Ikibazo gihuza ibibazo · Hagati Urwego

Guhindura Ishusho Guhindura hamwe na Modimodal Models Ikibazo

Gerageza uburyo bwo kuganira kuri multimodal itandukanye no gushushanya gushingiye kubishushanyo mbonera, kuki ibisobanuro bigenda bisimburana, hamwe nuburyo bufatika bwo kubigabanya.

Ikibazo 1 Bya 8

Ni irihe tandukaniro nyamukuru riri hagati yo gukwirakwiza ibishushanyo mbonera no gutunganya ibiganiro bya multimodal kavukire?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Ni irihe tandukaniro nyamukuru riri hagati yo gukwirakwiza ibishushanyo mbonera no gutunganya ibiganiro bya multimodal kavukire?
    • Inpainting only works on black-and-white images
    • Inpainting regenerates only a masked region, while multimodal models usually generate a whole new image from the conversation
    • Multimodal models require you to paint a mask first
    • Inpainting cannot use a text prompt
  2. Ni ukubera iki ibisobanuro utigeze usaba guhindura shift mugihe cyo guhindura ibintu byinshi?
    • The model deliberately adds random changes
    • The chat interface compresses text
    • The original file is deleted after upload
    • The model typically generates a whole new image, so preservation is learned rather than guaranteed
  3. Niki kugera kumateka y'ibiganiro reka reka moderi ya multimodal ikora?
    • Follow requests that refer to earlier turns, such as returning to a previous color
    • Guarantee pixel-identical output
    • Edit the layers of your original file
    • Increase the image resolution automatically
  4. Nubuhe buryo 2022 bwubushakashatsi bwerekanye ko amashusho ashobora guhindurwa uhereye kumabwiriza yanditse adafite mask?
    • SynthID
    • C2PA
    • InstructPix2Pix
    • Nano Banana
  5. Niyihe ntambwe nziza mugihe drift yubatse hejuru yibiganiro byinshi?
    • Keep adding more instructions in the same chat indefinitely
    • Start a fresh chat from the best version or refer back to the original upload
    • Lower the image to a thumbnail
    • Switch to black and white
  6. A retoucher needs every untouched pixel to stay exactly the same. Which approach fits best?
    • Asking a chat model to 'change nothing else'
    • Running the edit many times and picking one at random
    • Generating a new image from scratch
    • Mask-based inpainting or compositing the edited region back over the original
  7. Which provenance signal does Google add to images from its models?
    • An invisible SynthID watermark
    • A visible logo in every corner
    • A QR code
    • No signal of any kind
  8. Which statement about conversational editing is a misconception?
    • The output may not match your original resolution
    • Preservation of unchanged areas is learned, not guaranteed
    • The model edits your original file and its layers directly
    • Stating what must stay the same can help reduce drift