Summary

When a pipeline is JPEG→orient→JPEG with no pixel-level operations, zenpipe currently fully decodes to pixels, applies orientation, and re-encodes — introducing generation loss and wasting ~10× the CPU. JPEG orientation can be done losslessly in the DCT coefficient domain via zenjpeg::lossless::transform().

imageflow4's imageflow_core/src/pipeline.rs:1084-1230 has a working implementation (~150 lines).

What to add

Pipeline classification

Before compilation, check if the processing steps can be expressed as a single lossless JPEG transform:

1. Source must be JPEG — check via zencodecs::from_bytes() format detection
2. Output must be JPEG — check encode format (explicit JPEG, Keep, or Auto resolving to JPEG)
3. All steps must map to lossless transforms — FlipH, FlipV, Rotate90/180/270, Transpose, Transverse, AutoOrient, Orient(exif)
4. If any step is non-lossless (resize, crop, filter, etc.) → fall through to normal pixel pipeline

D4 dihedral group composition

Multiple orientation transforms compose into a single transform via the D4 group (8 elements: 4 rotations × 2 flips). imageflow4 has compose_transforms() using (rotation_steps: 0-3, flip_h: bool) representation.

Execution

let config = zenjpeg::lossless::TransformConfig {
    transform,
    edge_handling: zenjpeg::lossless::EdgeHandling::TrimPartialBlocks,
};
let result = zenjpeg::lossless::transform(source_data, &config, stop)?;

Integration point

Best place is in orchestrate.rs before calling compile() — check if the graph qualifies for lossless, and if so, bypass the entire streaming pipeline. Return the transformed bytes directly as output.

Pull reference

• Source: imageflow4/imageflow_core/src/pipeline.rs lines 1084-1230
  • try_jpeg_lossless() — main entry point (~60 lines)
  • classify_lossless_steps() — step classification (~30 lines)
  • orientation_to_lossless() — EXIF→LosslessTransform mapping
  • compose_transforms() — D4 group composition via to_d4()/from_d4()
• Tests: imageflow4 has integration tests for lossless passthrough

Dependency

zenjpeg already has the lossless module with transform(), LosslessTransform enum, TransformConfig, and EdgeHandling. No new dependencies needed.

Priority

High — this is a common real-world pattern (thumbnail services applying EXIF orientation to JPEG uploads) and the speedup is ~10×. Zero generation loss is a correctness improvement, not just performance.