Diffusion News
The latest Diffusion coverage curated for game developers.
Researchers have extended grid-free Monte Carlo diffusion solvers to transient heat equations, removing the need for volumetric meshing and time stepping. The method targets time-dependent diffusion with initial conditions plus …
A diffusion pipeline is being used to synthesize sand-boil inspection imagery for earthen levees when real defect annotations are scarce. The system fine-tunes Stable Diffusion XL with DreamBooth, then steers …
PixGS pushes 3D Gaussian splat generation into a single-stage, pixel-space diffusion pipeline, aiming to cut the artifacts and error buildup common in latent-based workflows. The method directly denoises splat attributes …
For graphics programmers working on skin, wax, or other thin translucent materials, this paper puts numbers on a long-standing dipole problem: the geometry error falls like about 1.99e^-2τ for diffuse …
HiMat tackles a real bottleneck for material authoring: generating 4K SVBRDFs without blowing up memory or losing map alignment. It uses a high-compression latent space plus a diffusion transformer with …
MeshFlow replaces slow auto-regressive mesh tokenization with a continuous latent VAE plus a rectified-flow transformer, and it claims 18x faster generation than the fastest AR baseline. For graphics and tools …
AvatarMix tackles a real pain point in avatar outfit transfer: keeping identity and clothing intact without seam or intersection artifacts. It composes head and body from two high-fidelity Gaussian avatars, …
A new paper tackles a hard animation problem: turning a 2D sketch plus text into usable 3D motion without paired motion data. The interesting bit for animation and tech art …
A new diffusion framework can learn fine-grained object parts from ordinary 2D images, then let you swap, interpolate, and recombine those parts without 3D scans or camera poses. For teams …
A 5D point-cloud garment representation tries to bridge the gap between fast, high-level generation and the messy reality of editing both sewing patterns and draped 3D cloth. For teams working …
A new paper argues that subject-driven image generation works better when the diffusion model is conditioned on a multimodal LLM that sees text and reference images together. The authors also …
For level and tools teams, the interesting bit is the move from loose bounding proxies to explicit voxel occupancy for indoor scene layout. VoxScene claims that this cuts down collisions …
A diffusion-based material transfer paper is trying to cut out two common pain points for graphics teams: text prompts and extra reference networks. DealMaTe uses depth, normal, and lighting inputs …
SURF tackles a practical bottleneck for teams experimenting with video generation: 720p inference can take 50+ minutes on Wan2.1. The paper’s two-stage approach keeps more of the base model’s layout, …
The introduction of DiffusionLight-Turbo offers graphics programmers a significant leap in light probe estimation, reducing processing time from 30 minutes to just 30 seconds. This efficiency gain is achieved with …
DiffusionBrowser introduces a game-changing interactive preview system for video diffusion models, allowing developers to generate previews in real-time. This model-agnostic framework enhances control during the denoising process, making it particularly …
The introduction of Phase-Preserving Diffusion (?C6-PD) marks a significant advancement in structure-aligned generation, particularly beneficial for graphics programmers and artists. By preserving input phase while randomizing magnitude, this method enhances …