Meta tags:
description=Ivan Skorokhodov — research scientist working on generative models for images, video and 3D.;
author=Ivan Skorokhodov;

Headings (most frequently used words):

of, video, and, improving, the, 3d, for, 2021, 2025, in, diffusion, 2024, to, generation, 2022, continuous, image, ivan, skorokhodov, selected, research, programming, projects, alphaflow, understanding, meanflow, modelsiclr, 2026, diffusability, autoencodersicml, ac3d, analyzing, camera, control, transformerscvpr, hierarchical, patch, models, high, resolution, generationcvpr, snap, scaled, spatiotemporal, transformers, text, synthesiscvpr, on, imageneticlr, 2023, oral, epigraf, rethinking, training, gansneurips, stylegan, generator, with, price, quality, perks, stylegan2cvpr, aligning, latent, spaces, connect, unconnectableiccv, adversarial, imagescvpr, class, normalization, continual, generalized, zero, shot, learningiclr, loss, patterns, neural, networksbeyond, first, order, methods, ml, workshop, neurips, 2019, dl, reasonerrust, firelabpythonpytorch, omniplan, web, appjavascriptreact, rtrsrust, non, uniform, interpolationcuda,

Text of the page (most frequently used words):
the (73), and (60), that (30), for (21), video (21), this (18), model (18), models (15), with (14), more (14), ivan (14), skorokhodov (14), #quality (13), based (12), arxiv (12), read (12), #generation (11), code (11), our (11), generator (11), camera (11), diffusion (11), latent (10), training (10), first (9), from (9), into (8), state (8), art (8), page (8), 256 (8), image (8), sergey (8), tulyakov (8), using (7), normalization (7), datasets (7), work (7), but (6), which (6), research (6), times (6), two (6), show (6), high (6), aliaksandr (6), siarohin (6), learning (5), very (5), consistency (5), class (5), mohamed (5), elhoseiny (5), 2021 (5), scale (5), cvpr (5), motion (5), trained (5), resolution (5), 2024 (5), willi (5), menapace (5), built (4), interpolation (4), one (4), like (4), these (4), non (4), was (4), have (4), over (4), framework (4), across (4), loss (4), are (4), showed (4), faster (4), images (4), inr (4), continuous (4), codes (4), single (4), them (4), synthesis (4), 2022 (4), geometry (4), patch (4), imagenet (4), text (4), control (4), 2025 (4), existing (3), techniques (3), point (3), positions (3), uniform (3), grid (3), often (3), works (3), rust (3), years (3), pytorch (3), similar (3), different (3), neural (3), many (3), zero (3), shot (3), achieves (3), trains (3), than (3), approach (3), iclr (3), proposed (3), generative (3), end (3), stylegan2 (3), just (3), allows (3), generate (3), generating (3), frames (3), spaces (3), then (3), layers (3), simple (3), demonstrate (3), improve (3), results (3), low (3), strategy (3), peter (3), wonka (3), scales (3), videos (3), snap (3), propose (3), improving (3), 2026 (2), nearest (2), etc (2), form (2), cuda (2), rtrs (2), both (2), used (2), any (2), web (2), their (2), omniplan (2), past (2), few (2), deep (2), projects (2), without (2), reasoner (2), concept (2), all (2), most (2), general (2), 2019 (2), ideas (2), landscapes (2), networks (2), visual (2), patterns (2), turned (2), out (2), can (2), found (2), inside (2), contain (2), methods (2), neurips (2), 2020 (2), signal (2), there (2), performance (2), sota (2), also (2), novel (2), continual (2), scaled (2), gan (2), generates (2), pixel (2), parameters (2), multi (2), lsun (2), 1024 (2), properties (2), coordinate (2), embeddings (2), representation (2), positioning (2), plane (2), when (2), instead (2), connect (2), large (2), build (2), time (2), components (2), temporal (2), information (2), finally (2), sampling (2), scheme (2), clip (2), ucf (2), 101 (2), manipulation (2), nerf (2), gans (2), location (2), discriminator (2), patches (2), convergence (2), resulting (2), epigraf (2), efficient (2), four (2), 512 (2), resolutions (2), fidelity (2), same (2), scenes (2), 3dgp (2), new (2), regularization (2), knowledge (2), pre (2), explore (2), recent (2), terms (2), jian (2), ren (2), 2023 (2), redundant (2), significant (2), architecture (2), higher (2), transformers (2), fine (2), foundational (2), recently (2), analyze (2), frequency (2), only (2), fid (2), meanflow (2), step (2), objective (2), flow (2), matching (2), alphaflow (2), universome (2), yandex (2), last, updated, jun, claude, hugo, while, neighbour, bilinear, lanczos, hamming, assume, known, not, always, case, moreover, would, backpropagate, through, project, implemented, kernel, gaussian, mixture, small, ray, tracing, rasterization, engine, written, meshes, traditional, quadrics, has, some, cool, features, distributed, bvhs, arcball, rotations, extensively, previous, job, didn, interface, annoyed, everyone, official, api, react, javascript, app, been, building, running, experiments, lightning, hydra, proper, documentation, testing, python, firelab, tableau, alcq, description, logic, supporting, abox, checking, subsumption, tbox, implements, eight, expansion, rules, conjunction, disjunction, universal, existential, restrictions, least, cardinality, choice, inclusions, programming, searched, due, extreme, overparametrization, pattern, surface, indicates, complex, irregularities, mode, connectivity, mikhail, burtsev, beyond, order, workshop, paper, dived, zsl, how, influence, variance, deeper, need, other, procedures, developed, batch, applied, dimension, mlp, x50, 200, current, formulated, problem, tested, setup, attribute, cosine, similarity, generalized, function, takes, coordinates, input, predicts, value, way, hypernetwork, such, real, world, factorized, multiplicative, modulation, fmm, inrs, achieved, decent, churches, bedrooms, ffhq, lot, interesting, decoders, day, 1x1, convolutions, neighbor, upsampling, implicit, savva, ignatyev, adversarial, idea, means, each, sampled, associated, position, computes, color, neighboring, global, infinite, size, infinitely, unrelated, arbitrarily, panorama, grigory, sotnikov, iccv, aligning, unconnectable, autoregressive, rethink, fundamental, redesign, structuring, positional, drop, usage, expensive, conv3d, aggregate, concatenation, could, sparse, per, modifications, greatly, efficiency, achieve, strong, faceforensics, sky, timelapse, rainbow, jelly, mead, projecting, its, space, frame, editing, acyclic, stylegan, price, perks, several, months, appeared, speed, separate, decoder, upsample, produced, solution, comes, cost, breaks, view, learns, possible, obtain, simply, wise, revisit, optimization, ways, designing, aware, proportions, spatial, modifying, annealed, beta, distribution, stabilize, accelerate, named, pure, test, introduced, obtains, approx, upsampler, counterparts, yiqun, wang, rethinking, generators, typically, designed, well, curated, category, where, objects, approximately, orientation, makes, inapplicable, diverse, wild, alignable, rendered, arbitrary, poses, develop, generic, priors, assumptions, about, data, challenging, three, off, shelf, depth, estimator, guide, flexible, learnable, paired, distillation, transfer, external, feature, extractor, outperforms, texture, yinghao, hsin, ying, lee, oral, contemporary, remarkable, versatility, swayed, advantages, community, repurposes, since, content, highly, argue, naively, bringing, advances, domain, reduces, impairs, scalability, systematically, addresses, challenges, extend, edm, take, account, spatially, temporally, pixels, naturally, support, second, net, workhorse, behind, poorly, requiring, computational, overhead, hence, transformer, nets, inference, efficiently, train, billions, reach, number, benchmarks, substantially, complexity, user, studies, favored, margin, ekaterina, deyneka, tsai, shien, chen, anil, kag, yuwei, fang, aleksei, stoliar, elisa, ricci, spatiotemporal, dominant, paradigms, cascaded, structured, hierarchically, denoise, random, various, coarse, manner, conditional, promising, scaling, 288, hierarchical, integrated, imprecise, suffers, principles, perspective, obtaining, insights, enable, precise, compromising, inherently, motivates, apply, pose, conditioning, beginning, trajectory, next, implicitly, builds, coherent, representations, initial, providing, final, collect, dataset, still, cameras, moving, mitigate, static, bias, sherwin, bahmani, guocheng, qian, andrea, tagliasacchi, david, lindell, ac3d, analyzing, unwanted, degrade, introduces, equivariance, aligns, rgb, frequencies, minimal, tuning, delivers, improvements, reducing, fvd, kinetics, 700, making, sharath, girish, benran, yanyu, rameen, abdal, icml, diffusability, autoencoders, scratch, showing, decomposed, vanilla, generalization, decomposable, name, supports, convenient, curriculum, schedule, leads, regime, dit, 256x256, design, better, scores, huijie, zhang, michael, vasilkovsky, qing, understanding, selected, google, scholar, isskoro, iskorokhodov, gmail, com, before, spent, language, machine, translation, landscape, analysis, earlier, software, engineer, team, market, mipt, foundation, robotics, previously, scientist, working, mostly, obtained, phd, advised, kaust, rhoda,

Text of the page (random words):
ivan skorokhodov ivan skorokhodov i work on foundation models for robotics at rhoda ai previously i was a research scientist at snap research working on generative models mostly video diffusion i obtained my phd from kaust in 2023 advised by peter wonka and mohamed elhoseiny before that i spent two years on research at mipt first on language models and machine translation then on loss landscape analysis earlier i was a software engineer at yandex for 1 5 years on the yandex market team iskorokhodov gmail com universome universome isskoro google scholar cv selected research alphaflow understanding and improving meanflow models iclr 2026 huijie zhang aliaksandr siarohin willi menapace michael vasilkovsky sergey tulyakov qing qu ivan skorokhodov we explore the recently proposed meanflow framework for few step from scratch trained diffusion models we first analyze the training objective showing that it can be decomposed into vanilla flow matching and consistency terms we then propose a simple generalization of this decomposable objective which we name alphaflow it supports a convenient curriculum schedule that leads it from the flow matching into the consistency training regime we show that for the same dit architecture on imagenet 256x256 this design achieves 10 20 better fid scores for one two step generation read more 2025 arxiv code improving the diffusability of autoencoders icml 2025 ivan skorokhodov sharath girish benran hu willi menapace yanyu li rameen abdal sergey tulyakov aliaksandr siarohin latent diffusion models are the state of the art for image and video generation but their latent spaces often contain unwanted high frequency components that degrade quality our work introduces scale equivariance a simple regularization strategy that aligns latent and rgb spaces across frequencies with minimal fine tuning this approach delivers significant improvements reducing fid by 19 on imagenet 1k and fvd by over 44 on kinetics 700 making generative models more efficient and higher quality read more 2025 arxiv code ac3d analyzing and improving 3d camera control in video diffusion transformers cvpr 2025 sherwin bahmani ivan skorokhodov guocheng qian aliaksandr siarohin willi menapace andrea tagliasacchi david b lindell sergey tulyakov many works have recently integrated 3d camera control into foundational text to video models but the resulting camera control is often imprecise and video generation quality suffers in this work we analyze camera motion from a first principles perspective obtaining insights that enable precise 3d camera manipulation without compromising synthesis quality first we show that camera motion is an inherently low frequency signal which motivates us to apply pose conditioning only at the beginning of the diffusion trajectory next we found that a pre trained diffusion model implicitly builds coherent camera representations in the initial layers by providing the camera information to only these layers we improve the training convergence and final quality finally we collect a dataset with still cameras but moving scenes to mitigate the static bias of existing camera control datasets read more 2024 page arxiv code hierarchical patch diffusion models for high resolution video generation cvpr 2024 ivan skorokhodov willi menapace aliaksandr siarohin sergey tulyakov the two dominant diffusion paradigms are latent diffusion and cascaded diffusion we propose a novel patch based diffusion model that allows end to end training for high resolutions the model is structured hierarchically and trained to denoise random patches of various scales in a coarse to fine manner it achieves state of the art results on class conditional ucf 101 64 times 256 2 generation and promising scaling performance as a foundational 64 times 288 times 512 text to video model read more 2024 page arxiv snap video scaled spatiotemporal transformers for text to video synthesis cvpr 2024 willi menapace aliaksandr siarohin ivan skorokhodov ekaterina deyneka tsai shien chen anil kag yuwei fang aleksei stoliar elisa ricci jian ren sergey tulyakov contemporary models for generating images show remarkable quality and versatility swayed by these advantages the research community repurposes them to generate videos since video content is highly redundant we argue that naively bringing advances of image models to the video generation domain reduces motion fidelity visual quality and impairs scalability in this work we build snap video a video first model that systematically addresses these challenges to do that we first extend the edm framework to take into account spatially and temporally redundant pixels and naturally support video generation second we show that a u net a workhorse behind image generation scales poorly when generating videos requiring significant computational overhead hence we propose a new transformer based architecture that trains 3 31 times faster than u nets and is 4 5 faster at inference this allows us to efficiently train a text to video model with billions of parameters for the first time reach state of the art results on a number of benchmarks and generate videos with substantially higher quality temporal consistency and motion complexity the user studies showed that our model was favored by a large margin over the most recent methods read more 2024 page arxiv 3d generation on imagenet iclr 2023 oral ivan skorokhodov aliaksandr siarohin yinghao xu jian ren hsin ying lee peter wonka sergey tulyakov existing 3d from 2d generators are typically designed for well curated single category datasets where all the objects have approximately the same scale 3d location and orientation this makes them inapplicable to diverse in the wild datasets of non alignable scenes rendered from arbitrary camera poses in this work we develop a 3d generator with generic priors 3dgp a 3d synthesis framework with more general assumptions about the training data and show that it scales to very challenging datasets like imagenet our model is based on three new ideas 1 using an off the shelf depth estimator to guide the learning of 3d geometry 2 a flexible learnable camera generator paired with a regularization strategy and 3 knowledge distillation into the discriminator to transfer external knowledge from a pre trained feature extractor we explore our model on four datasets and demonstrate that 3dgp outperforms the recent state of the art in terms of both texture and geometry quality read more 2022 page arxiv code epigraf rethinking training of 3d gans neurips 2022 ivan skorokhodov sergey tulyakov yiqun wang peter wonka in the past several months there appeared 10 works that speed up nerf based gans by training a separate 2d decoder to upsample a low resolution 3d representation produced from the nerf generator this solution comes at a cost it breaks multi view consistency and learns the geometry at a low resolution instead we show that it is possible to obtain a high resolution 3d generator with sota image quality by simply training the model patch wise we revisit and improve this optimization scheme in two ways 1 by designing a location and scale aware discriminator to work on patches of different proportions and spatial positions and 2 modifying the patch sampling strategy based on an annealed beta distribution to stabilize training and accelerate the convergence the resulting model named epigraf is an efficient high resolution pure 3d generator and we test it on four datasets two introduced in this work at 256 2 and 512 2 resolutions it obtains state of the art image quality high fidelity geometry and trains approx 2 5 times faster than the upsampler based counterparts read more 2022 page arxiv code stylegan v a continuous video generator with the price image quality and perks of stylegan2 cvpr 2022 ivan skorokhodov sergey tulyakov mohamed elhoseiny we build a non autoregressive video generator which is continuous in time it is based on stylegan2 and we rethink fundamental components of video synthesis models first we redesign the motion codes to be continuous by structuring them as acyclic positional embeddings then we drop the usage of expensive conv3d layers and aggregate the temporal information across frames by simple concatenation finally we demonstrate that a state of the art video generator could be trained with a very sparse sampling scheme using just 2 3 frames per clip our modifications greatly improve the training efficiency of our model and we achieve strong state of the art results on faceforensics 256 2 sky timelapse 256 2 ucf 101 256 2 rainbow jelly 256 2 and mead 1024 2 we also demonstrate the video manipulation properties of our generator like projecting a video into its latent space using just a single frame and clip based editing read more 2021 page arxiv code aligning latent and image spaces to connect the unconnectable iccv 2021 ivan skorokhodov grigory sotnikov mohamed elhoseiny we proposed the idea of positioning a gan s latent codes on the coordinate plane this means that each latent code when sampled is associated with an x y position on the 2d image plane and our generator computes the color of a pixel from an interpolation of the neighboring latent codes instead of just a single global one this allows us 1 to generate images of infinite size by generating infinitely many latent codes and positioning them on the grid and 2 to connect unrelated frames into a single arbitrarily large panorama read more 2021 page arxiv code adversarial generation of continuous images cvpr 2021 ivan skorokhodov savva ignatyev mohamed elhoseiny we built a gan model that generates images in the implicit neural representation inr form an inr is a function f c which takes coordinates c x y as input and predicts a pixel value v r g b in this way our generator is a hypernetwork that generates parameters for f c we proposed two techniques to scale such a model to real world datasets factorized multiplicative modulation fmm and multi scale inrs we achieved decent for inr based models generative quality on lsun churches 256 2 lsun bedrooms 256 2 and ffhq 1024 2 and showed a lot of interesting properties of inr based decoders at the end of the day our approach turned out to be very similar to stylegan2 with 1x1 convolutions coordinate embeddings and nearest neighbor upsampling read more 2020 page arxiv code class normalization for continual generalized zero shot learning iclr 2021 ivan skorokhodov mohamed elhoseiny in this paper we dived into normalization techniques used in zero shot learning zsl we showed how scaled cosine similarity and attribute normalization influence the signal s variance inside a model we showed that for deeper models there is a need for other normalization procedures and developed class normalization which is similar to batch normalization but applied across the class dimension using class normalization we built an mlp model that achieves state of the art performance and trains x50 200 times faster than the current sota we also formulated a novel continual zero shot learning problem and tested our approach in that setup read more 2020 arxiv code loss patterns of neural networks beyond first order methods in ml workshop neurips 2019 ivan skorokhodov mikhail burtsev using mode connectivity ideas we searched loss landscapes of different neural networks for different visual patterns due to the extreme overparametrization it turned out that any pattern can be found inside the surface this indicates that the loss landscapes of deep models are very complex and contain many irregularities read more 2019 arxiv code programming projects dl reasoner rust a tableau based reasoner for the alcq description logic supporting abox consistency checking with model generation and concept subsumption over a tbox it implements all eight expansion rules conjunction disjunction universal existential restrictions at least at most cardinality choice and general concept inclusions firelab python pytorch over the past few years i have been building a framework for running deep learning experiments in pytorch and using it across my research projects it is very similar to pytorch lightning hydra but without proper documentation or testing _ ツ _ omniplan web app javascript react omniplan was used extensively at my previous job but didn t have any web interface which annoyed everyone so i built one using their official api rtrs rust rtrs is a small ray tracing rasterization engine written in rust it works on both meshes and traditional quadrics and has some cool features like distributed rt bvhs arcball rotations etc non uniform interpolation cuda while existing interpolation techniques nearest neighbour bilinear lanczos hamming etc assume that the known point positions form a uniform grid this is not always the case moreover one would often like to backpropagate through these point positions in this project i implemented a cuda kernel for point interpolation on a non uniform grid based on a gaussian mixture model built with hugo by claude last updated jun 2026