AUTO PASS 6 · LOAD

Pass 6 is the FINAL UPSCALE / COLOR / SAVE stage - the last stop on the line, where the image gets its final resolution bump, any color work, and gets written out for real. SayaImagePhase6Load opens it, and like every fixed LOAD node in the family it has no inputs: it restores phase 5's validated checkpoint from disk and hands this phase everything it needs to finish the job.

The six outputs are the family standard: image (the fully-detailed phase-5 result as a tensor), manifest_json (phase 5's full record), checkpoint_path, seed, positive_prompt, and negative_prompt. Files come from ComfyUI/output/image/checkpoints/, where phase 5's STOP promoted phase_5_detailers.png and its manifest.

What makes this phase different

Two things, both downstream of the LOAD rather than inside it.

First, the resolution context is richer here than anywhere else in the pipeline. This is the phase where you're working toward the final pixel dimensions - the pack pairs it with upscale-target and final-upscale utilities - so the loaded image gives you the actual detailed canvas to size from. There's also more freedom in how you use the seed and prompts: pass 6 can legitimately diverge more than earlier phases (new seed for a color pass, edited prompt for a final touch) because you're at the finish line and can eyeball the result before saving. The earlier phases exist to keep continuity; this one is where you're allowed to finish.

Second, the failure mode softens. Pass 6 is the only phase where loading isn't strictly fatal in spirit - if phase 5 hasn't validated yet, this node still throws the same "previous phase not validated" error as its siblings (the pipeline won't fabricate a source), but it's a two-minute fix: run pass 5, let its STOP promote, re-run. The dependency rule and the mtime-based cache invalidation work exactly as on the other LOAD nodes.

Should you use it?

Only inside the author's automated multi-pass architecture. If your whole job is "generate, detail, upscale, save" as one graph, a normal LoadImage + upscale model + SaveImage is simpler and you should use that. The fixed-phase LOAD nodes are the reproducible-entry-point machinery for a pipeline where every pass is an independently queueable prompt - and pass 6 is where that pays off by letting you queue the final expensive upscale with confidence that it's consuming exactly the detailed image you approved.

Install: ComfyUI Manager search saya-comfy-couple-plus, or git clone https://github.com/alphaziod/saya-comfy-couple-plus into custom_nodes, then restart. Deps are numpy and Pillow; no model downloads. WIP pack - back up workflows before updating.