Prompt chaining workflows
Great AI images rarely come from a single generation. A real pipeline chains stages: generate a base image, restore the face, upscale, inpaint problem areas, then export. Get the order wrong — upscale before face restore, inpaint before upscale — and you fight your own tools. This planner assembles the correct order and estimates the total processing time per image.
How it works
Each stage has a rough per-image cost that scales with your GPU tier. The planner holds the proven execution order and sums only the stages you enable:
generate → face-restore → upscale → inpaint → export
total_time = Σ stage_time(enabled stages, gpu)Generation and upscaling dominate the budget; export is effectively free. Running face restore on the small native face (before upscaling) and inpainting at the final resolution (after upscaling) is what keeps quality high.
Why stage order is not negotiable
Each stage leaves the image in a state the next stage expects. Run the stages out of order and quality falls apart:
- Face-restoring after upscaling wastes the restoration on a large, blurry face region and can introduce double-sharpening artifacts.
- Inpainting before upscaling repairs detail at low resolution; the upscaler then softens the repair, forcing a redo.
- Exporting before inpaint obviously exports the broken version.
The planner bakes the proven order in so you cannot accidentally wire the stages wrong before committing to a full batch.
Typical GPU time ranges
These are rough indicative ranges to help you understand relative cost; actual times depend on resolution, step count, and model size:
| Stage | Relative cost |
|---|---|
| Generation | High — the most compute-intensive step |
| Upscaling | Medium-high at 4× — lower for a 2× pass |
| Face restore | Low — fast, local operation |
| Inpaint | Medium — depends on mask size and steps |
| Export | Negligible |
Tips for building the chain
- Restore early, inpaint late. Fix faces at native size, fix everything else after upscaling.
- Upscale once. A single good 2×–4× pass beats stacking several small ones, which compounds softness.
- Mask tight for inpaint. Smaller masks at high resolution blend more cleanly and run faster.
- Build it in ComfyUI. A saved graph runs the whole chain unattended across a batch — ideal once the planner confirms your time budget.
- Use the time estimate as a batch budget. Multiply the per-image total by your target batch size to see how long the run will take before you commit.
- Test the chain on one image first. Run the full pipeline on a single image before committing to a 50-image overnight batch. Catching a misconfigured inpaint mask costs minutes, not hours.
- Save your ComfyUI graph. Once you have a working pipeline, the saved JSON graph is your repeatable workflow — changes to the planner settings should map back to your graph before the next batch.