Report provides results for novel audio, video and image tokenizers. They were developed specifically for subsequent generation by LDM in text-conditioned setting. Series of experiments confirm their ability to surpass leading opensource methods. Inference code available at: <a href=\"https://github.com/kandinskylab/kvae\" rel=\"nofollow\">https://github.com/kandinskylab/kvae</a> with corresponding weights at:</p>\n<ul>\n<li>Audio -- <a href=\"https://huggingface.co/kandinskylab/KVAE-Audio\">https://huggingface.co/kandinskylab/KVAE-Audio</a></li>\n<li>Video<ul>\n<li>4x16x16 -- <a href=\"https://huggingface.co/kandinskylab/KVAE-3D-2.0-t4s16\">https://huggingface.co/kandinskylab/KVAE-3D-2.0-t4s16</a></li>\n<li>4x8x8 -- <a href=\"https://huggingface.co/kandinskylab/KVAE-3D-2.0-t4s8\">https://huggingface.co/kandinskylab/KVAE-3D-2.0-t4s8</a></li>\n</ul>\n</li>\n<li>Image -- <a href=\"https://huggingface.co/kandinskylab/KVAE-2D-2.0\">https://huggingface.co/kandinskylab/KVAE-2D-2.0</a></li>\n</ul>\n","updatedAt":"2026-08-07T18:30:17.354Z","author":{"_id":"68cc16dc1b7328ba01df2e45","avatarUrl":"/avatars/5ab0c5027b33a0680e03d3f4c92ebad3.svg","fullname":"Kirill Chernyshev","name":"rtyasdf","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7709617614746094},"editors":["rtyasdf"],"editorAvatarUrls":["/avatars/5ab0c5027b33a0680e03d3f4c92ebad3.svg"],"reactions":[],"isReport":false},"replies":[{"id":"6a7637007e4e34d640d41810","author":{"_id":"63177d85f957903db971a173","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1665094764329-63177d85f957903db971a173.png","fullname":"Artem","name":"kabachuha","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":78,"isUserFollowing":false},"createdAt":"2026-08-07T19:50:24.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"Minimax-destroying Kandinsky-video 6.0 when?1!!1 /s\n\nAwesome work, hope you accomplish what you started and show that we also have something to show the world, not only the West or the East","html":"<p>Minimax-destroying Kandinsky-video 6.0 when?1!!1 /s</p>\n<p>Awesome work, hope you accomplish what you started and show that we also have something to show the world, not only the West or the East</p>\n","updatedAt":"2026-08-07T19:50:24.236Z","author":{"_id":"63177d85f957903db971a173","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1665094764329-63177d85f957903db971a173.png","fullname":"Artem","name":"kabachuha","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":78,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9158686399459839},"editors":["kabachuha"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1665094764329-63177d85f957903db971a173.png"],"reactions":[],"isReport":false,"parentCommentId":"6a76243974be585b61ef7a7a"}}]}],"primaryEmailConfirmed":false,"paper":{"id":"2608.05798","authors":[{"_id":"6a75b8d7e1228e04b3238475","name":"Andrey Shutkin","hidden":false},{"_id":"6a75b8d7e1228e04b3238476","name":"Denis Parkhomenko","hidden":false},{"_id":"6a75b8d7e1228e04b3238477","name":"Ivan Kirillov","hidden":false},{"_id":"6a75b8d7e1228e04b3238478","user":{"_id":"68cc16dc1b7328ba01df2e45","avatarUrl":"/avatars/5ab0c5027b33a0680e03d3f4c92ebad3.svg","isPro":false,"fullname":"Kirill Chernyshev","user":"rtyasdf","type":"user","name":"rtyasdf"},"name":"Kirill Chernyshev","status":"claimed_verified","statusLastChangedAt":"2026-08-07T16:45:28.207Z","hidden":false},{"_id":"6a75b8d7e1228e04b3238479","name":"Kirill Malakhov","hidden":false},{"_id":"6a75b8d7e1228e04b323847a","name":"Ilia Vasiliev","hidden":false},{"_id":"6a75b8d7e1228e04b323847b","name":"Ilia Trushkin","hidden":false},{"_id":"6a75b8d7e1228e04b323847c","name":"Valeriya Kobenko","hidden":false},{"_id":"6a75b8d7e1228e04b323847d","user":{"_id":"664ca5856970d6ba864e0b6f","avatarUrl":"/avatars/4f8f6b0755d8b1ceb48eda1127444e0d.svg","isPro":false,"fullname":"David Chikovani","user":"davidchick","type":"user","name":"davidchick"},"name":"David Chikovani","status":"claimed_verified","statusLastChangedAt":"2026-08-07T16:45:28.202Z","hidden":false},{"_id":"6a75b8d7e1228e04b323847e","user":{"_id":"662b63e9e58d277c811f9cf9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/662b63e9e58d277c811f9cf9/EKqOzkWnUQEO9m59-a-7u.jpeg","isPro":false,"fullname":"Alexander Ivanov","user":"ivnvalex","type":"user","name":"ivnvalex"},"name":"Alexander Ivanov","status":"claimed_verified","statusLastChangedAt":"2026-08-07T16:45:28.213Z","hidden":false},{"_id":"6a75b8d7e1228e04b323847f","name":"Azat Saginbaev","hidden":false},{"_id":"6a75b8d7e1228e04b3238480","name":"Egor Silvestrov","hidden":false},{"_id":"6a75b8d7e1228e04b3238481","name":"Ivan Mikheev","hidden":false},{"_id":"6a75b8d7e1228e04b3238482","name":"Konstantin Zakharov","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/68cc16dc1b7328ba01df2e45/k_ok8P8J5yeNbLcv1ixIE.png","https://cdn-uploads.huggingface.co/production/uploads/68cc16dc1b7328ba01df2e45/JV1gK1OZ8w5hT4VNruj_8.png","https://cdn-uploads.huggingface.co/production/uploads/68cc16dc1b7328ba01df2e45/6fofDFnZ4QKdaaJcQCBj4.png","https://cdn-uploads.huggingface.co/production/uploads/68cc16dc1b7328ba01df2e45/RLx9vuqbks5HZZsDuRpJ4.png"],"publishedAt":"2026-08-06T00:00:00.000Z","submittedOnDailyAt":"2026-08-07T00:00:00.000Z","title":"KVAE: Family of Tokenizers for Multimodal Generative Models","submittedOnDailyBy":{"_id":"68cc16dc1b7328ba01df2e45","avatarUrl":"/avatars/5ab0c5027b33a0680e03d3f4c92ebad3.svg","isPro":false,"fullname":"Kirill Chernyshev","user":"rtyasdf","type":"user","name":"rtyasdf"},"summary":"Latent diffusion modeling (LDM), a prominent paradigm, utilizes tokenizers to map input signal to compressed representation. This dependency positions tokenizer as an integral part of generation process itself, since it affects learning speed, quality of synthesized samples and lay foundation for later applications. This report presents series of KVAE tokenizers for audio, image and video, all designed for subsequent text-conditioned generation: KVAE-Audio, a continuous full-band 48 kHz tokenizer with a 50 Hz latent of 64 channels; KVAE-3D -- two causal video tokenizers for 4x16x16 and 4x8x8 compression; KVAE-2D, an image model, compressing input by factor of 8 with 32 channels. We demonstrate that reconstruction (PSNR, LPIPS, PESQ, etc.) and generation results on objective (Frechet Distance, CLIP score, CLAP score, etc.) and subjective (side-by-side evaluation) metrics matches or surpasses frontier opensource tokenizers, such as VAEs from Wan-2.2, HunyuanVideo-1.5, FLUX.2, MovieGen, StableAudio and MMAudio. Considering difficulty of development, we share with community training details, model selection method and ablation on design choices. The code is publicly available at https://github.com/kandinskylab/kvae and https://github.com/kandinskylab/kvae-audio.","upvotes":10,"discussionId":"6a75b8d7e1228e04b3238483","githubRepo":"https://github.com/kandinskylab/kvae","githubRepoAddedBy":"user","githubStars":48,"organization":{"_id":"690f6cebeed4fe1e4fafae4c","name":"kandinskylab","fullname":"Kandinsky Lab","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/60991602f7c9c7bf29603a88/oDAZ2f49nmDJKICzgHglO.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"62653f745f6f2e14d6ae128c","avatarUrl":"/avatars/944b564ab810a5b31fa5e45f63bdf4ee.svg","isPro":false,"fullname":"Ivan Kirillov","user":"funnylittleman","type":"user"},{"_id":"691efbe1623e254dabf255eb","avatarUrl":"/avatars/453aba27bca45ae24ff348694a272685.svg","isPro":false,"fullname":"Valery","user":"svaleryk","type":"user"},{"_id":"68cc16dc1b7328ba01df2e45","avatarUrl":"/avatars/5ab0c5027b33a0680e03d3f4c92ebad3.svg","isPro":false,"fullname":"Kirill Chernyshev","user":"rtyasdf","type":"user"},{"_id":"628a57f7042f83532f40c37d","avatarUrl":"/avatars/4b175b1b6f42259ea47a2d804f1070ec.svg","isPro":false,"fullname":"Konstantin","user":"Nikkisora","type":"user"},{"_id":"662b63e9e58d277c811f9cf9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/662b63e9e58d277c811f9cf9/EKqOzkWnUQEO9m59-a-7u.jpeg","isPro":false,"fullname":"Alexander Ivanov","user":"ivnvalex","type":"user"},{"_id":"6a1094bddb3c66d898cd8b3e","avatarUrl":"/avatars/5a7f546ee3b10d612fba96bb40c176b5.svg","isPro":false,"fullname":"M","user":"kirill-malakhov","type":"user"},{"_id":"664ca5856970d6ba864e0b6f","avatarUrl":"/avatars/4f8f6b0755d8b1ceb48eda1127444e0d.svg","isPro":false,"fullname":"David Chikovani","user":"davidchick","type":"user"},{"_id":"63177d85f957903db971a173","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1665094764329-63177d85f957903db971a173.png","isPro":false,"fullname":"Artem","user":"kabachuha","type":"user"},{"_id":"6687f9a71309e08b1f84bdc6","avatarUrl":"/avatars/f947ec9fe620ae4cffa83b371acdd571.svg","isPro":false,"fullname":"MeiYi","user":"natalie5","type":"user"},{"_id":"66872630abca3bbcec480c94","avatarUrl":"/avatars/b1e4ce1808fddaeb9975f7ba298c86f0.svg","isPro":false,"fullname":"Cris","user":"Bubmy","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"690f6cebeed4fe1e4fafae4c","name":"kandinskylab","fullname":"Kandinsky Lab","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/60991602f7c9c7bf29603a88/oDAZ2f49nmDJKICzgHglO.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.05798.md","query":{}}">