Idanwo ti o sopọ mọ itọsọna · Lile Ipele

Ìdánwò Ṣiṣatunṣe Ifarabalẹ Agbelebu Tọ-si-Kia

Explore how Prompt-to-Prompt edits generated images through cross-attention control, localized blending, and inversion for real photos.

Jẹmọ itọsọna awọn ọnaTọ Lati Tọ Ifarabalẹ Agbelebu Ṣatunkọ
Ibeere 1 ti 8

Alaye inu wo ni Tọ-si-Tara lo tun lo lati ṣe itọju iṣẹlẹ kan?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Alaye inu wo ni Tọ-si-Tara lo tun lo lati ṣe itọju iṣẹlẹ kan?
    • Awọn maapu akiyesi agbelebu ti o so awọn ọrọ pọ si awọn agbegbe aworan
    • Ik o wu awọn piksẹli
    • Ipo optimizer awoṣe
    • Metadata funmorawon aworan naa
  2. Ṣe Tọ-si-Tara nilo atunṣe awoṣe daradara bi?
    • Bẹẹni, atunṣe kikun fun atunṣe
    • Rara, o jẹ ilana ti ko ni ikẹkọ
    • Bẹẹni, ṣugbọn koodu koodu nikan
    • Nikan fun fifi awọn ọrọ titun kun
  3. Lati paarọ 'keke' kan fun 'alupupu' lakoko ti o tọju ibi iṣẹlẹ, ọna naa:
    • Boju agbegbe keke pẹlu ọwọ
    • Ṣe ikẹkọ LoRA tuntun fun awọn alupupu
    • Ṣe akiyesi maapu laarin awọn ami ti o baamu kọja awọn itọka meji naa
    • Ṣe ina aworan ominira patapata
  4. Kí ni àtúnyẹ̀wò àfiyèsí àmi ń ṣe?
    • O ṣe ayipada ipinnu aworan
    • O npa ami naa kuro patapata
    • O ṣe atunṣe awọn ọrọ inu itọka naa
    • Ó ń fúnni lókun tàbí kí agbára ìríran ti ọ̀rọ̀ náà di aláìlágbára
  5. Nigbati o ba ṣafikun ọrọ kan si tọ, Tọ-si-Tara:
    • Ṣe itọju akiyesi fun awọn ami-ami ti ko yipada ati jẹ ki aami tuntun nikan ṣe agbekalẹ akiyesi tuntun
    • Tun-randomizes gbogbo aworan
    • Fojusi ọrọ ti a ṣafikun
    • Yipada si awoṣe ipilẹ ti o yatọ
  6. Which separate model was trained to follow written image-editing instructions from an input image?
    • 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. What does the Prompt-to-Prompt LocalBlend feature use to localize a change?
    • 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