\n","updatedAt":"2024-06-09T05:08:03.136Z","author":{"_id":"6186ddf6a7717cb375090c01","avatarUrl":"/avatars/716b6a7d1094c8036b2a8a7b9063e8aa.svg","fullname":"Julien BLANCHON","name":"blanchon","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":142}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.46547985076904297},"editors":["blanchon"],"editorAvatarUrls":["/avatars/716b6a7d1094c8036b2a8a7b9063e8aa.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2202.09778","authors":[{"_id":"6411c7806b75ddced3890fe4","user":{"_id":"63e8a1fc46574e63a2bf3e88","avatarUrl":"/avatars/b91f587920c9102cabeca77d72d33995.svg","isPro":false,"fullname":"Luping Liu","user":"luping-liu","type":"user"},"name":"Luping Liu","status":"claimed_verified","statusLastChangedAt":"2024-10-16T09:45:37.934Z","hidden":false},{"_id":"6411c7806b75ddced3890fe5","name":"Yi Ren","hidden":false},{"_id":"6411c7806b75ddced3890fe6","name":"Zhijie Lin","hidden":false},{"_id":"6411c7806b75ddced3890fe7","name":"Zhou Zhao","hidden":false}],"publishedAt":"2022-02-20T10:37:52.000Z","title":"Pseudo Numerical Methods for Diffusion Models on Manifolds","summary":"Denoising Diffusion Probabilistic Models (DDPMs) can generate high-quality\nsamples such as image and audio samples. However, DDPMs require hundreds to\nthousands of iterations to produce final samples. Several prior works have\nsuccessfully accelerated DDPMs through adjusting the variance schedule (e.g.,\nImproved Denoising Diffusion Probabilistic Models) or the denoising equation\n(e.g., Denoising Diffusion Implicit Models (DDIMs)). However, these\nacceleration methods cannot maintain the quality of samples and even introduce\nnew noise at a high speedup rate, which limit their practicability. To\naccelerate the inference process while keeping the sample quality, we provide a\nfresh perspective that DDPMs should be treated as solving differential\nequations on manifolds. Under such a perspective, we propose pseudo numerical\nmethods for diffusion models (PNDMs). Specifically, we figure out how to solve\ndifferential equations on manifolds and show that DDIMs are simple cases of\npseudo numerical methods. We change several classical numerical methods to\ncorresponding pseudo numerical methods and find that the pseudo linear\nmulti-step method is the best in most situations. According to our experiments,\nby directly using pre-trained models on Cifar10, CelebA and LSUN, PNDMs can\ngenerate higher quality synthetic images with only 50 steps compared with\n1000-step DDIMs (20x speedup), significantly outperform DDIMs with 250 steps\n(by around 0.4 in FID) and have good generalization on different variance\nschedules. Our implementation is available at\nhttps://github.com/luping-liu/PNDM.","upvotes":0,"discussionId":"641192373ea54b1aa7e2f508","ai_summary":"PNDMs, a pseudo numerical method for diffusion models, accelerate the generation of high-quality synthetic images while maintaining or improving sample quality compared to existing methods.","ai_keywords":["Denoising Diffusion Probabilistic Models","DDPMs","variance schedule","denoising equation","Denoising Diffusion Implicit Models","DDIMs","differential equations on manifolds","pseudo numerical methods","pseudo linear multi-step method","FID","Cifar10","CelebA","LSUN"]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[],"acceptLanguages":["*"]}">
PNDMs, a pseudo numerical method for diffusion models, accelerate the generation of high-quality synthetic images while maintaining or improving sample quality compared to existing methods.
AI-generated summary
Denoising Diffusion Probabilistic Models (DDPMs) can generate high-quality
samples such as image and audio samples. However, DDPMs require hundreds to
thousands of iterations to produce final samples. Several prior works have
successfully accelerated DDPMs through adjusting the variance schedule (e.g.,
Improved Denoising Diffusion Probabilistic Models) or the denoising equation
(e.g., Denoising Diffusion Implicit Models (DDIMs)). However, these
acceleration methods cannot maintain the quality of samples and even introduce
new noise at a high speedup rate, which limit their practicability. To
accelerate the inference process while keeping the sample quality, we provide a
fresh perspective that DDPMs should be treated as solving differential
equations on manifolds. Under such a perspective, we propose pseudo numerical
methods for diffusion models (PNDMs). Specifically, we figure out how to solve
differential equations on manifolds and show that DDIMs are simple cases of
pseudo numerical methods. We change several classical numerical methods to
corresponding pseudo numerical methods and find that the pseudo linear
multi-step method is the best in most situations. According to our experiments,
by directly using pre-trained models on Cifar10, CelebA and LSUN, PNDMs can
generate higher quality synthetic images with only 50 steps compared with
1000-step DDIMs (20x speedup), significantly outperform DDIMs with 250 steps
(by around 0.4 in FID) and have good generalization on different variance
schedules. Our implementation is available at
https://github.com/luping-liu/PNDM.