Implement computer vision solutions 3.1 Design and implement image- and video-generation solutions
You use the Image Edit API with a gpt-image-1 deployment (limited access preview) to…
Study guide skill: Select and apply appropriate generation and editing controls provided by the platform
Question B2-022
You use the Image Edit API with a gpt-image-1 deployment (limited access preview) to restyle a portrait photo as a watercolor painting. The people's faces must stay recognizable. Which parameter should you set?
Select one answer.
Answer: B
The input_fidelity parameter controls how much effort the model puts into matching the style and features, especially facial features, of the input images, and high input fidelity preserves faces more accurately than standard mode. It isn't supported by gpt-image-1-mini. Quality, streaming partial images, and transparent backgrounds don't control how closely faces are preserved.
Deep dive
How it works
Restyling a portrait is an image edit: you send the photo plus a prompt, and the model returns a new image. The risk is that the model "re-imagines" the people. The parameter built for this is input_fidelity.
Per the image generation page, input_fidelity controls how much effort the model puts into matching the style and features, especially facial features, of the input images. It lets you make subtle edits without changing unrelated areas. With high input fidelity, faces are preserved more accurately than in standard mode.
Two limits to remember:
- Input fidelity isn't supported by the gpt-image-1-mini model, so check which deployment you're calling.
- It's about how closely the output follows the input. The style change itself still comes from your prompt ("as a watercolor painting").
If you also want to restrict which part of the image changes, the edit API takes a mask: a PNG with the same dimensions as the input, where fully transparent pixels mark the area to edit.
The other options, and when you'd use them
- partial_images set to 3: Streaming option. The partial_images parameter controls how many intermediate images are returned before the final result, for showing progress. It doesn't affect fidelity.
- quality set to low: Quality trades detail for speed; low is optimized for latency-sensitive use cases. Lowering it won't make faces more recognizable.
- background set to transparent: Background accepts auto or transparent, and transparency requires PNG output. It's for cut-out images, not face preservation.
Exam tip
"Keep faces or features recognizable during an edit" = input_fidelity high (not on gpt-image-1-mini). "Edit only this region" = mask.
References
Last reviewed: