AUTO PASS 5 · STOP / UNLOAD

The detailer pass is the longest and fussiest stage in this pack's pipeline - crop after crop re-rendered and pasted back - so when it's finally done, the last thing you want is to re-run it by accident. SayaImagePhase5Stop exists to make that impossible: it seals the fully-detailed image as the phase-5 validated checkpoint, frees VRAM, and queues the final pass. One pass left after this.

How the STOP works

Mechanically it's the same STOP as its siblings: image, seed, positive_prompt, negative_prompt, the models_json/vaes_json/samplers_json metadata notes (record-keeping only), optional source_path, and a save_receipt the node accepts and ignores. It writes phase_5_detailers.candidate.png/.json, atomically promotes them to phase_5_detailers.png/.json under ComfyUI/output/image/checkpoints/, and emits the completion event. Outputs: image, validated_path, manifest_json. Unload waits for execution_success, then pass 6 queues.

Two phase-5 specifics

First, detailer bookkeeping. The fixed phase-5 STOP records detailer: "none" because, in the automated flow, the detailers run as one chained pass and the whole result is what gets checkpointed. Per-detailer candidate tracking exists, but it lives on the manual CheckpointStop node when you drive detailers one at a time - in the auto pipeline, this STOP is the single authoritative "detailers finished" marker. The manifest's detailer field tells later readers that.

Second, and this is the one that matters if anything ever looks wrong: pass 5 always re-runs from the phase-4 source. When this STOP validates, it becomes the input to pass 6 - the final upscale/color stage - not to another detailer run. If you discard this candidate via the redo endpoint, the next pass-5 attempt rebuilds the whole detailer chain from phase_4_predetail.png, never from an older detailer result. That's a feature: no compounding stale detail, at the cost that re-rolling detailers means paying for the full chain again. Fine if you're iterating on a face, annoying if you wanted a cheap partial redo.

Operational notes don't change across the family: passes must validate in order, headless runs keep the checkpoint but lose the auto-queue, and the pack is explicit WIP - back up workflows before updating, and fully restart ComfyUI if a saved workflow shows stale ports.

Install: ComfyUI Manager search saya-comfy-couple-plus, or git clone https://github.com/alphaziod/saya-comfy-couple-plus into custom_nodes, then restart. Just numpy and Pillow.