Field note
Replicate's littletry/compress-image utility converts JPEG, PNG, WebP, and AVIF on CPU. It can optimize for quality or a target file size, resize dimensions, remove EXIF metadata, and preserve ICC color profiles. PNG and WebP support lossless modes; animated images are not supported.
The feature is intentionally ordinary. Image conversion and compression are deterministic preprocessing jobs, so using generative compute for them wastes money and makes outcomes harder to reproduce. A stable CPU step can run before storage, a vision API, or Telegram delivery.
The useful workflow boundary is explicit: inspect the input, transform it, verify output size and format, then invoke a model only when semantic work is required. Packaging that step as a callable service makes it convenient, but teams should still retain hashes and validate the returned artifact.