Maswali yanayounganishwa na mwongozo · Ngumu Kiwango

Maswali ya Kuhariri ya Uhakika-kwa-Uhakika

Gundua jinsi uhariri wa Upesi-kwa-Uarifu ulivyozalisha picha kupitia udhibiti wa umakini, uchanganyaji uliojanibishwa, na ubadilishaji wa picha halisi.

Njia za mwongozo zinazohusianaUliza Kuhimiza Uhariri wa Usikivu Mtambuka
Swali 1 ya 8

Je, ni taarifa gani ya ndani ambayo Prompt-to-Prompt hutumia tena kuhifadhi tukio?

Hakiki kila swali8 maswali

Majibu na maelezo huonekana unapojibu maswali.

  1. Je, ni taarifa gani ya ndani ambayo Prompt-to-Prompt hutumia tena kuhifadhi tukio?
    • Ramani za umakinifu zinazounganisha maneno na maeneo ya picha
    • Saizi za mwisho za pato
    • Hali ya kiboreshaji cha mfano
    • Metadata ya mbano ya picha
  2. Je, Uhakika wa Kuagiza unahitaji kusawazisha muundo?
    • Ndiyo, mafunzo kamili kwa kila hariri
    • Hapana, ni mbinu isiyo na mafunzo
    • Ndiyo, lakini kisimbaji maandishi pekee
    • Kwa kuongeza maneno mapya tu
  3. Kubadilisha 'baiskeli' kwa 'pikipiki' wakati wa kuweka tukio, mbinu:
    • Hufunika eneo la baiskeli kwa mkono
    • Hutoa mafunzo kwa LoRA mpya kwa pikipiki
    • Huweka umakini kati ya tokeni zinazolingana katika vidokezo viwili
    • Inazalisha picha ya kujitegemea kabisa
  4. Je, kuweka upya uzani wa ishara kunafanikisha nini?
    • Inabadilisha azimio la picha
    • Inafuta ishara kabisa
    • Hupanga upya maneno katika kidokezo
    • Huimarisha au kudhoofisha ushawishi wa kuona wa neno hilo
  5. Unapoongeza neno kwenye kidokezo, haraka-kwa-Ufafanuzi:
    • Huhifadhi umakini kwa ishara ambazo hazijabadilika na huruhusu tu ishara mpya kuunda umakini mpya
    • Hubadilisha upya picha nzima
    • Hupuuza neno lililoongezwa
    • Hubadilisha hadi muundo tofauti wa msingi
  6. Ni modeli gani tofauti iliyofunzwa kufuata maagizo ya maandishi ya uhariri wa picha kutoka kwa picha ya ingizo?
    • InstructPix2Pix, a conditional diffusion editor trained on generated instruction-edit pairs
    • Prompt-to-Prompt, which edits by controlling cross-attention maps during denoising
    • Null-text inversion, which optimizes unconditional embeddings for image inversion
    • LocalBlend, which creates a selected-token mask from attention maps
  7. Je, kipengele cha Prompt-to-Prompt LocalBlend kinatumia nini kuleta mabadiliko?
    • A hand-drawn mask required for every Prompt-to-Prompt edit
    • The EXIF metadata of the original image
    • Attention maps for selected prompt words to construct a spatial mask
    • A classifier that identifies image copyright
  8. What extra step can enable Prompt-to-Prompt editing of a real photograph?
    • Train a new diffusion model from scratch
    • Convert the photograph into an audio spectrogram
    • Replace every cross-attention layer with a classifier
    • Invert the image into the diffusion model’s latent trajectory first