Iliyan Georgiev
I am a research scientist at Adobe. My recent work explores generative imaging models, including decomposition, synthesis, and editing with material- and lighting-aware control. I also work on neural and classical 3D representations—such as 3D Gaussian splatting, radiance fields, and geometry compression—as well as appearance modeling, understanding, and selection. My research further spans differentiable and inverse rendering, and I have a long-standing interest in Monte Carlo light transport simulation, importance sampling, and physically based rendering, with applications in film production and visual content creation.
Bio. I obtained my master's and PhD degrees from Saarland University with fellowships from the Max-Planck Institute and support from the Intel Visual Computing Instute. During my PhD, I did internships at Disney Research and Weta Digital, and a consultancy at Chaos Group (contributing to the V-Ray renderer). After graduating, I joined the Arnold renderer team, initially at Solid Angle and then Autodesk. My PhD thesis received the Eurographics PhD Thesis Award, and my contributions to the Arnold renderer were recognized by an Engineering Emmy Award. I am currently at Adobe Research in London.
Collaboration. I am always looking to connect with strong PhD students to collaborate with and/or to host as summer interns at Adobe Research. If you believe your research interests align with mine, do not hesitate to get in touch.
Publications
2026
CVPR 2026
An end-to-end framework for intrinsic-aware video editing that unifies video inverse rendering, photorealistic synthesis, and keyframe-based editing conditioned on intrinsic scene properties like albedo, normals, and materials.
CVPR 2026
A sorting-free differentiable stochastic formulation for ray-traced 3D Gaussian splatting, enabling efficient reconstruction and rendering with accurate ray-traced shadows and shading.
CVPR 2026
A joint latent space that unifies multiple lighting representations (environment maps, irradiance, spherical harmonics, text) for cross-modal retrieval, generation, and lighting control.
3DV 2026
3D-consistent material selection across representations (NeRFs, 3D Gaussians, meshes) based on fine-tuned SAM2, enabling click-to-selection in about 2 seconds.
WACV 2026
A diffusion distillation technique derived from mean-shift mode seeking that serves as a drop-in replacement for score distillation sampling (SDS), with provably better mode alignment and improved convergence.
2025
SIGGRAPH Asia 2025 (journal)
Fine-grained material selection in images at two granularity levels (texture and subtexture), using a multi-resolution ViT feature processing strategy robust to lighting and reflectance variations.
SIGGRAPH Asia 2025 (journal), best paper award!
Automatic boundary sampling for discontinuities in differentiable shaders via a program transformation that converts discontinuous functions into piecewise-constant ones, removing the need for specialized boundary-sampling routines.
ICCV 2025
The first scene-level 3D variational autoencoder that encodes many Gaussian primitives into a low-dimensional latent embedding, enabling image-to-3DGS and text-to-3DGS generation.
SIGGRAPH 2025 (journal)
Precise editing via local, pixel-level manipulations to the intrinsic channels of an image; a diffusion model recombines the channels into an edited result, handling global illumination effects and preserving input identity.
EGSR 2025
Unbiased stochastic ray tracing of scenes with many transparent primitives (e.g. 3D Gaussians), without explicit per-ray sorting, suitable for both low- and and high-end GPUs.
ICPRAM 2025 (short paper)
A framework for combining multiple importance sampling distributions tailored to specific parameter gradients, yielding superior gradient estimates and faster training convergence.
ICPRAM 2025, best student paper award!
An on-the-fly importance sampling algorithm for stochastic gradient optimization that prioritizes influential data points, improving accuracy and enabling online data pruning.
2024
SIGGRAPH Asia 2024 (conference)
A normalizing-flow-based method for approximate product importance sampling of environment lighting with material and cosine terms, composing a learned head warp with a discretized emitter tail warp.
arXiv, 2024
A physics-based inverse rendering framework for glossy objects that jointly optimizes shape, reflectance, and a parallax-aware non-distant environment lighting representation.
SIGGRAPH 2024 (conference)
Diffusion models for decomposing images into intrinsic channels (albedo, roughness, metallicity, lighting) and synthesizing realistic images from full or partial intrinsic channels.
SIGGRAPH 2024 (conference)
A neural compression architecture for answering arbitrary ray queries in 3D, achieving over an order of magnitude compression while integrating seamlessly into standard ray-tracing pipelines.
EGSR 2024 (journal), best paper award!
A formulation of the image difference between scene edits as a correlated light-transport integral, enabling efficient incremental re-rendering by focusing on paths affected by scene changes.
CVPR 2024 (highlight)
A feature-grid-based angular encoding for NeRF that improves modeling of high-frequency view-dependent effects including interreflections, enabling real-time inference.
2023
SIGGRAPH Asia 2023 (journal)
A neural-mesh-based 2D representation that preserves sharp discontinuities along specified curves at arbitrary resolution, applicable to image compression, diffusion curves, and PDE solutions.
SIGGRAPH Asia 2023 (conference)
paper
supplemental document
poster
supplemental results
slides
supplemental video
fast-forward video
talk video
Extension of perceptual error optimization to the spatio-temporal domain for Monte Carlo animation rendering, distributing pixel error as blue noise across both space and time.
SIGGRAPH Asia 2023 (conference)
Efficient displacement-map ray tracing via inverse mapping and on-the-fly surface-bound computation, achieving over 10x speedup without pre-tessellation.
SIGGRAPH 2023 (conference)
Evaluation and comparison of pdf-learning approaches (analytic-lobe mixtures, normalizing flows, histogram prediction) for importance sampling spatially varying neural materials.
2022
SIGGRAPH Asia 2022 (journal)
A multi-class point optimization formulation based on continuous Wasserstein barycenters, scaling to thousands of classes for applications in stippling, object placement, and blue-noise error distribution.
SIGGRAPH Asia 2022 (conference)
A generalization of multiple importance sampling to combine samples from marginal probability distributions, enabling new path sampling techniques such as multi-vertex path-space filtering.
SIGGRAPH 2022 (journal)
A general framework for transforming biased estimators into unbiased and consistent ones, applied to transmittance estimation, photon mapping, and finite differences.
SIGGRAPH 2022 (journal)
A method to improve MIS efficiency by cheaply estimating the efficiencies of various technique and sample-count combinations, automatically adapting bidirectional rendering to scene complexity.
ACM Transactions on Graphics, 2022 (presented at SIGGRAPH 2022)
paper
supplemental document
paper (ACM version)
supplemental results
supplemental video
fast-forward video
talk video
talk video subtitles
arXiv preprint
A perception-oriented framework that optimizes the error distribution of Monte Carlo renderings to produce visually pleasing blue-noise characteristics in image space.
EGSR 2022
A single-pass algorithm for stratified importance resampling that orders candidates along a space-filling curve, combining the benefits of stratification with resampling-based rendering.
2021
SIGGRAPH 2021
A taxonomy and analysis of Monte Carlo estimators for differential light transport, exploring the large design space of choices for sampling, differentiation, and MIS in differentiable rendering.
EGSR 2021
Complementary importance sampling techniques for volumetric single scattering that account for transmittance, phase-function anisotropy, foreshortening, and distance fall-off from emitters.
EUROGRAPHICS 2021
Incorporating path-prefix correlation knowledge into the MIS balance heuristic to improve technique combination in bidirectional and path-reuse rendering algorithms.
2020
SIGGRAPH 2020
A generalization of multiple importance sampling to uncountably infinite sets of sampling techniques, with a provably optimal balance heuristic and a practical stochastic estimator.
SIGGRAPH 2020
A simple and general sampling strategy for specular light paths that unifies caustic and glint rendering, finding specular subpaths connecting given endpoints via manifold walks.
SIGGRAPH 2020 (course)
A SIGGRAPH course recapping the research contributions of Jaroslav Křivánek toward finding a single robust and efficient light transport simulation algorithm.
SIGGRAPH 2020 (course)
A SIGGRAPH course on physically based shading, presenting further research and practical advice from film and game production.
2019
SIGGRAPH Asia 2019
Enhancing the MIS balance heuristic with variance estimates of individual sampling techniques to improve efficiency when combining stratified and unstratified techniques.
SIGGRAPH Asia 2019
Novel integral formulations of volumetric transmittance that reveal existing null-scattering estimators as direct Monte Carlo estimators and enable new estimator types with complementary tradeoffs.
SIGGRAPH 2019
A path integral formulation of null-scattering that reveals sampling PDFs of null-collision methods, enabling their combination with other techniques via multiple importance sampling.
SIGGRAPH 2019 (course)
A SIGGRAPH course on the algorithms and characteristics of different classes of uniformly distributed point sets and sequences for Monte Carlo and quasi-Monte Carlo rendering.
Autodesk white paper, 2019
Specification of an uber surface shader for production rendering, capable of modeling the vast majority of materials used in visual effects and feature animation.
2018
ACM Transactions on Graphics, 2018 (presented at SIGGRAPH 2018)
Description of the architecture of the Arnold production path tracer, focusing on design choices for handling massive geometry, detailed shading, and billions of incoherent rays.
SIGGRAPH 2018 (course)
A SIGGRAPH course on Monte Carlo methods for volume rendering, covering free-path sampling, transmittance estimation, and light-path construction in participating media.
EUROGRAPHICS 2018 (state-of-the-art report)
A comprehensive survey of Monte Carlo methods for light transport in participating media, covering free-path sampling, transmittance estimation, and path-construction algorithms.
2017
EGSR 2017
Area-preserving mappings from the unit square to spherical ellipses for stratified solid-angle sampling of direct illumination from disk-shaped light sources.
EUROGRAPHICS 2017
A stochastic particle model for unbiased rendering of heterogeneous media that does not require a-priori knowledge of the maximum medium density.
2016
SIGGRAPH 2016 (talk)
Blue-noise dither masks that correlate pixel estimates across the image to minimize low-frequency error content, producing more faithful images at low sampling rates.
2015
Computer Graphics Forum, 2015 (presented at EGSR 2016)
Blue-noise sampling patterns that retain their spectral properties under projection onto lower-dimensional subspaces, improving Monte Carlo integration and image reconstruction.
SIGGRAPH 2015 (course)
A SIGGRAPH course on how path tracing is revolutionizing movie production, covering architectures, workflows, and examples from recent films.
2014
SIGGRAPH 2014
paper
supplemental document
poster
slides (pptx)
slides (pdf)
supplemental results
code (SmallUPBP)
fxguide article
A unified rendering algorithm combining volumetric point and beam density estimators with Monte Carlo path sampling via extended multiple importance sampling.
SIGGRAPH 2014 (course)
A SIGGRAPH course reviewing the path integral formulation of light transport and its applications, with practical guidance on advanced rendering in VFX and architectural visualization.
EUROGRAPHICS 2014 (tutorial)
A Eurographics tutorial on path integral methods for light transport simulation, covering bidirectional techniques, MCMC methods, and practical applications.
2013
SIGGRAPH Asia 2013
Joint importance sampling of sequences of vertices in participating media, constructing single- and double-scattering subpaths that account for all scattering and geometry terms simultaneously.
SIGGRAPH Asia 2013 (course)
A SIGGRAPH Asia course on the latest advances in photon density estimation, covering progressive methods, bandwidth adaptation, beam-based estimators, and generalized formulations.
SIGGRAPH 2013 (course)
A SIGGRAPH course reviewing the path integral formulation of light transport, Markov chain Monte Carlo methods, and practical applications in VFX.
EUROGRAPHICS 2013
A global visibility caching algorithm that adaptively quantizes and caches visibility queries between arbitrary scene points, reducing shadow ray counts to a fraction of the original.
2012
SIGGRAPH Asia 2012
paper
supplemental document
slides
supplemental results
supplemental images
code (SmallVCM)
technical report
A reformulation of photon mapping as a bidirectional path sampling technique, enabling seamless combination with bidirectional path tracing via multiple importance sampling.
Technical report, November 2012 (revision 2, July 2013)
A technical report on the efficient implementation of MIS weighting for vertex connection and merging, using partial weight sums accumulated at path vertices.
Graphics Interface 2012
A unification of ray tracing and rasterization via 3D edge functions, enabling non-planar viewports for rasterization and consistent anti-aliasing for ray tracing.
2011
SIGGRAPH Asia 2011 (technical sketch)
Early formulation of vertex merging as a bidirectional path sampling technique that brings the advantages of photon mapping to the path integral framework.
2010
International Conference on Information Visualisation (IV) 2010
Integration of real-time ray tracing into molecular visualization for interactive high-quality rendering with shadows, reflections, and correct light attenuation.