Effectively Merging LoRAs

ZipLoRA introduces a novel approach to merging low-rank adaptations (LoRAs) for generative models, allowing developers to generate any subject in any style without sacrificing fidelity. This method is particularly relevant …

Research arXiv cs.GR · 7 months, 3 weeks ago
EditYourself: Audio-Driven Generation and Manipulation of Talking Head Videos with Diffusion Transformers

EditYourself introduces a groundbreaking method for editing talking head videos, crucial for video post-production. This tool allows developers to modify spoken content while ensuring lip sync and motion coherence, a …

Research arXiv cs.GR · 8 months ago
PromptVFX: Text-Driven Fields for Open-World 3D Gaussian Animation

PromptVFX introduces a groundbreaking text-driven framework for 3D Gaussian animation, significantly reducing the complexity of creating visual effects. This innovation allows both novice and expert users to generate compelling animations …

Research arXiv cs.GR · 8 months ago
LouvreSAE: Sparse Autoencoders for Interpretable and Controllable Style Transfer

LouvreSAE introduces a novel approach to artistic style transfer that significantly reduces computational overhead. By utilizing a Sparse Autoencoder, developers can achieve style transfer without the need for model fine-tuning …

Research arXiv cs.GR · 9 months, 1 week ago
What matters for Representation Alignment: Global Information or Spatial Structure?

Recent findings challenge the conventional wisdom that global semantic performance is key for generative model training. Instead, spatial structure has emerged as the driving factor for generation performance, which could …

Research arXiv cs.GR · 9 months, 2 weeks ago