Question 1

Is this feature production-ready?

Accepted Answer

This feature is currently in beta. Output quality can vary depending on the reference images and prompt. We recommend testing before using it in critical workflows. Credits are charged per generation regardless of output quality, as the generation process itself runs to completion.

Question 2

Can I use photos of real people or celebrities?

Accepted Answer

No. This feature is designed for fictional, original, and illustrated characters only. Using photos of real people, public figures, or celebrities is not supported and may produce poor results. Such use may also raise legal and ethical concerns.

Question 3

What kind of reference images work best?

Accepted Answer

Clear, well-lit images of your original or illustrated character with a simple background. The character should be fully visible with no heavy obstructions. Illustrated or stylized art often works well.

Question 4

Why didn't my character look accurate in the video?

Accepted Answer

Character accuracy depends on the quality of your reference images and how specifically you describe the outfit. This is an experimental feature and exact likeness is not guaranteed. Try using a clearer reference image or a more detailed outfit description.

Question 5

How do I use the character_id in video generation?

Accepted Answer

Pass the `request_id` returned from a completed `sd-2-character` job as the `character_id` field in `/sd-2.0-omni-reference`. Your character will be used in the video generation.

Question 6

How many reference images should I use?

Accepted Answer

1 clear image is enough for most cases. You can upload up to 3 images for more detail. Quality and clarity matter more than quantity.

Question 7

Can I reuse the same character_id multiple times?

Accepted Answer

Yes. Once created, a character_id can be used in as many video generations as you like at no additional charge.

Question 8

Are credits refunded if the output quality is poor?

Accepted Answer

Credits are charged when the generation completes, regardless of output quality. This is consistent with how all AI generation models work — the compute cost is incurred whether or not the result matches your expectations. We recommend testing with a low-stakes prompt first.