I'm supporting a collaborator as he implements an image transform and image proxy using libvips via Elixir Image which I maintain.

To take best advantage of libvips pipelining I want to make sure the order of transforms is as good as it can be. I know I've seen something on this in both libvips and sharp but I can find neither now. ChatGPT offers the following - does that seem correct?

1. Load image with region/size hint if possible
   Use VipsImage.thumbnail() or load with shrink-on-load flags (e.g., VipsForeignLoadOption::shrink with JPEG).
   ✅ Reduces input size early, saving downstream processing.
2. Crop / Extract ROI
   Reduce working data early in the pipeline.
   ✅ Efficiently drops unneeded pixels from processing.
3. Affine transformations (rotate, flip, transpose)
   Perform these after reducing size but before color transforms.
   ⚠️ Expensive in memory if done too early; can block streaming.
4. Resize / scale
   Downsample images early (ideally at load) if possible.
   ✅ Minimizes pixel data for later operations.
5. Color space conversion
   Convert to the working color space (e.g., sRGB → linear) only if necessary, and delay until needed.
   ⚠️ Color conversions can’t stream and consume more memory.
6. Pixel-level operations (e.g., brightness, contrast, gamma)
   These operate per-pixel and are fast and parallel-friendly.
7. Alpha compositing / blending
   If compositing, try to minimize the resolution and use the same color space beforehand.
8. Save to file / encode
   Save as the last step. Use streaming-friendly formats when possible (JPEG, PNG).