PNG, JPEG or WebP for ad stills: measure file size with Pillow

Request png, jpeg or webp with output_format, and compress locally if a platform wants smaller files, because output_compression is in the schema but no model serves it. Per the Image API docs, sending it returns 400 unsupported_parameter. So your quality dial lives in Pillow, not in the request.

The script below takes one downloaded PNG and prints how big the same picture is as JPEG and WebP at a few quality levels, so you pick a setting from your own images instead of a rule of thumb.

What does the API give me for format?

The output_format field lists png, jpeg, webp or svg, and the response labels each file with media_type, for example image/webp. A model only accepts what its catalog descriptor lists, so read supported_parameters first. A parameter that is not listed is rejected, not ignored.

How do I compare sizes?

Save one result as a PNG, then re-encode it in memory. Photos usually shrink a lot as JPEG or WebP; flat graphics with sharp text may not, so test both kinds.

import io
from PIL import Image

def sizes(path):
    img = Image.open(path).convert("RGB")
    out = {"png": len(open(path, "rb").read())}
    for q in (95, 85, 75):
        for fmt in ("JPEG", "WEBP"):
            buf = io.BytesIO()
            img.save(buf, fmt, quality=q)
            out[f"{fmt.lower()}-q{q}"] = buf.tell()
    return out

# Demo file; replace with a downloaded ad still.
Image.effect_noise((600, 600), 30).convert("RGB").save("demo.png")
for name, n in sizes("demo.png").items():
    print(f"{name:10} {n / 1024:8.1f} KiB")

Which format should I ship?

Keep the PNG as the master, since each re-encode of a JPEG or WebP loses detail, and publish a copy in whichever format meets the upload limit. Open the smallest copy at full size and check text edges before you accept it.

What should I record?

Store the master URL, the format you shipped and the quality number next to the job id, so a later re-export starts from the master and not from the compressed copy.