Meta tags:

Headings (most frequently used words):

diffusion, sds, fix, with, mean, shift, distillation, for, mode, seeking, illustrating, the, shortcomings, of, and, how, we, it, results, stable, bibtex, what, does, this, look, like, our,

Text of the page (most frequently used words):
the (9), and (9), sds (7), #diffusion (6), with (6), image (5), this (4), distillation (4), mean (3), shift (3), for (3), mode (3), seeking (3), code (2), based (2), from (2), vikas (2), thamizharasan (2), nikitas (2), chatzis (2), iliyan (2), georgiev (2), matthew (2), fisher (2), difan (2), liu (2), nanxuan (2), zhao (2), evangelos (2), kalogerakis (2), michal (2), 2502 (2), 15989 (2), arxiv (2), results (2), distribution (2), well (2), technique (2), fix (2), simulate (2), optimization (2), samples (2), score (2), model (2), z_t (2), sampling (2), models (2), 2022 (2), website, heavily, template, visual, brown, edu, misc, thamizharasan2025meanshiftdistillationdiffusionmode, title, author, lukac, year, 2025, eprint, archiveprefix, primaryclass, url, https, org, abs, bibtex, coming, soon, stable, putting, all, together, propose, gradient, proxy, known, our, let, multiple, points, densely, initialized, along, grid, across, canvas, after, several, steps, observe, optimized, fail, fit, what, does, look, like, begin, training, simple, epsilon_, theta, approx, sigma_t, nabla_, log, then, inference, time, draw, via, ddim, song, 2021a, popular, first, order, algorithm, without, guidance, cfg, salimans, 2021, suffers, significant, bias, variance, yielding, inaccurate, gradients, manifests, over, smoothened, when, optimizing, text, illustrate, pitfalls, using, small, denoising, network, watch, blog, video, mins, mp4, poole, wang, has, emerged, useful, leveraging, priors, learned, large, scale, beyond, raster, images, provides, procedure, estimate, parameters, differentiable, generator, such, that, rendered, pushed, towards, higher, probability, region, pre, trained, prompt, conditioned, illustrating, shortcomings, how, paper, crete, ntua, adobe, research, university, massachusetts, amherst, lukáč, meanshiftdistillation,

Text of the page (random words):
meanshiftdistillation mean shift distillation for diffusion mode seeking vikas thamizharasan 1 2 nikitas chatzis 3 iliyan georgiev 2 matthew fisher 2 difan liu 2 nanxuan zhao 2 evangelos kalogerakis 1 4 michal lukáč 2 1 university of massachusetts amherst 2 adobe research 3 ntua 4 tu crete paper code illustrating the shortcomings of sds and how we fix it score distillation sampling sds poole et al 2022 wang et al 2022 has emerged as a useful technique for leveraging the priors learned by large scale image models beyond 2d raster images sds provides an optimization procedure to estimate the parameters of a differentiable image generator such that the rendered image is pushed towards a higher probability region of a pre trained prompt conditioned image diffusion model sds suffers from significant bias as well as variance yielding inaccurate gradients this manifests as over smoothened results when optimizing with text to image diffusion models to illustrate the pitfalls of sds we simulate it in 2d using a small denoising diffusion network watch blog as video 3 mins mp4 we begin my training a simple score model epsilon_ theta approx sigma_t nabla_ z_t log p z_t c then at inference time we draw samples via ddim song et al 2021a a popular first order sampling algorithm without and with guidance cfg ho salimans 2021 what does this look like with sds let s simulate this with multiple points densely initialized along a grid across the canvas after several optimization steps we observe samples optimized with sds fail to fit the distribution our fix we propose mean shift distillation a distribution gradient proxy based on a well known mode seeking technique putting it all together results with stable diffusion coming soon bibtex misc thamizharasan2025meanshiftdistillationdiffusionmode title mean shift distillation for diffusion mode seeking author vikas thamizharasan and nikitas chatzis and iliyan georgiev and matthew fisher and difan liu and nanxuan zhao and evangelos kalogerakis and michal lukac year 2025 eprint 2502 15989 archiveprefix arxiv primaryclass cs lg url https arxiv org abs 2502 15989 the code of this website is heavily based on the template from visual cs brown edu