Sɛnea mpɔtam hɔ akyi a woyi fi hɔ yɛ adwuma
Eyi ne bathroom-diagram version — a ɛdɔɔso sɛ wobɛgye pipeline no adi, ɛnyɛ krataa.
1. Input no bɛyɛ piksel
Wode fael anaa bitmap (drop, picker, paste, anaa krataafa capture) ma. Editor no decode no kɔ mfonini bitmap so. Formats te sɛ HEIC hia decode anammɔn foforo; huammɔdi gyae ntɛm denam mfomso bi so sen sɛ ɛbɛyɛ canvas a ɛnyɛ hwee.
2. Nhwɛsode a ɛyɛ ahoboa
Inference de on-device segmentation model (ISNet-class ONNX) a wɔde Web Worker ne browser ONNX runtime di dwuma. Sɛ wɔannya nyɛɛ weights no cached a, wubehu download splash. Caching kwati sɛ wobɛsan ayɛ saa hit no bere biara.
3. Nkyekyɛmu → akatasoɔ
Odwumayɛni no san de alpha mask (nea ɛyɛ asɛmti vs akyi). UI no boaboa asɛmti no ano wɔ checkerboard anaa akyi a woapaw (kɔla, gradient, mfonini) so.
4. Onipa refine
Add / Erase brushes sesa adwuma mask no wɔ CPU canvas no so. Dwodwo sesa sɛnea ɔyare no si fam denneennen no. Export de mask a wohu no di dwuma — ɛnyɛ automatic-only snapshot a wo werɛ afi.
5. Abakɔsɛm ne nneɛma a wɔde kɔ amannɔne
IndexedDB abakɔsɛm tiawa betumi akora nnansa yi mfitiase/nea efii mu bae abien. Download na Copy emit PNG a mprempren transparency anaa hyɛ no ma.
Performance notes wobɛhyɛ no nsow
- a edi kan run: network + init ka ma model no.
- mfonini akɛse: betumi ayɛ adwuma wɔ emu nsusuwii a wɔatew so a kɔkɔbɔ ka ho; export da so ara de n’ani si ne kɛse nyinaa so.
- UI thread tra hɔ a wotumi de di dwuma efisɛ inference a emu yɛ duru wɔ odwumayɛni mu — sɛ odwumayɛni no hwe ase a, wubenya odwumayɛni mfomso, ɛnyɛ tab a wɔahyɛ no nwini daa.
Wɔakata kokoam nsɛm a ɛfa “mpɔtam hɔ” ho wɔ Privacy so. Da biara da kliki: Fi ase.