PixRestore Uyasebenza

I-PixRestore iyimodeli ehlanganisiwe yokulungisa izithombe eyethulwe ngokubambisana yi-Hong Kong Polytechnic University kanye ne-OPPO Research Institute. Imodeli isebenzisa i-Diffusion Transformer esikhaleni samaphikseli ukwenza i-flow matching ngqo ezingeni lamaphikseli, igweme ukulahleka kwemininingwane. I-PixRestore isebenzisa indlela yokuhambisa izici eguquguqukayo eqhutshwa yi-DINO; le modeli elula enamapharamitha angu-53.7M kuphela ingacubungula kahle imisebenzi engu-8 yokuwohloka, okuhlanganisa ukukhulisa ukulungiswa, ukususa umsindo, ukufiphala, imizila yemvula kanye nenkungu. Ukuqagela ngesinyathelo esisodwa kuthatha ama-millisecond angu-44 kuphela, futhi imodeli ifinyelela ezingeni elihamba phambili ekugcineni ukwethembeka nasekusebenzeni kahle.

Ukukhulisa ukulungiswa: Iguqula izithombe ezinokulungiswa okuphansi zibe izithombe ezicacile ezinokulungiswa okuphezulu nemininingwane eminingi.
Ukususa umsindo ezithombeni: Isusa izinhlayiya zomsindo ezingahleliwe, ibuyise ukuthungwa okuhlanzekile.
Ukususa ukufiphala kokunyakaza: Isusa ukufiphala okubangelwa ukunyakaza noma ukungagxili, ibuyise imiphetho ebukhali.
Ukususa imizila yemvula: Isusa ukuphazamiseka kwemizila yemvula eminyene, ibuyise okuqukethwe kwangemuva.
Ukususa amaconsi emvula: Isusa ukuvinjwa okubangelwa amaconsi emvula engilazini noma elensini, ibuyise umbono ocacile.
Ukususa inkungu ezithombeni: Isusa inkungu ebangelwa yisimo somoya, ithuthukise ukucaca nemibala yesithombe.
Ukususa iqhwa: Isusa amachashazi eqhwa ezithombeni, ibuyise ukubukeka kwangempela kwesigcawu.
Ukuthuthukisa ukukhanya okuphansi: Kwenyusa ukukhanya nemininingwane ezindaweni ezimnyama, kuthuthukise ukubonakala kwesithombe.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI avulekile
I-Pixel-space Diffusion Transformer: I-PixRestore yenza i-flow matching ngqo ezingeni lamaphikseli, iyilahle ngokuphelele i-VAE encoder ukuze igweme ukulahleka kwemininingwane yemvamisa ephezulu okubangelwa ukucindezelwa esikhaleni esifihlekile. Iphinde igweme ukuncika kumamodeli aqeqeshwe kusengaphambili e-T2I, iqinisekise ukuthi imiphumela yokulungisa ihambisana kakhulu nokuqukethwe kwasekuqaleni.
Ukuhanjiswa kwezici ngokwezendlalelo okuvumelana nezimo: Imodeli isebenzisa ukufana kwezici ze-DINO zezithombe ze-LQ–HQ ukuze ihlole ngamandla ukwethembeka kwezici ezivela ezendlalelweni ezihlukene zesifaki khodi esibonakalayo; ihlanganisa izendlalelo ezithembekile zibe yizimo eziminyene ezijovwa ku-DiT, kuyilapho izendlalelo ezingathembekile zithola ukuqondiswa okunamandla kwezici zekhwalithi ephezulu, ngaleyo ndlela isuse ngokuzenzakalelayo izinhlobo ezihlukene zokuwohloka.
Ukuqagela okusheshayo ngesinyathelo esisodwa: Ngemva kokuqeqeshwa okuphelele, i-PixRestore icushwa kabusha ngomgomo wokuphikisana osuselwe ku-DINO ukuze ibe yijeneretha yesinyathelo esisodwa. Ngamapharamitha angu-53.7M kuphela, ingaqedela ukuqagela phakathi kwama-millisecond angu-44, ifinyelele ibhalansi engcono kakhulu phakathi kokwethembeka, ikhwalithi ebonakalayo nokusebenza kahle kokubala.

Ukulungisa indawo: Klona inqolobane yekhodi ye-PixRestore ku-GitHub bese ufaka okuncikile okudingekayo.
Ukuthola imodeli: Landa amafayela ezisindo zemodeli eziqeqeshwe kusengaphambili ku-Hugging Face.
Ukulungisa izithombe: Lungisa izithombe ezisezingeni eliphansi okufanele zilungiswe; kusekelwa izinhlobo ezingu-8 zokuwohloka, okuhlanganisa ukukhulisa ukulungiswa, ukususa umsindo nokususa ukufiphala.
Ukwenza ukuqagela: Qalisa iskripthi sokuqagela. Imodeli isebenzisa ukuqagela kwesinyathelo esisodwa ngokuzenzakalelayo futhi ikhiphe umphumela wokulungisa cishe ngama-millisecond angu-44.
Ukugcina imiphumela: Thola futhi ulondoloze izithombe ezilungiswe ngekhwalithi ephezulu ukuze zisetshenziswe ngokuqondile kwezinye izinhlelo.

Ukwethembeka ezingeni lamaphikseli: Imodeli yokusabalalisa yenziwa ngqo esikhaleni samaphikseli, igweme ngokuphelele ukulahleka kwemininingwane yemvamisa ephezulu nezinto ezingezona ezangempela zokuqukethwe okubangelwa ukufakwa kwekhodi kwe-VAE.
Ilula kakhulu futhi isebenza kahle: Inamapharamitha angu-53.7M kuphela, futhi ukuqagela ngesinyathelo esisodwa kuthatha ama-millisecond angu-44 kuphela, okwenza izindleko zokubala zibe phansi kakhulu kunezixazululo ezivamile zokulungisa ezisebenzisa ukusabalalisa esikhaleni esifihlekile.
Imboza imisebenzi eminingi ngemodeli eyodwa: Imodeli eyodwa ing处理 ukukhulisa ukulungiswa, ukususa umsindo, ukufiphala, imvula, inkungu, iqhwa kanye nokuthuthukisa ukukhanya okuphansi, phakathi kwezinye izinhlobo ezingu-8 zokuwohloka.
Ukubona ukuwohloka okuvumelana nezimo: Ihlola ngamandla ukwethembeka kwezendlalelo ezihlukene ngokufana kwezici ze-DINO, yenze ibhalansi ehlakaniphile phakathi kokususa ukuwohloka nokugcina imininingwane.
Indlela yokuqeqesha ezimele: Ayincikile kunoma iyiphi imodeli yombhalo uye esithombeni eqeqeshwe kusengaphambili; iqeqeshwa ngokuphelele kusukela ekuqaleni ukuze igweme ukuchema kokuqukethwe okungeniswa ulwazi lokuqala lokukhiqiza.
Ukusebenza okuhamba phambili ngokuphelele: Kumabhentshimakhi asesidlangalaleni nasemasethini okuhlola omhlaba wangempela, idlula amamodeli akhona ahlanganisiwe okulungisa ekwethembekeni, kwikhwalithi ebonakalayo nasekhonweni lokususa ukuwohloka.

Iwebhusayithi esemthethweni: https://csslc.github.io/pixrestore-page
Inqolobane ye-GitHub: https://github.com/csslc/PixRestore
Inqolobane yemodeli ku-Hugging Face: https://huggingface.co/VCLab-PolyU/PixRestore
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2608.16793

Ukuthuthukisa izithombe ngesikhathi sangempela kumafoni: Ngezinga lokubambezeleka eliphansi kakhulu lama-millisecond angu-44, ilungisa ngesikhathi sangempela umsindo wasebusuku, ukufiphala kokunyakaza namaconsi emvula kumadivayisi eselula afana ne-OPPO, ithuthukise ikhwalithi yezithombe ezithathiwe.
Umbono wezimoto ezizishayelayo namadrone: Esimweni sezulu esibi esinemvula, inkungu noma iqhwa, isusa ukuwohloka kwezithombe ngendlela ehlanganisiwe, ibuyise imininingwane yezimpawu zomgwaqo nezithiyo, ithuthukise ukwethembeka kohlelo lokubona.
Ukulungisa ikhwalithi yamavidiyo okuqapha: Isebenzisa ukukhulisa ukulungiswa nokususa umsindo kumavidiyo okuqapha amadala noma acindezelwe kakhulu, ibuyise imininingwane ebalulekile efana nobuso namapuleti ezimoto ukuze isize ekuqoqeni ubufakazi bezokuphepha.
Ukufakwa kwedijithali kwezithombe ezindala namarekhodi omlando: Ilungisa ngokuhlanganisiwe ukufiphala kwemibala, imihuzuko, amabala esikhunta nezinkinga zokukhanya okuphansi, iguqule izithombe zomlando eziyigugu zibe amarekhodi edijithali athembeke kakhulu.
Ukuthuthukisa izithombe zezokwelapha: Ithuthukisa ikhwalithi yezithombe zamakhamera angaphakathi emzimbeni, amakroskopu nezinye izinsiza ezisebenza ekukhanyeni okuphansi, isuse umsindo nokufiphala ukuze isize odokotela babone imininingwane yezilonda ngokucacile.