AUTO PASS 3 · LOAD

Pass 3 in this pack's automated pipeline is the REFINER HIDREAM stage - the pass that runs the hires result through a refiner model to clean up and polish what pass 2 produced. And like every pass after the first, it starts the same way: at a fixed LOAD node that pulls the previous phase's validated checkpoint off the disk and hands its work to this phase.

SayaImagePhase3Load is that node, and it's a strict sibling of the other AUTO PASS N · LOAD nodes: zero inputs, six outputs. There are no sockets because there's nothing to wire - the state lives in ComfyUI/output/image/checkpoints/, where phase 2's STOP node promoted phase_2_hires.png and its manifest. This phase just reads them.

Outputs

• image - phase 2's validated hires image, ready for the refiner pass.
• manifest_json - phase 2's record: seed, prompts, models/VAEs/samplers, paths.
• checkpoint_path - where that validated image lives on disk.
• seed - the seed behind the hires result.
• positive_prompt / negative_prompt - what produced it.

The seed and prompt outputs are the quiet workhorses here. A refiner usually wants to continue from the hires image rather than reinvent it, and re-exposing the exact seed + prompts from phase 2 lets the refiner's sampler keep continuity instead of drifting into a different image. Your refiner stage can denoise lightly with the same seed and prompt, or intentionally diverge - but at least it knows what came before.

The dependency you can't skip

Like every LOAD node in the family, this one refuses to run if phase 2 hasn't validated. No phase_2_hires.png/.json marked validated means an error - the pipeline won't feed a refiner an image that was never approved or that you discarded with a Restart. So the golden rule is the same for all six passes: run them in order and let each STOP (or the phase-1 Review Continue) promote its candidate before the next phase's LOAD fires. Its change-detection watches phase 2's manifest mtime, so regenerating phase 2 automatically invalidates a cached phase-3 run - no manual cache clearing when you redo an earlier pass.

If you're not running this six-pass system, a LoadImage plus your own refiner setup is all you need; this node's entire value is the on-disk checkpoint handoff between isolated passes. And since the whole pack is explicit work-in-progress, expect the occasional port or behavior change - 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, then restart. Only numpy/Pillow as dependencies - nothing to download.