AUTO PASS 5 · LOAD
Detailers start from the pre-detail upscale — on purpose, every time
- image
- manifest_json
- checkpoint_path
- seed
- positive_prompt
- negative_prompt
Pass 5 is the DETAILERS stage - the detect-crop-refine-paste loop that fixes faces, hands, eyes and hair on the upscaled image. And SayaImagePhase5Load is the fixed-phase node that opens it, reading phase 4's validated checkpoint from disk and handing this phase everything it needs to work: image, manifest_json, checkpoint_path, seed, positive_prompt, negative_prompt. No inputs on the node itself - the state arrives via the checkpoint files under ComfyUI/output/image/checkpoints/.
The detail worth understanding: phase 5 starts fresh from phase 4
There's a deliberate asymmetry buried in the pack's source comments: phase 5 always restarts from phase 4 and never reuses an older phase-5 result as an implicit input. Why? Because the detailers run as one normal chain inside a single phase - each region is detected, cropped, refined, and pasted back in sequence within pass 5 itself. An old "phase 5 output" isn't a valid starting point for a new run; the source of truth for the next attempt is always the pre-detail image from phase 4. So if you re-run pass 5, you re-run the whole detailer chain against the same phase-4 source - which is what makes re-rolls clean and reproducible instead of compounding stale detail.
This is also the phase where the pack's couple philosophy pays off. If your image has two characters, the detailer crops shouldn't be re-prompted with the whole scene - they want character-local prompts (the pack's detailer_positive / person_1_positive outputs exist for exactly this). The seed and prompt outputs of this LOAD give the detailer stage the anchor it needs to keep identity intact while it refines small regions.
The dependency rule, restated
Like every LOAD node in the family, this one throws if phase 4 hasn't validated (phase_4_predetail.png/.json must exist with status: validated). The pipeline will not feed detailers an unapproved upscale. Run the passes in order, and let phase 4's STOP promote its candidate before pass 5 fires. Because the node's cache key watches the phase-4 manifest's modification time, redoing phase 4 automatically invalidates the phase-5 graph - no manual cache clearing.
If you're not running the automated six-pass pipeline, you likely don't need this node - the detailers in the pack can run in a normal graph off an ordinary LoadImage. The fixed LOAD earns its keep specifically when each pass must be independently queueable and start from a verified checkpoint. WIP pack, so expect changes between versions; back up workflows before updating.
Install: ComfyUI Manager search saya-comfy-couple-plus, or git clone https://github.com/alphaziod/saya-comfy-couple-plus into custom_nodes, restart. Deps: numpy, Pillow; no downloads.
Inputs (0)
No inputs
Outputs (6)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | — |
| manifest_json | STRING | — |
| checkpoint_path | STRING | — |
| seed | INT | — |
| positive_prompt | STRING | — |
| negative_prompt | STRING | — |