{"id":1146,"date":"2026-06-10T12:16:10","date_gmt":"2026-06-10T04:16:10","guid":{"rendered":"https:\/\/imgedits.net\/"},"modified":"2026-06-10T12:17:45","modified_gmt":"2026-06-10T04:17:45","slug":"understanding-ai-image","status":"publish","type":"post","link":"https:\/\/imgedits.net\/vi\/post\/understanding-ai-image\/","title":{"rendered":"Understanding the Foundations of AI Image Generation"},"content":{"rendered":"

Kh\u1ea3 n\u0103ng c\u1ee7a tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o trong vi\u1ec7c t\u1ea1o ra h\u00ecnh \u1ea3nh ch\u00e2n th\u1ef1c, c\u00e1c t\u00e1c ph\u1ea9m ngh\u1ec7 thu\u1eadt ph\u1ee9c t\u1ea1p v\u00e0 c\u00e1c thi\u1ebft k\u1ebf th\u1ecb gi\u00e1c tinh x\u1ea3o t\u1eeb c\u00e1c l\u1eddi nh\u1eafc v\u0103n b\u1ea3n \u0111\u01a1n gi\u1ea3n l\u00e0 m\u1ed9t trong nh\u1eefng b\u01b0\u1edbc nh\u1ea3y v\u1ecdt c\u00f4ng ngh\u1ec7 \u0111\u00e1ng kinh ng\u1ea1c nh\u1ea5t c\u1ee7a th\u1ebf k\u1ef7 21. Trong nhi\u1ec1u th\u1eadp k\u1ef7, \u0111\u1ed3 h\u1ecda m\u00e1y t\u00ednh ph\u1ee5 thu\u1ed9c nghi\u00eam ng\u1eb7t v\u00e0o c\u00e1c thu\u1eadt to\u00e1n th\u1ee7 t\u1ee5c, m\u00f4 h\u00ecnh h\u00ecnh h\u1ecdc v\u00e0 s\u1ef1 thao t\u00e1c th\u1ee7 c\u00f4ng c\u1ee7a c\u00e1c ngh\u1ec7 s\u0129 con ng\u01b0\u1eddi. Ng\u00e0y nay, c\u00e1c m\u00f4 h\u00ecnh AI t\u1ea1o sinh c\u00f3 th\u1ec3 t\u1ed5ng h\u1ee3p nh\u1eefng h\u00ecnh \u1ea3nh ho\u00e0n to\u00e0n m\u1edbi ch\u1ec9 trong v\u00e0i gi\u00e2y. S\u1ef1 chuy\u1ec3n d\u1ecbch m\u00f4 h\u00ecnh n\u00e0y kh\u00f4ng xu\u1ea5t ph\u00e1t t\u1eeb m\u1ed9t s\u1ef1 hi\u1ec3u bi\u1ebft k\u1ef3 di\u1ec7u v\u1ec1 ngh\u1ec7 thu\u1eadt, m\u00e0 t\u1eeb c\u00e1c khung to\u00e1n h\u1ecdc ti\u00ean ti\u1ebfn, c\u01a1 s\u1edf h\u1ea1 t\u1ea7ng t\u00ednh to\u00e1n kh\u1ed5ng l\u1ed3 v\u00e0 ph\u00e2n t\u00edch th\u1ed1ng k\u00ea s\u00e2u s\u1eafc v\u1ec1 d\u1eef li\u1ec7u th\u1ecb gi\u00e1c. \u0110\u1ec3 th\u1ef1c s\u1ef1 hi\u1ec3u c\u00e1ch AI t\u1ea1o ra m\u1ed9t h\u00ecnh \u1ea3nh, ng\u01b0\u1eddi ta ph\u1ea3i nh\u00ecn xa h\u01a1n giao di\u1ec7n ng\u01b0\u1eddi d\u00f9ng v\u00e0 kh\u00e1m ph\u00e1 c\u1ea5u tr\u00fac n\u1ec1n t\u1ea3ng c\u1ee7a c\u00e1c m\u1ea1ng th\u1ea7n kinh, kh\u00f4ng gian bi\u1ec3u di\u1ec5n v\u00e0 m\u00f4 h\u00ecnh x\u00e1c su\u1ea5t. C\u1ed1t l\u00f5i c\u1ee7a m\u1ecdi AI t\u1ea1o sinh n\u1eb1m \u1edf kh\u00e1i ni\u1ec7m h\u1ecdc m\u00e1y t\u1eeb d\u1eef li\u1ec7u l\u1edbn (big data). Tr\u01b0\u1edbc khi m\u1ed9t h\u1ec7 th\u1ed1ng c\u00f3 th\u1ec3 t\u1ea1o ra m\u1ed9t h\u00ecnh \u1ea3nh ch\u1ea5t l\u01b0\u1ee3ng cao v\u1ec1 m\u1ed9t \u201cch\u00fa ch\u00f3 golden retriever \u0111ang ch\u01a1i trong c\u00f4ng vi\u00ean v\u00e0o l\u00fac ho\u00e0ng h\u00f4n\u201d, n\u00f3 ph\u1ea3i ph\u00e2n t\u00edch h\u00e0ng tri\u1ec7u ho\u1eb7c th\u1eadm ch\u00ed h\u00e0ng t\u1ef7 h\u00ecnh \u1ea3nh hi\u1ec7n c\u00f3 v\u00e0 c\u00e1c m\u00f4 t\u1ea3 v\u0103n b\u1ea3n t\u01b0\u01a1ng \u1ee9ng c\u1ee7a ch\u00fang. Giai \u0111o\u1ea1n n\u00e0y \u0111\u01b0\u1ee3c g\u1ecdi l\u00e0 hu\u1ea5n luy\u1ec7n. Trong qu\u00e1 tr\u00ecnh n\u00e0y, m\u1ed9t m\u1ea1ng th\u1ea7n kinh \u2014 m\u1ed9t c\u1ea5u tr\u00fac t\u00ednh to\u00e1n ph\u1ee9c t\u1ea1p l\u1ea5y c\u1ea3m h\u1ee9ng t\u1eeb c\u00e1c t\u1ebf b\u00e0o th\u1ea7n kinh li\u00ean k\u1ebft trong n\u00e3o ng\u01b0\u1eddi \u2014 s\u1ebd qu\u00e9t t\u1eadp d\u1eef li\u1ec7u \u0111\u1ec3 x\u00e1c \u0111\u1ecbnh c\u00e1c m\u1eabu, k\u1ebft c\u1ea5u, h\u00ecnh d\u1ea1ng v\u00e0 m\u00e0u s\u1eafc. M\u1ea1ng l\u01b0\u1edbi h\u1ecdc c\u00e1ch li\u00ean k\u1ebft c\u00e1c s\u1eafp x\u1ebfp pixel c\u1ee5 th\u1ec3 v\u1edbi c\u00e1c kh\u00e1i ni\u1ec7m ng\u1eef ngh\u0129a, ch\u1eb3ng h\u1ea1n nh\u01b0 k\u1ebft c\u1ea5u m\u1ec1m m\u1ea1i c\u1ee7a l\u00f4ng \u0111\u1ed9ng v\u1eadt, \u0111\u1eb7c t\u00ednh ph\u1ea3n chi\u1ebfu c\u1ee7a n\u01b0\u1edbc, ho\u1eb7c c\u00e1c t\u00f4ng m\u00e0u \u1ea5m \u00e1p \u0111\u1eb7c tr\u01b0ng c\u1ee7a b\u1ea7u tr\u1eddi bu\u1ed5i chi\u1ec1u. Theo th\u1eddi gian, h\u1ec7 th\u1ed1ng chuy\u1ec3n t\u1eeb vi\u1ec7c ch\u1ec9 nh\u1eadn di\u1ec7n c\u00e1c \u0111\u1ed1i t\u01b0\u1ee3ng sang vi\u1ec7c hi\u1ec3u c\u00e1c m\u1ed1i quan h\u1ec7 th\u1ed1ng k\u00ea gi\u1eefa ch\u00fang.<\/p>\n\n\n\n

\"M\u1ed9t
M\u1ed9t t\u00e1c ph\u1ea9m ngh\u1ec7 thu\u1eadt k\u1ef9 thu\u1eadt s\u1ed1 tr\u1eebu t\u01b0\u1ee3ng, \u0111\u1ea7y m\u00e0u s\u1eafc \u0111\u01b0\u1ee3c t\u1ea1o ra b\u1edfi AARON, m\u1ed9t h\u1ec7 th\u1ed1ng ngh\u1ec7 thu\u1eadt AI ti\u00ean phong do ngh\u1ec7 s\u0129 Harold Cohen ph\u00e1t tri\u1ec3n, v\u1edbi c\u00e1c h\u00ecnh d\u1ea1ng hoa v\u00e0 h\u00ecnh h\u1ecdc ph\u1ee9c t\u1ea1p.<\/figcaption><\/figure>\n\n\n\n

Tuy nhi\u00ean, m\u1ed9t m\u00f4 h\u00ecnh AI kh\u00f4ng ch\u1ec9 \u0111\u01a1n gi\u1ea3n l\u01b0u tr\u1eef m\u1ed9t c\u01a1 s\u1edf d\u1eef li\u1ec7u kh\u1ed5ng l\u1ed3 c\u00e1c h\u00ecnh \u1ea3nh trong b\u1ed9 nh\u1edb \u0111\u1ec3 sao ch\u00e9p v\u00e0 d\u00e1n ch\u00fang sau n\u00e0y. C\u00e1ch ti\u1ebfp c\u1eadn nh\u01b0 v\u1eady s\u1ebd c\u1ef1c k\u1ef3 k\u00e9m hi\u1ec7u qu\u1ea3 v\u00e0 kh\u00f4ng c\u00f3 kh\u1ea3 n\u0103ng t\u1ea1o ra c\u00e1c t\u00e1c ph\u1ea9m ngh\u1ec7 thu\u1eadt th\u1ef1c s\u1ef1 nguy\u00ean b\u1ea3n. Thay v\u00e0o \u0111\u00f3, qu\u00e1 tr\u00ecnh hu\u1ea5n luy\u1ec7n bu\u1ed9c m\u00f4 h\u00ecnh ph\u1ea3i n\u00e9n \u0111\u1ea1i d\u01b0\u01a1ng th\u00f4ng tin th\u1ecb gi\u00e1c kh\u1ed5ng l\u1ed3 n\u00e0y th\u00e0nh m\u1ed9t kh\u00e1i ni\u1ec7m \u0111\u01b0\u1ee3c t\u1ed5 ch\u1ee9c theo to\u00e1n h\u1ecdc \u0111\u01b0\u1ee3c g\u1ecdi l\u00e0 kh\u00f4ng gian ti\u1ec1m \u1ea9n (latent space). Kh\u00f4ng gian ti\u1ec1m \u1ea9n c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c kh\u00e1i ni\u1ec7m h\u00f3a nh\u01b0 m\u1ed9t h\u1ec7 th\u1ed1ng t\u1ecda \u0111\u1ed9 \u0111a chi\u1ec1u v\u00f4 h\u00ecnh, n\u01a1i c\u00e1c kh\u00e1i ni\u1ec7m t\u01b0\u01a1ng t\u1ef1 \u0111\u01b0\u1ee3c nh\u00f3m l\u1ea1i g\u1ea7n nhau. V\u00ed d\u1ee5, trong c\u00f5i to\u00e1n h\u1ecdc \u1ea9n gi\u1ea5u n\u00e0y, vect\u01a1 \u0111\u1ea1i di\u1ec7n cho m\u1ed9t \"con ch\u00f3\" n\u1eb1m g\u1ea7n vect\u01a1 cho m\u1ed9t \"con m\u00e8o\" trong m\u1ed9t c\u1ee5m \"\u0111\u1ed9ng v\u1eadt\" r\u1ed9ng h\u01a1n, trong khi t\u1ecda \u0111\u1ed9 cho \"ho\u00e0ng h\u00f4n\" \u0111\u01b0\u1ee3c \u0111\u1eb7t g\u1ea7n \"b\u00ecnh minh\" v\u00e0 \"ch\u1ea1ng v\u1ea1ng\". Khi ng\u01b0\u1eddi d\u00f9ng nh\u1eadp m\u1ed9t l\u1eddi nh\u1eafc, AI s\u1ebd \u0111i\u1ec1u h\u01b0\u1edbng kh\u00f4ng gian to\u00e1n h\u1ecdc n\u00e0y, t\u00ecm ra giao \u0111i\u1ec3m ch\u00ednh x\u00e1c c\u1ee7a c\u00e1c kh\u00e1i ni\u1ec7m \u0111\u01b0\u1ee3c y\u00eau c\u1ea7u \u0111\u1ec3 l\u00e0m b\u1ea3n thi\u1ebft k\u1ebf cho \u0111\u1ea7u ra.<\/p>\n\n\n\n

S\u1ef1 ti\u1ebfn h\u00f3a c\u1ee7a ki\u1ebfn tr\u00fac: T\u1eeb GAN \u0111\u1ebfn Variational Autoencoders<\/h2>\n\n\n\n

\u0110\u1ec3 hi\u1ec3u tr\u1ea1ng th\u00e1i hi\u1ec7n \u0111\u1ea1i c\u1ee7a vi\u1ec7c t\u1ea1o h\u00ecnh \u1ea3nh b\u1eb1ng AI, \u0111i\u1ec1u c\u1ea7n thi\u1ebft l\u00e0 ph\u1ea3i truy xu\u1ea5t ngu\u1ed3n g\u1ed1c ti\u1ebfn h\u00f3a c\u1ee7a c\u00e1c ki\u1ebfn tr\u00fac c\u1ed1t l\u00f5i c\u1ee7a n\u00f3. B\u01b0\u1edbc \u0111\u1ed9t ph\u00e1 ban \u0111\u1ea7u trong vi\u1ec7c t\u1ed5ng h\u1ee3p h\u00ecnh \u1ea3nh thuy\u1ebft ph\u1ee5c \u0111\u00e3 \u0111\u1ebfn v\u1edbi s\u1ef1 ra \u0111\u1eddi c\u1ee7a M\u1ea1ng \u0111\u1ed1i ngh\u1ecbch t\u1ea1o sinh (Generative Adversarial Networks), th\u01b0\u1eddng \u0111\u01b0\u1ee3c g\u1ecdi l\u00e0 GAN. \u0110\u01b0\u1ee3c h\u00ecnh th\u00e0nh v\u00e0o n\u0103m 2014, m\u1ed9t GAN ho\u1ea1t \u0111\u1ed9ng d\u1ef1a tr\u00ean nguy\u00ean t\u1eafc c\u1ea1nh tranh xu\u1ea5t s\u1eafc li\u00ean quan \u0111\u1ebfn hai m\u1ea1ng th\u1ea7n kinh ri\u00eang bi\u1ec7t \u0111\u00f3ng vai tr\u00f2 l\u00e0 \u0111\u1ed1i th\u1ee7: B\u1ed9 t\u1ea1o (Generator) v\u00e0 B\u1ed9 ph\u00e2n bi\u1ec7t (Discriminator). M\u1ee5c ti\u00eau duy nh\u1ea5t c\u1ee7a B\u1ed9 t\u1ea1o l\u00e0 t\u1ea1o ra m\u1ed9t h\u00ecnh \u1ea3nh t\u1eeb nhi\u1ec5u to\u00e1n h\u1ecdc ng\u1eabu nhi\u00ean, trong khi vai tr\u00f2 c\u1ee7a B\u1ed9 ph\u00e2n bi\u1ec7t l\u00e0 \u0111\u00e1nh gi\u00e1 h\u00ecnh \u1ea3nh \u0111\u00f3 d\u1ef1a tr\u00ean t\u1eadp d\u1eef li\u1ec7u g\u1ed3m c\u00e1c b\u1ee9c \u1ea3nh th\u1ef1c do con ng\u01b0\u1eddi t\u1ea1o ra v\u00e0 x\u00e1c \u0111\u1ecbnh xem h\u00ecnh \u1ea3nh \u0111\u01b0\u1ee3c t\u1ea1o ra l\u00e0 \"th\u1eadt\" hay \"gi\u1ea3\".<\/p>\n\n\n\n

\n\n\n
\n

Lu\u1ed3ng ki\u1ebfn tr\u00fac GAN<\/h3>\n \n
\n
Nhi\u1ec5u ng\u1eabu nhi\u00ean<\/div>\n
<\/div><\/div>\n
B\u1ed9 t\u1ea1o<\/div>\n
<\/div><\/div>\n
H\u00ecnh \u1ea3nh gi\u1ea3<\/div>\n <\/div>\n
<\/div>\n
\n
B\u1ed9 ph\u00e2n bi\u1ec7t<\/div>\n
<\/div><\/div>\n
H\u00ecnh \u1ea3nh th\u1eadt<\/div>\n <\/div>\n
\n \n \ud83d\udd04 V\u00f2ng l\u1eb7p ph\u1ea3n h\u1ed3i: Th\u1eadt\/Gi\u1ea3\n <\/span>\n <\/div>\n <\/div>\n\n \n<\/article>\n\n\n\n

M\u1ed1i quan h\u1ec7 \u0111\u1ed1i ngh\u1ecbch n\u00e0y t\u1ea1o ra m\u1ed9t v\u00f2ng l\u1eb7p ph\u1ea3n h\u1ed3i hi\u1ec7u qu\u1ea3 cao. Ban \u0111\u1ea7u, B\u1ed9 t\u1ea1o kh\u00f4ng t\u1ea1o ra g\u00ec ngo\u00e0i nh\u1eefng nhi\u1ec5u lo\u1ea1n kh\u00f4ng m\u1ea1ch l\u1ea1c. Tuy nhi\u00ean, khi B\u1ed9 ph\u00e2n bi\u1ec7t d\u1ec5 d\u00e0ng ph\u00e1t hi\u1ec7n ra nh\u1eefng sai s\u00f3t n\u00e0y v\u00e0 t\u1eeb ch\u1ed1i c\u00e1c k\u1ebft qu\u1ea3 \u0111\u1ea7u ra, B\u1ed9 t\u1ea1o bu\u1ed9c ph\u1ea3i \u0111i\u1ec1u ch\u1ec9nh c\u00e1c tham s\u1ed1 b\u00ean trong c\u1ee7a n\u00f3 \u0111\u1ec3 t\u1ea1o ra c\u00e1c c\u1ea5u tr\u00fac thuy\u1ebft ph\u1ee5c h\u01a1n. Ng\u01b0\u1ee3c l\u1ea1i, khi B\u1ed9 t\u1ea1o tr\u1edf n\u00ean th\u00e0nh th\u1ea1o h\u01a1n trong vi\u1ec7c b\u1eaft ch\u01b0\u1edbc th\u1ef1c t\u1ebf, B\u1ed9 ph\u00e2n bi\u1ec7t ph\u1ea3i tr\u1edf n\u00ean tinh vi h\u01a1n \u0111\u1ec3 ph\u00e1t hi\u1ec7n ra nh\u1eefng \u0111i\u1ec3m kh\u00f4ng nh\u1ea5t qu\u00e1n tinh t\u1ebf. Cu\u1ed9c ch\u1ea1y \u0111ua v\u0169 trang li\u00ean t\u1ee5c n\u00e0y cu\u1ed1i c\u00f9ng cho ph\u00e9p c\u00e1c GAN t\u1ea1o ra c\u00e1c khu\u00f4n m\u1eb7t v\u00e0 v\u1eadt th\u1ec3 c\u1ef1c k\u1ef3 s\u1eafc n\u00e9t, c\u00f3 \u0111\u1ed9 ph\u00e2n gi\u1ea3i cao. M\u1eb7c d\u00f9 th\u00e0nh c\u00f4ng, c\u00e1c GAN v\u1eabn g\u1eb7p ph\u1ea3i nh\u1eefng h\u1ea1n ch\u1ebf \u0111\u00e1ng k\u1ec3, ch\u1eb3ng h\u1ea1n nh\u01b0 \"s\u1ee5p \u0111\u1ed5 ch\u1ebf \u0111\u1ed9\" (mode collapse) \u2014 m\u1ed9t ch\u1ebf \u0111\u1ed9 th\u1ea5t b\u1ea1i n\u01a1i b\u1ed9 t\u1ea1o t\u00ecm th\u1ea5y m\u1ed9t k\u1ebft qu\u1ea3 \u0111\u1ea7u ra duy nh\u1ea5t c\u00f3 th\u1ec3 \u0111\u00e1nh l\u1eeba b\u1ed9 ph\u00e2n bi\u1ec7t v\u00e0 l\u1eb7p \u0111i l\u1eb7p l\u1ea1i c\u00f9ng m\u1ed9t h\u00ecnh \u1ea3nh ch\u00ednh x\u00e1c \u0111\u00f3, l\u00e0m h\u1ea1n ch\u1ebf nghi\u00eam tr\u1ecdng s\u1ef1 \u0111a d\u1ea1ng s\u00e1ng t\u1ea1o.<\/p>\n\n\n\n

\u0110\u1ed3ng th\u1eddi, c\u00e1c nh\u00e0 nghi\u00ean c\u1ee9u \u0111\u00e3 kh\u00e1m ph\u00e1 m\u1ed9t ki\u1ebfn tr\u00fac n\u1ec1n t\u1ea3ng kh\u00e1c \u0111\u01b0\u1ee3c g\u1ecdi l\u00e0 Variational Autoencoders (VAEs). Kh\u00f4ng gi\u1ed1ng nh\u01b0 khu\u00f4n kh\u1ed5 c\u1ea1nh tranh c\u1ee7a c\u00e1c GAN, c\u00e1c VAE t\u1eadp trung m\u1ea1nh v\u00e0o vi\u1ec7c n\u00e9n v\u00e0 t\u00e1i t\u1ea1o d\u1eef li\u1ec7u. M\u1ed9t VAE bao g\u1ed3m m\u1ed9t b\u1ed9 m\u00e3 h\u00f3a (encoder) nh\u1eadn h\u00ecnh \u1ea3nh \u0111\u1ea7u v\u00e0o v\u00e0 n\u00e9n n\u00f3 th\u00e0nh m\u1ed9t bi\u1ec3u di\u1ec5n ti\u1ec1m \u1ea9n (latent representation) c\u00f3 chi\u1ec1u th\u1ea5p, hi\u1ec7u qu\u1ea3 cao, ch\u1ec9 n\u1eafm b\u1eaft c\u00e1c \u0111\u1eb7c \u0111i\u1ec3m c\u1ea5u tr\u00fac quan tr\u1ecdng nh\u1ea5t. Th\u00e0nh ph\u1ea7n th\u1ee9 hai, b\u1ed9 gi\u1ea3i m\u00e3 (decoder), sau \u0111\u00f3 l\u1ea5y bi\u1ec3u di\u1ec5n n\u00e9n n\u00e0y v\u00e0 c\u1ed1 g\u1eafng m\u1edf r\u1ed9ng n\u00f3 tr\u1edf l\u1ea1i h\u00ecnh \u1ea3nh g\u1ed1c m\u1ed9t c\u00e1ch ch\u00ednh x\u00e1c nh\u1ea5t c\u00f3 th\u1ec3. B\u1eb1ng c\u00e1ch chu\u1ea9n h\u00f3a kh\u00f4ng gian n\u00e9n n\u00e0y, c\u00e1c VAE \u0111\u1ea3m b\u1ea3o r\u1eb1ng c\u1ea3nh quan ti\u1ec1m \u1ea9n (latent landscape) l\u00e0 m\u01b0\u1ee3t m\u00e0 v\u00e0 li\u00ean t\u1ee5c, ngh\u0129a l\u00e0 n\u1ebfu b\u1ea1n ch\u1ecdn m\u1ed9t \u0111i\u1ec3m ng\u1eabu nhi\u00ean gi\u1eefa c\u00e1c t\u1ecda \u0111\u1ed9 cho m\u1ed9t \"h\u00ecnh tr\u00f2n\" v\u00e0 m\u1ed9t \"h\u00ecnh vu\u00f4ng\", b\u1ed9 gi\u1ea3i m\u00e3 s\u1ebd hi\u1ec3n th\u1ecb m\u01b0\u1ee3t m\u00e0 m\u1ed9t h\u00ecnh vu\u00f4ng bo tr\u00f2n. M\u1eb7c d\u00f9 c\u00e1c VAE cung c\u1ea5p s\u1ef1 \u1ed5n \u0111\u1ecbnh v\u00e0 t\u00ednh \u0111a d\u1ea1ng tuy\u1ec7t v\u1eddi, c\u00e1c k\u1ebft qu\u1ea3 \u0111\u1ea7u ra cu\u1ed1i c\u00f9ng c\u1ee7a ch\u00fang th\u01b0\u1eddng b\u1ecb m\u1edd \u0111\u00e1ng k\u1ec3, kh\u00f4ng th\u1ec3 n\u1eafm b\u1eaft \u0111\u01b0\u1ee3c c\u00e1c chi ti\u1ebft s\u1eafc n\u00e9t, tinh x\u1ea3o m\u00e0 ng\u01b0\u1eddi xem mong \u0111\u1ee3i t\u1eeb ngh\u1ec7 thu\u1eadt \u0111\u1ed9 trung th\u1ef1c cao.<\/p>\n\n\n\n

Si\u00eau n\u0103ng l\u1ef1c hi\u1ec7n \u0111\u1ea1i: C\u00e1c m\u00f4 h\u00ecnh khu\u1ebfch t\u00e1n v\u00e0 c\u01a1 ch\u1ebf nhi\u1ec5u<\/h2>\n\n\n\n

B\u1ed1i c\u1ea3nh hi\u1ec7n \u0111\u1ea1i c\u1ee7a vi\u1ec7c t\u1ea1o h\u00ecnh \u1ea3nh b\u1eb1ng AI \u2014 \u0111\u01b0\u1ee3c th\u1ed1ng tr\u1ecb b\u1edfi c\u00e1c h\u1ec7 th\u1ed1ng d\u1eabn \u0111\u1ea7u ng\u00e0nh nh\u01b0 Midjourney, DALL-E v\u00e0 Stable Diffusion \u2014 \u0111\u01b0\u1ee3c th\u00fac \u0111\u1ea9y b\u1edfi m\u1ed9t b\u01b0\u1edbc \u0111\u1ed9t ph\u00e1 ho\u00e0n to\u00e0n kh\u00e1c bi\u1ec7t \u0111\u01b0\u1ee3c g\u1ecdi l\u00e0 C\u00e1c m\u00f4 h\u00ecnh khu\u1ebfch t\u00e1n (Diffusion Models). L\u1ea5y c\u1ea3m h\u1ee9ng t\u1eeb c\u00e1c kh\u00e1i ni\u1ec7m trong nhi\u1ec7t \u0111\u1ed9ng l\u1ef1c h\u1ecdc kh\u00f4ng c\u00e2n b\u1eb1ng, c\u00e1c m\u00f4 h\u00ecnh khu\u1ebfch t\u00e1n \u0111\u00e3 ho\u00e0n to\u00e0n thay \u0111\u1ed5i c\u00e1c m\u00f4 h\u00ecnh t\u1ed5ng h\u1ee3p h\u00ecnh \u1ea3nh tr\u01b0\u1edbc \u0111\u00e2y. Thay v\u00ec c\u1ed1 g\u1eafng x\u00e2y d\u1ef1ng m\u1ed9t h\u00ecnh \u1ea3nh t\u1eeb \u0111\u1ea7u c\u00f9ng m\u1ed9t l\u00fac, c\u00e1c m\u00f4 h\u00ecnh n\u00e0y coi v\u1ea5n \u0111\u1ec1 l\u00e0 m\u1ed9t qu\u00e1 tr\u00ecnh thanh l\u1ecdc d\u1ea7n d\u1ea7n, h\u1ecdc c\u00e1ch t\u1ea1o ra c\u00e1c h\u00ecnh \u1ea3nh ph\u1ee9c t\u1ea1p b\u1eb1ng c\u00e1ch l\u00e0m ch\u1ee7 ngh\u1ec7 thu\u1eadt h\u1ee7y di\u1ec7t c\u00f3 ki\u1ec3m so\u00e1t v\u00e0 t\u00e1i thi\u1ebft k\u1ebf c\u00f3 h\u1ec7 th\u1ed1ng.<\/p>\n\n\n\n

C\u01a1 ch\u1ebf c\u1ee7a m\u1ed9t m\u00f4 h\u00ecnh khu\u1ebfch t\u00e1n \u0111\u01b0\u1ee3c chia th\u00e0nh hai giai \u0111o\u1ea1n ch\u00ednh: qu\u00e1 tr\u00ecnh khu\u1ebfch t\u00e1n xu\u00f4i v\u00e0 qu\u00e1 tr\u00ecnh khu\u1ebfch t\u00e1n ng\u01b0\u1ee3c. Trong qu\u00e1 tr\u00ecnh xu\u00f4i, h\u1ec7 th\u1ed1ng l\u1ea5y m\u1ed9t h\u00ecnh \u1ea3nh hu\u1ea5n luy\u1ec7n ho\u00e0n to\u00e0n r\u00f5 r\u00e0ng v\u00e0 c\u1ed1 t\u00ecnh th\u00eam c\u00e1c gia s\u1ed1 nh\u1ecf c\u1ee7a nhi\u1ec5u Gaussian qua h\u00e0ng tr\u0103m b\u01b0\u1edbc. Khi c\u00e1c b\u01b0\u1edbc ti\u1ebfn tri\u1ec3n, c\u1ea5u tr\u00fac ban \u0111\u1ea7u c\u1ee7a h\u00ecnh \u1ea3nh d\u1ea7n b\u1ecb suy gi\u1ea3m. \u0110\u1ebfn cu\u1ed1i chu\u1ed7i xu\u00f4i, h\u00ecnh \u1ea3nh b\u1ecb x\u00f3a s\u1ed5 ho\u00e0n to\u00e0n, bi\u1ebfn th\u00e0nh m\u1ed9t bi\u1ec3n nhi\u1ec5u pixel ng\u1eabu nhi\u00ean v\u00f4 ngh\u0129a, gi\u1ed1ng nh\u01b0 nhi\u1ec5u tr\u1eafng tr\u00ean m\u00e0n h\u00ecnh tivi c\u0169 kh\u00f4ng c\u00f3 t\u00edn hi\u1ec7u.<\/p>\n\n\n\n

\n

Qu\u00e1 tr\u00ecnh khu\u1ebfch t\u00e1n<\/h3>\n\n \n
\n

Khu\u1ebfch t\u00e1n xu\u00f4i<\/h4>\n
\n
H\u00ecnh \u1ea3nh r\u00f5 n\u00e9t<\/div>\n
<\/div><\/div>\n
Nhi\u1ec5u m\u1ed9t ph\u1ea7n<\/div>\n
<\/div><\/div>\n
Nhi\u1ec5u t\u0129nh to\u00e0n ph\u1ea7n<\/div>\n <\/div>\n <\/div>\n\n \n
\n

Khu\u1ebfch t\u00e1n ng\u01b0\u1ee3c<\/h4>\n
\n
Nhi\u1ec5u t\u0129nh to\u00e0n ph\u1ea7n<\/div>\n
<\/div><\/div>\n
Giai \u0111o\u1ea1n kh\u1eed nhi\u1ec5u<\/div>\n
<\/div><\/div>\n
H\u00ecnh \u1ea3nh cu\u1ed1i c\u00f9ng<\/div>\n <\/div>\n <\/div>\n<\/div>\n\n\n\n

Ph\u00e9p thu\u1eadt th\u1ef1c s\u1ef1 x\u1ea3y ra trong qu\u00e1 tr\u00ecnh khu\u1ebfch t\u00e1n ng\u01b0\u1ee3c, n\u01a1i vi\u1ec7c t\u1ea1o h\u00ecnh \u1ea3nh th\u1ef1c t\u1ebf di\u1ec5n ra. M\u1ea1ng th\u1ea7n kinh, th\u01b0\u1eddng s\u1eed d\u1ee5ng ki\u1ebfn tr\u00fac g\u1ecdi l\u00e0 U-Net, \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n \u0111\u1ec3 quan s\u00e1t m\u1ed9t h\u00ecnh \u1ea3nh ch\u1ee9a m\u1ed9t m\u1ee9c \u0111\u1ed9 nhi\u1ec5u c\u1ee5 th\u1ec3 v\u00e0 d\u1ef1 \u0111o\u00e1n ch\u00ednh x\u00e1c l\u01b0\u1ee3ng nhi\u1ec5u \u0111\u00e3 \u0111\u01b0\u1ee3c th\u00eam v\u00e0o \u1edf b\u01b0\u1edbc tr\u01b0\u1edbc. B\u1eb1ng c\u00e1ch hu\u1ea5n luy\u1ec7n m\u1ea1ng tr\u00ean h\u00e0ng t\u1ef7 v\u00ed d\u1ee5, n\u00f3 h\u1ecdc c\u00e1ch tr\u1eeb \u0111i nhi\u1ec5u \u0111\u00e3 d\u1ef1 \u0111o\u00e1n n\u00e0y v\u1edbi \u0111\u1ed9 ch\u00ednh x\u00e1c \u0111\u00e1ng kinh ng\u1ea1c. Do \u0111\u00f3, khi ng\u01b0\u1eddi d\u00f9ng y\u00eau c\u1ea7u m\u1ed9t h\u00ecnh \u1ea3nh m\u1edbi, AI b\u1eaft \u0111\u1ea7u v\u1edbi m\u1ed9t khung v\u1ebd ch\u1ee9a nhi\u1ec5u to\u00e1n h\u1ecdc ng\u1eabu nhi\u00ean, thu\u1ea7n t\u00fay. Sau \u0111\u00f3, n\u00f3 \u00e1p d\u1ee5ng m\u1ea1ng U-Net \u0111\u00e3 \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n c\u1ee7a m\u00ecnh m\u1ed9t c\u00e1ch l\u1eb7p \u0111i l\u1eb7p l\u1ea1i, lo\u1ea1i b\u1ecf d\u1ea7n c\u00e1c l\u1edbp nhi\u1ec5u t\u1eebng b\u01b0\u1edbc m\u1ed9t. V\u1edbi m\u1ed7i l\u1ea7n l\u1eb7p, nh\u1eefng h\u00ecnh d\u1ea1ng m\u01a1 h\u1ed3 b\u1eaft \u0111\u1ea7u k\u1ebft tinh t\u1eeb s\u1ef1 h\u1ed7n lo\u1ea1n, bi\u1ebfn nh\u1eefng kh\u1ed1i tr\u1eebu t\u01b0\u1ee3ng th\u00e0nh c\u00e1c \u0111\u01b0\u1eddng n\u00e9t, k\u1ebft c\u1ea5u r\u00f5 r\u00e0ng v\u00e0 cu\u1ed1i c\u00f9ng l\u00e0 m\u1ed9t h\u00ecnh \u1ea3nh cu\u1ed1i c\u00f9ng nh\u1ea5t qu\u00e1n, chi ti\u1ebft cao.<\/p>\n\n\n\n

K\u1ebft n\u1ed1i gi\u1eefa t\u1eeb ng\u1eef v\u00e0 \u0111i\u1ec3m \u1ea3nh: Vai tr\u00f2 c\u1ee7a vi\u1ec7c ti\u1ec1n hu\u1ea5n luy\u1ec7n ng\u00f4n ng\u1eef-h\u00ecnh \u1ea3nh t\u01b0\u01a1ng ph\u1ea3n (CLIP)<\/h2>\n\n\n\n

M\u1eb7c d\u00f9 c\u00e1c m\u00f4 h\u00ecnh khu\u1ebfch t\u00e1n v\u00f4 c\u00f9ng \u0111i\u00eau luy\u1ec7n trong vi\u1ec7c bi\u1ebfn nhi\u1ec5u ng\u1eabu nhi\u00ean th\u00e0nh h\u00ecnh \u1ea3nh c\u00f3 c\u1ea5u tr\u00fac, nh\u01b0ng ch\u00fang v\u1ed1n d\u0129 thi\u1ebfu kh\u1ea3 n\u0103ng hi\u1ec3u ng\u00f4n ng\u1eef ho\u1eb7c v\u0103n b\u1ea3n vi\u1ebft c\u1ee7a con ng\u01b0\u1eddi. \u0110\u1ec3 thu h\u1eb9p kho\u1ea3ng c\u00e1ch gi\u1eefa ng\u00f4n ng\u1eef con ng\u01b0\u1eddi v\u00e0 c\u00e1c \u0111i\u1ec3m \u1ea3nh tr\u1ef1c quan, c\u00e1c h\u1ec7 th\u1ed1ng t\u1ea1o h\u00ecnh \u1ea3nh hi\u1ec7n \u0111\u1ea1i d\u1ef1a v\u00e0o m\u1ed9t l\u1edbp chuy\u1ec3n \u0111\u1ed5i quan tr\u1ecdng, v\u00ed d\u1ee5 \u0111i\u1ec3n h\u00ecnh nh\u1ea5t l\u00e0 CLIP (Contrastive Language-Image Pre-training) c\u1ee7a OpenAI. N\u1ebfu kh\u00f4ng c\u00f3 m\u1ed9t c\u01a1 ch\u1ebf nh\u01b0 CLIP, m\u1ed9t m\u00f4 h\u00ecnh khu\u1ebfch t\u00e1n c\u00f3 th\u1ec3 t\u1ea1o ra c\u00e1c phong c\u1ea3nh ho\u1eb7c \u0111\u1ed1i t\u01b0\u1ee3ng \u0111\u1eb9p m\u1eaft nh\u01b0ng ng\u1eabu nhi\u00ean, m\u00e0 kh\u00f4ng c\u00f3 c\u00e1ch n\u00e0o bi\u1ebft \u0111\u01b0\u1ee3c l\u00e0m th\u1ebf n\u00e0o \u0111\u1ec3 c\u0103n ch\u1ec9nh nh\u1eefng s\u00e1ng t\u1ea1o \u0111\u00f3 v\u1edbi c\u00e1c l\u1ec7nh vi\u1ebft c\u1ee5 th\u1ec3 c\u1ee7a ng\u01b0\u1eddi d\u00f9ng.<\/p>\n\n\n\n

CLIP \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n tr\u00ean m\u1ed9t b\u1ed9 d\u1eef li\u1ec7u kh\u1ed5ng l\u1ed3 g\u1ed3m c\u00e1c c\u1eb7p h\u00ecnh \u1ea3nh-v\u0103n b\u1ea3n \u0111\u01b0\u1ee3c thu th\u1eadp t\u1eeb kh\u1eafp n\u01a1i tr\u00ean internet. M\u1ee5c ti\u00eau ch\u00ednh c\u1ee7a n\u00f3 l\u00e0 h\u1ecdc m\u1ed9t kh\u00f4ng gian nh\u00fang chung (shared embedding space) n\u01a1i m\u1ed9t m\u00f4 t\u1ea3 v\u0103n b\u1ea3n v\u00e0 h\u00ecnh \u1ea3nh t\u01b0\u01a1ng \u1ee9ng c\u1ee7a n\u00f3 \u0111\u01b0\u1ee3c \u00e1nh x\u1ea1 ch\u00ednh x\u00e1c v\u00e0o c\u00f9ng m\u1ed9t vect\u01a1 to\u00e1n h\u1ecdc. V\u00ed d\u1ee5, c\u00e2u \u201c\u0111\u01b0\u1eddng ch\u00e2n tr\u1eddi c\u1ee7a m\u1ed9t th\u00e0nh ph\u1ed1 cyberpunk t\u01b0\u01a1ng lai\u201d v\u00e0 m\u1ed9t b\u1ee9c tranh k\u1ef9 thu\u1eadt s\u1ed1 th\u1ef1c s\u1ef1 v\u1ec1 m\u1ed9t khu v\u1ef1c \u0111\u00f4 th\u1ecb neon r\u1ef1c r\u1ee1 \u0111\u01b0\u1ee3c h\u01b0\u1edbng \u0111\u1ebfn c\u00f9ng m\u1ed9t t\u1ecda \u0111\u1ed9 trong kh\u00f4ng gian \u0111a chi\u1ec1u n\u00e0y. M\u00f4 h\u00ecnh \u0111\u1ea1t \u0111\u01b0\u1ee3c \u0111i\u1ec1u n\u00e0y th\u00f4ng qua h\u1ecdc t\u01b0\u01a1ng ph\u1ea3n (contrastive learning), t\u1ed1i \u0111a h\u00f3a s\u1ef1 c\u0103n ch\u1ec9nh to\u00e1n h\u1ecdc gi\u1eefa c\u00e1c c\u1eb7p kh\u1edbp nhau trong khi gi\u1ea3m thi\u1ec3u quy\u1ebft li\u1ec7t s\u1ef1 c\u0103n ch\u1ec9nh gi\u1eefa c\u00e1c v\u0103n b\u1ea3n v\u00e0 h\u00ecnh \u1ea3nh kh\u00f4ng li\u00ean quan.<\/p>\n\n\n\n

\"M\u1ed9t
M\u1ed9t khung c\u1ea3nh gi\u1ea3 t\u01b0\u1edfng s\u1ed1ng \u0111\u1ed9ng, \u0111\u01b0\u1ee3c AI t\u1ea1o ra, v\u1edbi h\u00ecnh \u1ea3nh m\u1ed9t con ph\u01b0\u1ee3ng ho\u00e0ng r\u1ef1c r\u1ee1 \u0111ang dang r\u1ed9ng \u0111\u00f4i c\u00e1nh ph\u00eda tr\u00ean m\u1ed9t chi\u1ebfn binh \u0111\u01a1n \u0111\u1ed9c trong khu r\u1eebng huy\u1ec1n b\u00ed, th\u1ec3 hi\u1ec7n kh\u1ea3 n\u0103ng c\u1ee7a AI t\u1ea1o sinh hi\u1ec7n \u0111\u1ea1i.<\/figcaption><\/figure>\n\n\n\n

Khi ng\u01b0\u1eddi d\u00f9ng nh\u1eadp m\u1ed9t l\u1eddi nh\u1eafc (prompt) v\u00e0o m\u1ed9t tr\u00ecnh t\u1ea1o AI, v\u0103n b\u1ea3n \u0111\u00f3 ngay l\u1eadp t\u1ee9c \u0111\u01b0\u1ee3c \u0111\u01b0a v\u00e0o th\u00e0nh ph\u1ea7n m\u00e3 h\u00f3a v\u0103n b\u1ea3n c\u1ee7a m\u1ea1ng CLIP. B\u1ed9 m\u00e3 h\u00f3a n\u00e0y chuy\u1ec3n \u0111\u1ed5i c\u00e1c chu\u1ed7i t\u1eeb th\u00e0nh m\u1ed9t vect\u01a1 s\u1ed1 d\u00e0y \u0111\u1eb7c, t\u00f3m t\u1eaft \u00fd ngh\u0129a ng\u1eef ngh\u0129a c\u1ee7a y\u00eau c\u1ea7u. Vect\u01a1 v\u0103n b\u1ea3n n\u00e0y sau \u0111\u00f3 \u0111\u01b0\u1ee3c ti\u00eam v\u00e0o qu\u00e1 tr\u00ecnh khu\u1ebfch t\u00e1n ng\u01b0\u1ee3c nh\u01b0 m\u1ed9t l\u1ef1c h\u01b0\u1edbng d\u1eabn, th\u01b0\u1eddng th\u00f4ng qua m\u1ed9t c\u01a1 ch\u1ebf g\u1ecdi l\u00e0 ch\u00fa \u00fd ch\u00e9o (cross-attention). Khi ki\u1ebfn tr\u00fac U-Net ho\u1ea1t \u0111\u1ed9ng \u0111\u1ec3 lo\u1ea1i b\u1ecf nhi\u1ec5u kh\u1ecfi khung v\u1ebd t\u0129nh ban \u0111\u1ea7u, n\u00f3 li\u00ean t\u1ee5c ki\u1ec3m tra ti\u1ebfn \u0111\u1ed9 c\u1ee7a m\u00ecnh d\u1ef1a tr\u00ean vect\u01a1 v\u0103n b\u1ea3n CLIP. C\u00e1c c\u01a1 ch\u1ebf ch\u00fa \u00fd h\u01b0\u1edbng d\u1eabn qu\u00e1 tr\u00ecnh kh\u1eed nhi\u1ec5u, \u0111\u1ea3m b\u1ea3o r\u1eb1ng c\u00e1c c\u1ea5u tr\u00fac n\u1ed5i l\u00ean t\u1eeb nhi\u1ec5u t\u0129nh c\u0103n ch\u1ec9nh ch\u00ednh x\u00e1c v\u1edbi c\u00e1c kh\u00e1i ni\u1ec7m, phong c\u00e1ch v\u00e0 \u0111\u1ed1i t\u01b0\u1ee3ng \u0111\u01b0\u1ee3c y\u00eau c\u1ea7u trong l\u1eddi nh\u1eafc c\u1ee7a ng\u01b0\u1eddi d\u00f9ng.<\/p>\n\n\n\n

Khu\u1ebfch t\u00e1n ti\u1ec1m \u1ea9n v\u00e0 t\u1ed1i \u01b0u h\u00f3a: Gi\u00fap \u0111\u1ed9 ph\u00e2n gi\u1ea3i cao tr\u1edf n\u00ean d\u1ec5 ti\u1ebfp c\u1eadn<\/h2>\n\n\n\n

Trong giai \u0111o\u1ea1n \u0111\u1ea7u ph\u00e1t tri\u1ec3n c\u00e1c m\u00f4 h\u00ecnh khu\u1ebfch t\u00e1n, t\u00e0i nguy\u00ean t\u00ednh to\u00e1n c\u1ea7n thi\u1ebft \u0111\u1ec3 t\u1ea1o ra h\u00ecnh \u1ea3nh \u0111\u1ed9 ph\u00e2n gi\u1ea3i cao l\u00e0 r\u1ea5t l\u1edbn. Vi\u1ec7c x\u1eed l\u00fd t\u1eebng pixel c\u1ee7a h\u00ecnh \u1ea3nh 1024\u00d71024 qua h\u00e0ng tr\u0103m b\u01b0\u1edbc c\u1ee7a m\u1ed9t m\u1ea1ng th\u1ea7n kinh s\u00e2u y\u00eau c\u1ea7u l\u01b0\u1ee3ng VRAM kh\u1ed5ng l\u1ed3 v\u00e0 s\u1ee9c m\u1ea1nh x\u1eed l\u00fd kh\u1ee7ng khi\u1ebfp, khi\u1ebfn n\u00f3 ho\u00e0n to\u00e0n kh\u00f4ng th\u1ef1c t\u1ebf \u0111\u1ed1i v\u1edbi ph\u1ea7n c\u1ee9ng ti\u00eau d\u00f9ng ho\u1eb7c tri\u1ec3n khai r\u1ed9ng r\u00e3i. N\u00fat th\u1eaft n\u00e0y \u0111\u00e3 d\u1eabn \u0111\u1ebfn vi\u1ec7c ph\u00e1t minh ra c\u00e1c M\u00f4 h\u00ecnh Khu\u1ebfch t\u00e1n Ti\u1ec1m \u1ea9n (Latent Diffusion Models - LDM), m\u1ed9t k\u1ef9 thu\u1eadt t\u1ed1i \u01b0u h\u00f3a mang t\u00ednh c\u00e1ch m\u1ea1ng t\u1ea1o n\u00ean n\u1ec1n t\u1ea3ng cho c\u00e1c m\u00f4 h\u00ecnh ngu\u1ed3n m\u1edf nh\u01b0 Stable Diffusion. \u0110\u1ed5i m\u1edbi c\u1ed1t l\u00f5i c\u1ee7a Khu\u1ebfch t\u00e1n Ti\u1ec1m \u1ea9n l\u00e0 to\u00e0n b\u1ed9 qu\u00e1 tr\u00ecnh kh\u1eed nhi\u1ec5u kh\u00f4ng di\u1ec5n ra trong kh\u00f4ng gian pixel th\u1ef1c t\u1ebf kh\u1ed5ng l\u1ed3, nhi\u1ec1u chi\u1ec1u. Thay v\u00e0o \u0111\u00f3, h\u1ec7 th\u1ed1ng s\u1eed d\u1ee5ng m\u1ed9t b\u1ed9 t\u1ef1 m\u00e3 h\u00f3a (Autoencoder) m\u1ea1nh m\u1ebd \u0111\u1ec3 n\u00e9n h\u00ecnh \u1ea3nh ban \u0111\u1ea7u th\u00e0nh m\u1ed9t kh\u00f4ng gian ti\u1ec1m \u1ea9n nh\u1ecf h\u01a1n, c\u00f3 chi\u1ec1u th\u1ea5p h\u01a1n nhi\u1ec1u tr\u01b0\u1edbc khi b\u1ea5t k\u1ef3 s\u1ef1 khu\u1ebfch t\u00e1n n\u00e0o x\u1ea3y ra. V\u00ed d\u1ee5, m\u1ed9t h\u00ecnh \u1ea3nh th\u01b0\u1eddng bao g\u1ed3m h\u00e0ng tri\u1ec7u gi\u00e1 tr\u1ecb pixel \u0111\u1ecf, xanh l\u00e1 c\u00e2y v\u00e0 xanh d\u01b0\u01a1ng \u0111\u01b0\u1ee3c n\u00e9n th\u00e0nh m\u1ed9t bi\u1ec3u di\u1ec5n to\u00e1n h\u1ecdc nh\u1ecf g\u1ecdn ch\u1ec9 b\u1eb1ng m\u1ed9t ph\u1ea7n nh\u1ecf k\u00edch th\u01b0\u1edbc ban \u0111\u1ea7u, nh\u01b0ng v\u1eabn gi\u1eef l\u1ea1i t\u1ea5t c\u1ea3 c\u00e1c d\u1eef li\u1ec7u ng\u1eef ngh\u0129a v\u00e0 c\u1ea5u tr\u00fac thi\u1ebft y\u1ebfu.<\/p>\n\n\n\n

Khi h\u00ecnh \u1ea3nh \u0111\u01b0\u1ee3c l\u01b0u tr\u1eef an to\u00e0n trong kh\u00f4ng gian ti\u1ec1m \u1ea9n (latent space) hi\u1ec7u qu\u1ea3 n\u00e0y, c\u00e1c qu\u00e1 tr\u00ecnh khu\u1ebfch t\u00e1n xu\u00f4i v\u00e0 ng\u01b0\u1ee3c \u0111\u01b0\u1ee3c th\u1ef1c hi\u1ec7n. V\u00ec m\u1ea1ng th\u1ea7n kinh ch\u1ec9 thao t\u00e1c tr\u00ean m\u1ed9t kh\u00e1i ni\u1ec7m tr\u1eebu t\u01b0\u1ee3ng to\u00e1n h\u1ecdc \u0111\u00e3 \u0111\u01b0\u1ee3c c\u00f4 \u0111\u1ecdng cao thay v\u00ec h\u00e0ng tri\u1ec7u pixel ri\u00eang l\u1ebb, kh\u1ed1i l\u01b0\u1ee3ng c\u00f4ng vi\u1ec7c t\u00ednh to\u00e1n gi\u1ea3m \u0111\u00e1ng k\u1ec3. \u0110i\u1ec1u n\u00e0y cho ph\u00e9p m\u00f4 h\u00ecnh ch\u1ea1y hi\u1ec7u qu\u1ea3 tr\u00ean c\u00e1c card \u0111\u1ed3 h\u1ecda ti\u00eau d\u00f9ng ti\u00eau chu\u1ea9n. Khi qu\u00e1 tr\u00ecnh khu\u1ebfch t\u00e1n ng\u01b0\u1ee3c ho\u00e0n t\u1ea5t v\u00e0 nhi\u1ec5u \u0111\u00e3 \u0111\u01b0\u1ee3c lo\u1ea1i b\u1ecf th\u00e0nh c\u00f4ng trong kh\u00f4ng gian ti\u1ec1m \u1ea9n, vect\u01a1 ti\u1ec1m \u1ea9n cu\u1ed1i c\u00f9ng \u0111\u00e3 \u0111\u01b0\u1ee3c t\u1ed1i \u01b0u h\u00f3a s\u1ebd \u0111\u01b0\u1ee3c chuy\u1ec3n qua th\u00e0nh ph\u1ea7n gi\u1ea3i m\u00e3 c\u1ee7a b\u1ed9 t\u1ef1 m\u00e3 h\u00f3a (Autoencoder). B\u1ed9 gi\u1ea3i m\u00e3 d\u1ecbch c\u00e1c con s\u1ed1 tr\u1eebu t\u01b0\u1ee3ng tr\u1edf l\u1ea1i l\u0129nh v\u1ef1c pixel, ngay l\u1eadp t\u1ee9c \"ph\u00f3ng \u0111\u1ea1i\" vect\u01a1 nh\u1ecf g\u1ecdn th\u00e0nh m\u1ed9t h\u00ecnh \u1ea3nh s\u1eafc n\u00e9t, \u0111\u1ed9 ph\u00e2n gi\u1ea3i cao \u0111\u1ec3 ng\u01b0\u1eddi d\u00f9ng nh\u00ecn th\u1ea5y.<\/p>\n\n\n\n

\u0110\u1ea1o \u0111\u1ee9c, t\u00ednh hi\u1ec7n th\u1ef1c v\u00e0 t\u01b0\u01a1ng lai c\u1ee7a truy\u1ec1n th\u00f4ng t\u1ed5ng h\u1ee3p<\/h2>\n\n\n\n

Khi c\u00e1c khu\u00f4n kh\u1ed5 to\u00e1n h\u1ecdc v\u00e0 ki\u1ebfn tr\u00fac c\u1ee7a tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o t\u1ea1o sinh ng\u00e0y c\u00e0ng ho\u00e0n thi\u1ec7n, ranh gi\u1edbi ng\u0103n c\u00e1ch truy\u1ec1n th\u00f4ng t\u1ed5ng h\u1ee3p v\u1edbi th\u1ef1c t\u1ebf kh\u00e1ch quan \u0111ang nhanh ch\u00f3ng tan bi\u1ebfn. C\u00e1c nguy\u00ean t\u1eafc c\u01a1 b\u1ea3n v\u1ec1 khu\u1ebfch t\u00e1n, kh\u00f4ng gian ti\u1ec1m \u1ea9n v\u00e0 c\u0103n ch\u1ec9nh v\u0103n b\u1ea3n b\u1eb1ng s\u1ef1 ch\u00fa \u00fd ch\u00e9o \u0111\u00e3 ph\u00e1t tri\u1ec3n \u0111\u1ebfn m\u1ee9c c\u00e1c m\u00f4 h\u00ecnh AI hi\u1ec7n c\u00f3 th\u1ec3 t\u00e1i t\u1ea1o trung th\u1ef1c c\u00e1c hi\u1ec7n t\u01b0\u1ee3ng \u00e1nh s\u00e1ng ph\u1ee9c t\u1ea1p nh\u01b0 t\u00e1n x\u1ea1 d\u01b0\u1edbi b\u1ec1 m\u1eb7t, chi\u1ebfu s\u00e1ng to\u00e0n c\u1ee5c v\u00e0 \u0111\u1ed9 s\u00e2u tr\u01b0\u1eddng \u1ea3nh tinh vi. Nh\u1eefng g\u00ec b\u1eaft \u0111\u1ea7u nh\u01b0 m\u1ed9t lo\u1ea1t c\u00e1c th\u00ed nghi\u1ec7m \u0111\u1ed9 ph\u00e2n gi\u1ea3i th\u1ea5p trong c\u00e1c ph\u00f2ng th\u00ed nghi\u1ec7m h\u1ecdc thu\u1eadt \u0111\u00e3 n\u1edf r\u1ed9 th\u00e0nh m\u1ed9t cu\u1ed9c c\u00e1ch m\u1ea1ng c\u00f4ng nghi\u1ec7p t\u00e1c \u0111\u1ed9ng \u0111\u1ebfn thi\u1ebft k\u1ebf \u0111\u1ed3 h\u1ecda, l\u00e0m phim, ki\u1ebfn tr\u00fac v\u00e0 ph\u00e1t tri\u1ec3n tr\u00f2 ch\u01a1i \u0111i\u1ec7n t\u1eed.<\/p>\n\n\n\n

Tuy nhi\u00ean, s\u1ee9c m\u1ea1nh to l\u1edbn c\u1ee7a c\u00e1c nguy\u00ean t\u1eafc c\u01a1 b\u1ea3n n\u00e0y c\u0169ng mang l\u1ea1i nh\u1eefng c\u00e2n nh\u1eafc \u0111\u00e1ng k\u1ec3 v\u1ec1 m\u1eb7t x\u00e3 h\u1ed9i v\u00e0 \u0111\u1ea1o \u0111\u1ee9c. V\u00ec c\u00e1c m\u1ea1ng n\u00e0y h\u1ecdc b\u1eb1ng c\u00e1ch t\u00ecm c\u00e1c m\u1eabu th\u1ed1ng k\u00ea trong c\u00e1c t\u1eadp d\u1eef li\u1ec7u do con ng\u01b0\u1eddi t\u1ea1o ra, ch\u00fang d\u1ec5 d\u00e0ng h\u1ea5p th\u1ee5 v\u00e0 khu\u1ebfch \u0111\u1ea1i b\u1ea5t k\u1ef3 th\u00e0nh ki\u1ebfn x\u00e3 h\u1ed9i, \u0111\u1ecbnh ki\u1ebfn ho\u1eb7c s\u1ef1 thi\u1ebfu ch\u00ednh x\u00e1c v\u1ec1 l\u1ecbch s\u1eed n\u00e0o c\u00f3 trong d\u1eef li\u1ec7u \u0111\u00e0o t\u1ea1o c\u1ee7a ch\u00fang. H\u01a1n n\u1eefa, s\u1ef1 d\u1ec5 d\u00e0ng m\u00e0 c\u00e1c m\u00f4 h\u00ecnh n\u00e0y c\u00f3 th\u1ec3 thao t\u00fang qu\u00e1 tr\u00ecnh khu\u1ebfch t\u00e1n ng\u01b0\u1ee3c \u0111\u1ec3 t\u1ea1o ra c\u00e1c m\u00f4 t\u1ea3 t\u1ed5ng h\u1ee3p ho\u00e0n h\u1ea3o v\u1ec1 con ng\u01b0\u1eddi th\u1ef1c g\u00e2y ra nh\u1eefng lo ng\u1ea1i s\u00e2u s\u1eafc v\u1ec1 t\u00ednh x\u00e1c th\u1ef1c k\u1ef9 thu\u1eadt s\u1ed1, th\u00f4ng tin sai l\u1ec7ch, quy\u1ec1n s\u1edf h\u1eefu tr\u00ed tu\u1ec7 v\u00e0 s\u1ef1 x\u00f3i m\u00f2n ni\u1ec1m tin t\u1ed5ng th\u1ec3 v\u00e0o ph\u01b0\u01a1ng ti\u1ec7n truy\u1ec1n th\u00f4ng h\u00ecnh \u1ea3nh. Nh\u00ecn v\u1ec1 ph\u00eda tr\u01b0\u1edbc, s\u1ef1 ph\u00e1t tri\u1ec3n c\u1ee7a vi\u1ec7c t\u1ea1o h\u00ecnh \u1ea3nh b\u1eb1ng AI \u0111ang chuy\u1ec3n d\u1ecbch kh\u1ecfi h\u00ecnh \u1ea3nh 2D t\u0129nh v\u00e0 m\u1edf r\u1ed9ng linh ho\u1ea1t v\u00e0o c\u00e1c kh\u00f4ng gian \u0111a chi\u1ec1u. C\u00e1c nguy\u00ean t\u1eafc c\u01a1 b\u1ea3n t\u01b0\u01a1ng t\u1ef1 c\u1ee7a vi\u1ec7c t\u1ed5ng h\u1ee3p v\u0103n b\u1ea3n th\u00e0nh h\u00ecnh \u1ea3nh hi\u1ec7n \u0111ang \u0111\u01b0\u1ee3c \u0111i\u1ec1u ch\u1ec9nh \u0111\u1ec3 th\u00fac \u0111\u1ea9y c\u00e1c ki\u1ebfn tr\u00fac v\u0103n b\u1ea3n th\u00e0nh video ti\u00ean ti\u1ebfn, t\u1ea1o t\u00e0i s\u1ea3n 3D t\u1ef1 \u0111\u1ed9ng v\u00e0 c\u00e1c m\u00f4i tr\u01b0\u1eddng \u1ea3o t\u01b0\u01a1ng t\u00e1c. B\u1eb1ng c\u00e1ch x\u1eed l\u00fd th\u1eddi gian v\u00e0 \u0111\u1ed9 s\u00e2u nh\u01b0 c\u00e1c chi\u1ec1u to\u00e1n h\u1ecdc b\u1ed5 sung trong kh\u00f4ng gian ti\u1ec1m \u1ea9n, c\u00e1c m\u1ea1ng th\u1ea7n kinh \u0111ang h\u1ecdc c\u00e1ch duy tr\u00ec s\u1ef1 nh\u1ea5t qu\u00e1n v\u1ec1 c\u1ea5u tr\u00fac v\u00e0 th\u1eddi gian gi\u1eefa c\u00e1c khung h\u00ecnh. Khi hi\u1ec7u su\u1ea5t t\u00ednh to\u00e1n t\u0103ng l\u00ean v\u00e0 c\u00e1c ki\u1ebfn tr\u00fac thu\u1eadt to\u00e1n tr\u1edf n\u00ean tinh vi h\u01a1n, h\u00e0nh tr\u00ecnh t\u1eeb m\u1ed9t tia s\u00e1ng \u0111\u01a1n gi\u1ea3n c\u1ee7a tr\u00ed t\u01b0\u1edfng t\u01b0\u1ee3ng con ng\u01b0\u1eddi \u0111\u1ebfn m\u1ed9t th\u1ef1c t\u1ebf k\u1ef9 thu\u1eadt s\u1ed1 si\u00eau th\u1ef1c ho\u00e0n ch\u1ec9nh s\u1ebd ti\u1ebfp t\u1ee5c thu h\u1eb9p, l\u00e0m thay \u0111\u1ed5i m\u00e3i m\u00e3i b\u1ed1i c\u1ea3nh c\u1ee7a s\u1ef1 s\u00e1ng t\u1ea1o con ng\u01b0\u1eddi v\u00e0 bi\u1ec3u hi\u1ec7n c\u00f4ng ngh\u1ec7.<\/p>\n\n\n\n

<\/p>","protected":false},"excerpt":{"rendered":"

The ability of artificial intelligence to generate photorealistic imagery, intricate artwork, and complex visual designs from simple textual prompts stands as one of the most remarkable technological leaps of the twenty-first century. For decades, computer graphics relied strictly on procedural algorithms, geometric modeling, and manual manipulation by human artists. Today, generative AI models can synthesize […]<\/p>\n","protected":false},"author":1,"featured_media":1148,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_seopress_titles_title":"How Does AI Create Images? A Simple Guide to Generative Art","_seopress_titles_desc":"Curious about how AI turns text into stunning images? Learn how diffusion models, latent space, and AI \"translators\" work together to turn your prompts into reality.","_seopress_robots_index":"","_seopress_robots_follow":"","_seopress_robots_imageindex":"","_seopress_robots_snippet":"","_seopress_robots_primary_cat":"","_seopress_robots_breadcrumbs":"","_seopress_robots_freeze_modified_date":"","_seopress_robots_custom_modified_date":"","_seopress_robots_canonical":"","_seopress_social_fb_title":"","_seopress_social_fb_desc":"","_seopress_social_fb_img":"","_seopress_social_fb_img_attachment_id":0,"_seopress_social_fb_img_width":0,"_seopress_social_fb_img_height":0,"_seopress_social_twitter_title":"","_seopress_social_twitter_desc":"","_seopress_social_twitter_img":"","_seopress_social_twitter_img_attachment_id":0,"_seopress_social_twitter_img_width":0,"_seopress_social_twitter_img_height":0,"_seopress_redirections_value":"","_seopress_redirections_enabled":"","_seopress_redirections_enabled_regex":"","_seopress_redirections_logged_status":"","_seopress_redirections_param":"","_seopress_redirections_type":0,"_seopress_analysis_target_kw":"","footnotes":""},"categories":[1],"tags":[],"class_list":["post-1146","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-post"],"_links":{"self":[{"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/posts\/1146","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/comments?post=1146"}],"version-history":[{"count":4,"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/posts\/1146\/revisions"}],"predecessor-version":[{"id":1153,"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/posts\/1146\/revisions\/1153"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/media\/1148"}],"wp:attachment":[{"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/media?parent=1146"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/categories?post=1146"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/imgedits.net\/vi\/wp-json\/wp\/v2\/tags?post=1146"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}