Nea on-device cutout model hu

By Technical specialist

On-device cutout model nhu onipa, mpaboa, afieboa, anaa nkongua wɔ onipa ntease mu. Ɛnya mfonini no ho akontaahyɛde na ɛbu akontaa, piksel biara, mmeae bɛn na ɛda adi sɛ ɛyɛ asɛmti titiriw no dea. Nea mfaso wɔ so a efi mu ba ne akataso: asase mfonini a ɛyɛ grayscale a ebetumi abɛyɛ nea efi mu ba no kwan a ɛda adi pefee.

Saa nsonsonoe no kyerɛkyerɛ ahoɔhare a wɔde twa nneɛma a ɛnyɛ adwuma ne nea ɛwɔ anohyeto nyinaa mu. Nhwɛsode no betumi de nhwɛso ahorow a wɔde aniwa hu a woasua bere a wɔretete no adi dwuma, nanso ennim w’adwene. Ebia ade a ano yɛ nnam a ɛbɛn mfinimfini no bɛyɛ te sɛ nea ɛho hia; ahwehwɛ a ɛyɛ mmerɛw anaa ti nhwi a ayɛ mmerɛw betumi ayɛ te sɛ akyi.

Efi mfonini fael so kɔ model input so

Ɛsɛ sɛ wo decode fael a wudi kan paw no mu. Wɔmfa mfonini a wɔahyɛ no den nkɔ model no mu tẽẽ sɛ fael baiti ahorow a wɔabɔ ho ban. Browser no dan no piksel a ne tɛtrɛtɛ, ne sorokɔ, ne kɔla botae ahorow wom.

Afei inference pipeline no siesie model-sized input. Ɛgyina model ne runtime so no, ahosiesie betumi ayɛ:

• sesa decoded mfonini no kɛseɛ ma ɛyɛ dimensions a model no hwɛ kwan;
• a wohyehy3 kɔkɔɔ, ahabammono, ne bruu botaeɛ sɛ akontabuo akwan;
• scaling anaase normalizing saa values ​​no;
• packaging won mu tensor, array format a inference runtime no de di dwuma.

Nsusuwii pɔtee ne normalization mmara no yɛ model implementation no dea. Asɛm a ɛho hia a mfaso wɔ so ne sɛ nsusuwii taa yɛ adwuma fi gyinabea a wɔasiesie mu, na ɛnyɛ fibea piksel biara a ɛwɔ ne kɛse a edi kan no mu.

Easy Background Remover de on-device segmentation model na ɛyɛ saa inference anammɔn yi wɔ browser no mu. Ebia ɛho behia sɛ wɔtwe model fael no ansa na wɔatwa no kan, nanso wɔayɛ mfonini-dwumadi kwan no sɛnea ɛbɛyɛ a ɛbɛkɔ so atra hɔ bere a model no ayɛ krado no. Model download no yɛ agyapadeɛ a wɔbisa; ɛnyɛ ade koro ne mfonini a woapaw no a wode bɛto so ama akyirikyiri nsusuwii.

Nea efi mu ba no yɛ akataso a ɛte sɛ ahotoso

Wɔ beae biara a ɛwɔ nea efi mu ba no mu no, nhwɛso no ma wonya botae bi a wobetumi akyerɛ ase sɛ ɛte sɛ asɛmti kɛse anaasɛ ɛte sɛ akyigyina kɛse. Mfaso wɔ so sɛ wobɛfa eyi ho mfonini sɛ mfonini a ɛyɛ fitaa:

• fitaa kyere se fa saa bea yi sie;
• tuntum kyere se yi bea yi;
• grey kyerε sε εkata fã bi so anaase εntumi nsi pi a εbɛn anoano bi.

Browser no dane saa output no kɔ alpha mask a ɛreyɛ adwuma no mu. Afei wobetumi ayɛ mfitiase kɔla piksel ahorow no wɔ checkerboard, solid fill, gradient, anaa mfonini foforo so. Export a ɛda adi pefee de mask no sie sɛ alpha sen sɛ ɛbɛ paint checkerboard no akɔ fael no mu.

Gyinapɛn ahorow a ɛyɛ mmerɛw no ho hia. Sɛ wɔhyɛ akataso bo biara ma ɛyɛ tuntum anaa fitaa kronkron ntɛm ara a, anka ti nhwi, aboa nhoma, ntama nhama, kankyee a ɛyɛ mmerɛw, ne anoano a wɔantwe adwene asi so no bɛyɛ den wɔ ɔkwan a ɛnyɛ awosu so. Alpha fã bi betumi akora nsakrae a ɛkɔ so yiye so. Ebetumi nso ada kɔla fringe a efi akyi dedaw no adi, na ɛno nti na nea efi mu ba no da so ara fata sɛ wɔde aniwa hwɛ no.

Adanse bɛn na aniwa hu na ɛboa nhwɛsode no

Nkyekyɛmu nhwɛso betumi atwe adwene asi nsɛnkyerɛnne bɛboro biako so prɛko pɛ. Nsɛnkyerɛnne a emu yɛ den taa bi ne:

• nsonsonoe a ɛda asɛmti ne akyigyina ntam;
• ahye a etoa so twa ade a edi so ho hyia;
• a wonim no yiye a woaka abom a efa nsusuie, nwene, ne gyinabea ho;
• a emu dɔ te sɛ nsonsonoeɛ a ɛwɔ adwene a wɔde si biribi so;
• mpɔtam a wɔde aniwa hu no yɛ biako dea.

Saa nsɛnkyerɛnne yi mu biara nni hɔ a ɛyɛ awerɛhyem. Jacket tuntum a wɔde abɔ ɔfasu a ɛyɛ tumm no ma ɔhye so nsonsonoe a ɛyɛ mmerɛw. Ebia nkongua a ntam yɛ teateaa no wɔ akyi mmeae a ɛte sɛ ade no. Akyi mfonini a wɔayɛ ho nsusuwso betumi ne asɛmti ankasa no asi akan ama adwene a wɔde besi so.

Model no reyɛ akontaabu a eye sen biara afi piksel ahorow a ɛwɔ hɔ no mu. Entumi nsan nnya nsɛm a ɛkɔ akyiri a mfonini no ankyere da.

Nea model no nnim

Model no nhu ade a na wopɛ sɛ wode sie bere a nsɛmti pii wɔ hɔ no ara kwa. Ennim sɛ ahwehwɛ a wɔnom a ɛda adi pefee ho hia wɔ aguadi mu, ɛsɛ sɛ sunsuma tra hɔ ma ɛyɛ nokware, anaasɛ abɔso ketewaa bi yɛ bag dea.

Ɛnnya fibea a ɛyɛ blurry a ɛyɛ fɛ a ahintaw nso. Compression blocks, clipped highlights, motion blur, ne nsɛm nketenkete paa to adanse a ɛwɔ hɔ no ano hye. Sɛ woma preview zoom no kɔ soro a, ɛma piksel yɛ akɛse wɔ screen so; ɛmma nsɛm a ɛyera mma.

Eyi nti na “AI hui” ne “edge no ayɛ krado sɛ wobetintim” yɛ nsɛm a ɛsono no. Detection fi ase yɛ edit no. Review wie no.

Nea enti a nsɛm a ɛyɛ tratraa na ɛda adi pefee no yɛ den

Nneɛma a ɛyɛ tratraa gye piksel kakraa bi. Sɛ wɔsesa mfonini bi kɛse de yɛ nsusuwii a, wobetumi de waya anaa ti nhwi a ɛyɛ piksel biako ayɛ nkyɛmu a akyi atwa ho ahyia. Nneɛma foforo nti na ɛyɛ den sɛ nneɛma a ɛda adi pefee: wɔn kɔla a wotumi hu no fã bi yɛ akyi dedaw no. Ebia ɔhye no mu da hɔ ama obi a ɔte ade no ase, bere a piksel nhyehyɛe no da so ara yɛ nea emu nna hɔ.

Sunsuma a ɛyɛ mmerɛw ma adwenem naayɛ a wɔahyɛ da ayɛ. Wɔn akyi ano no yera nkakrankakra, na biribiara nni hɔ a wɔatwa no yiye wɔ botae mu. Sunsuma a ɛyɛ ma a wobɛma atra hɔ no betumi ayɛ te sɛ nea ɛyɛ abɔde wɔ akyi a ɛte saa ara nanso ɛho agu fĩ wɔ ade foforo a ɛhyerɛn a wɔde ahyɛ mu ma so. Sɛ woyi fi hɔ a, ebia ɛbɛma asɛm no ayɛ te sɛ nea ɛrehuruhuruw.

Eyinom yɛ atemmu ho haw ahorow te sɛ nhwɛso ho haw ahorow no.

Model mask, adwuma mask, ne export mask

Ɛboa ma wɔtetew akwan abiɛsa mu:

1. Model mask: automatic akontabuo a wɔde san ba denam inference so.
2. Adwuma akatasoɔ: akatasoɔ a wɔakyerɛ wɔ editor no mu wɔ Add anaa Erase refinement biara akyi.
3. Export mask: adwumayɛ mask a wɔde di dwuma bere a wɔatwe mprempren aba no anaasɛ wɔayɛ ho mfonini no.

Wɔ Easy Background Remover mu no, brush refinement sesa adwuma mask no. Export de mprempren aba a wɔasesa no di dwuma sen sɛ ɛbɛsan akɔ komm akɔ automatic mask a wɔanka no so. Eyi ma model no di mpaapaemu a ɛtrɛw no ho dwuma bere a woteɛteɛ nsɛm a egyina adwene so no.

Wobetumi ayɛ preview a wɔada no adi no kɛse ma ɛne editor no ahyia. Saa ɔyɛkyerɛ kɛse no nyɛ ade a wotumi de ho to so a wɔde susuw akataso no mu nsɛm nketenkete. Kenkan Mask resolution vs preview size ma nsonsonoe no.

Local inference da so ara wɔ nhyehyɛe anammɔn

Sɛ wode model bi tu mmirika wɔ mfiri so a, ɛhwehwɛ sɛ model no mu duru ne browser runtime a ɛne no hyia. Nkaribo no sõ sen interface code a ɛyɛ mmerɛw koraa, enti sɛ wode biribiara bɛhyɛ editor launch biara so a, ɛbɛyɛ ɔsɛe. Ebia nhyiam a edi kan bi atwe model no na wɔde asie; akyiri yi nhyiam ahorow a wɔyɛ no daa no taa betumi asan de adi dwuma bio.

Dɛn nti na edi kan run downloads a model kyerɛkyerɛ saa hye no mu a ɛnfrafra model agyapade download ne mfonini upload. Heavy inference nso guan fi interface thread titiriw no so; hwɛ Web Worker mfonini dwumadie mfitiaseɛ.

Ɔkwan a mfaso wɔ so a wɔfa so bu nea ebefi mu aba no atɛn

Mma nhwehwɛ asɛm no mfinimfini nkutoo mu. Hwɛ akatasoɔ no wɔ baabi a adanseɛ yɛ mmerɛ paa:

• ti nhwi, aboa nhoma, nwi, ne ntama nhama;
• ahwehwɛ, nhama, nhama, ne nhama teateaa;
• a ewo nsensanee a esua;
• a ewo semi-transparent nneɛma;
• a wode di nkitaho sunsuma;
• ntwea so a ebia akyi kɔla dedaw no bɛtena hɔ.

Hwɛ nea ebefi mu aba no wɔ akyi a ɛyɛ hann ne esum nyinaa so sɛ wonnim baabi a etwa to a wobɛkɔ a. Halo a ɛyɛ mmerɛw betumi ayera wɔ fitaa so na ada adi wɔ tuntum so. Afei siesie mmeae a ɛka sɛnea wɔpɛ sɛ wɔde bedi dwuma no nkutoo.

Automatic segmentation so wɔ mfaso kɛse bere a wɔfa no sɛ ɛyɛ nea edi kan a ɛyɛ den, na ɛnyɛ ntease a edi mũ ho asɛm. Sɛ wopɛ sɛ wosɔ mpɔtam hɔ cutout ne refinement flow no nyinaa hwɛ a, hwɛ Get started.