5.5 Post-Processing and Rescuing Shots
A published frame is rarely one generation. It is a generation, an edit or two, an upscale, and a grade. The operators whose work looks effortless are usually not using a better model — they are running a short, disciplined chain and knowing when to stop.
This chapter is that chain, plus the rule that saves the most time in it.
Inpainting: what to fix, in order of payoff
Regenerating only a masked region; the surgical tool for a hand, an eye, or one bad detail. regenerates a masked region, using the surrounding image and the identity layer as context.
Hands. The most reliable tell in synthetic imagery (1.4) and the most fixable. Mask the hand, mask generously, and re-generate. If two attempts fail, change the pose or crop the hand out — a re-posed frame at the second attempt is cheaper than the fourth attempt at the same hand.
Ears and jewellery. Earrings are a frequent failure: the model renders a plausible object in the wrong place, or two of them at different heights. Necklaces merge into collarbones. These are small masks with high success rates.
Background intrusions. A person, a limb, or an object appearing where the scene should be empty. Especially common in crowds or in reflections, and often fixable with a mask smaller than you expect.
Wardrobe errors. A garment that changes mid-frame, an impossible fold, a sleeve that merges with the body. Fix the mask, not the prompt — re-generating the whole frame loses the composition you chose.
Eyes and teeth. Small, high-consequence regions. Asymmetry and malformed teeth are noticed in a way that a slightly wrong background never is. Fix, or reject the frame; do not publish a compromise here.
What not to inpaint: an off-identity face. If the face does not match, the frame is a reject. Inpainting the face region produces a different person wearing the same scene, which is worse than the original problem.
The upscale stage
Two families of upscaler are in current use for this work. [AS OF 2026-10] Descriptions are directional; both are active projects and specifications change.
Restoration-oriented upscalers. Strong at recovering plausible detail in soft frames, at the cost of inventing it. Excellent for a slightly soft face; dangerous for a frame with artifacts, which it dutifully enlarges.
Diffusion-based upscalers. Add detail by re-generating at higher resolution with reference conditioning. More control, more compute, and better preservation of identity when the reference is wired in. The cost is time, and the temptation to use them for every frame.
The decision rule: upscale everything you publish at least once, but do not run the expensive route on frames that will be discarded. Quantity first, upscale the survivors. Chapter 5.4's ordering rule already places crop and edit before this stage.
Resolution target, not maximum resolution. Aim for the platform's requirement plus headroom, and no more. Instagram does not display 8K, and the extra pixels cost time and money that the audience never sees. Keep a high-resolution master for the archive; export to the platform's size.
The skin detail pass
Where it belongs: after the upscale, at final resolution.
Why after: a detailer sharpens skin texture, pores, and fine hair. An upscaler then re-interprets those pixels and, depending on the algorithm, partially smooths them. Running the detailer last means its work survives to export.
What it fixes: the smooth, waxy quality that reads as synthetic (1.4). This is the single most effective counter to the plastic-skin tell, and it is a few seconds per frame.
What it does not fix: an inconsistent face, badly lit skin, or the wrong person. A A face-detail operation run last at final resolution and low denoise; over-detailing reads as synthetic. on a frame that fails identity QA is effort spent making the wrong face look more real.
How far to push it. The failure mode is over-detailing: visible texture in the wrong regions (a smooth wall acquiring pores), exaggerated skin contrast, and a crunchy, sharpened look. Score the result against the close-up frame in your 4.6 battery, which exists partly to catch exactly this. Chapter 4.6's rule applies: a poreless close-up is a 3 wearing a 5's clothes, and a crunchy one is the same problem in the other direction.
Color: the preset, applied last
The grade is the strongest consistency lever in the kit (2.5), and here is where it is enforced.
- One preset, applied identically to every exported frame. Including the ones you are unsure about. Especially those: a frame that needed rescuing and then received a different grade is a frame that reads as different.
- Grade after geometry, before export. Crop and upscale change how the curve lands on the image.
- Never grade per frame by eye. The whole value of the preset is that it does not respond to your mood on Thursday.
A recorded exception: promotional and ad assets often use a brighter, flatter grade (5.2). That deviation is legitimate, and it should be a second preset with a name, not an improvisation.
Exporting with provenance
One last step, before the archive, that costs seconds and answers most of the questions you will have later.
Write the frame's provenance into the exported file's metadata, or into a sidecar beside it: the identity version and base hash, the prompt One prompt-and-settings combination inside a batch, tracked so its survival can be measured., the reference versions used, the date, and which edits were applied. Chapter 11.5 covers the Provider-side marking as C2PA credentials or pixel watermarks; not your obligation and not reliable. obligations; this is the practical layer underneath them, and it is also how you answer three questions that will certainly come up.
Which reference produced this? When a frame turns out to have the wrong outfit, you need to know whether the fault is in the reference or the prompt.
What did we publish under version 3? The disclosure record in 11.3 requires knowing what went out when, and the retirement procedure in 4.7 depends on it.
Can this be reproduced? Same references, same prompt, same model version, same The random starting point; the same seed with the same settings reproduces the same image.. Without the provenance record, no.
The archive of full-resolution masters plus this metadata is the part of your pipeline that survives a model change, a tool change, and a mistake — which is more than can be said for the frames themselves.
The three-attempt rule
Three attempts, then decide. A frame that fails three fixes is not a frame that needs a fourth. It is a frame to reject, or a pose to change.
This rule is the difference between a two-hour production session and an eight-hour one. The specific failure it prevents is the loop that everyone falls into: the mask is nearly right, the hand is almost fixed, one more attempt. Three attempts is roughly when the probability of success stops improving — the remaining problem is usually structural (the pose, the crop, the reference), and no amount of iteration on the pixels addresses it.
Two alternatives when the third attempt fails, both faster than a fourth:
Change the frame, not the pixels. Crop it out, re-pose, or pick a different generation from the One production run, sized from the survival rate rather than from ambition.. The batch has 150 frames in it; you are not short of options.
Change the input. If hands keep failing in the same pose type, the fix belongs one layer up — in the reference set or the prompt variant list, permanently, not in this frame.
Realistic effort per frame
A useful budget, from batch of 150 to published set:
| Stage | Time per surviving frame |
|---|---|
| Selection from batch | Seconds |
| Identity QA (4.6 rubric) | ~30 s |
| Edits / inpaint when needed | 2–5 min |
| Crop | Seconds |
| Upscale | Machine time, no attention |
| Detail pass | Seconds |
| Grade + export | Seconds |
The shape of that table is the useful part. Attention concentrates in the edit step and in selection, and everything else is machine time or seconds. This is why chapter 5.4's ordering matters so much: putting the expensive-attention step before the cheap ones is what makes a batch affordable.
Honest accounting: what a portfolio image actually is
The images that make operators look talented are almost never single generations. They are:
Generated at volume, with most candidates discarded. Picking is a skill and it is invisible in the output.
Edited, usually in one or two places, sometimes more.
Upscaled, once, at the end.
Graded identically to everything else in the feed.
Selected against a QA rubric rather than a feeling.
None of those five steps is difficult. The reason the results look impressive is that most people skip two or three of them, and the reason the process looks like magic from outside is that nobody posts the 149 frames they discarded.
Where this goes wrong
Inpainting an off-identity face. You have made a different person in the same room.
Raising resolution on pixels rather than latents, so it runs after composition, identity, and repair. before editing. Chapter 5.4's ordering rule, in its most expensive form.
Detail pass before upscale. Its work gets partially undone.
Grading by eye. The preset exists precisely so this does not happen.
A fourth attempt. The three-attempt rule exists because the fourth rarely works.
Maximum resolution for everything. Cost with no visible return.
Over-detailing. Crunchy skin is the same tell as smooth skin, reversed.
Publishing a compromise on eyes or teeth. The two regions where a compromise is visible.
The gate
- Do you know which six defect classes you inpaint, and which one you refuse to (the face)?
- Is the upscale stage after all geometry changes?
- Is the skin detail pass after the upscale, at final resolution?
- Is there one saved preset, with named exceptions only for promotional sets?
- Do you apply the three-attempt rule, and do you know your two fallbacks?
- Is your attention budget concentrated in selection and edits?
What you leave with
A priority-ordered list of what to inpaint and one rule about what never to, the two upscaler families with a selection rule and a resolution target, the skin detail pass with its placement and its over-detailing failure mode, the preset as the enforced consistency lever, and the three-attempt rule with two faster alternatives — plus an honest account of what a finished portfolio image actually is.