4.4Retouching & upscaling (Photoshop, Topaz)
The essentials in 30 seconds
- No generated image is deliverable raw: plan a finishing step — flaw fixing then enlarging.
- Fix with two hands: Generative Fill (Photoshop, via Firefly) to regenerate, manual tools (clone stamp, healing brush) where the AI fails to blend.
- Sonia's reflex: fix BEFORE enlarging, check at 100% after upscaling — and beware creative upscalers on faces, text and logos.
Running case — Sonia and the festival banner
Sonia, freelance graphic designer in Algiers, must deliver a three-meter banner for a festival: her Midjourney base is superb but has two flaws — an illegible shop-sign text in the background and a visible inpainting seam on a hand — and its native resolution is far too low for large-format printing. She fixes first: Photoshop Generative Fill to replace the faulty sign, clone stamp and healing brush for the seam. Then she enlarges with Topaz Gigapixel on a conservative setting, and inspects faces and text at 100%. Fourth milestone of the running case: generation was only the start — human finishing makes the deliverable.
Sequel to cases 4.1-4.3 — Sonia moves from generation to a deliverable finish.
1. Fixing flaws: generative + manual
A generated image almost always needs finishing. Flaw fixing — approximate hands or fingers, illegible text, inconsistent details, stray elements, imperfect inpainting seams (section 4.1) — is done with Photoshop and its Generative Fill (powered by Firefly), which selects an area to cleanly regenerate or remove it, plus classic tools (clone stamp, healing brush, dodge and burn).
Photoshop's advantage is manual control: where the AI fails to blend, a few seconds of clone stamp or a local adjustment fixes the problem for good. Like the faulty sign on Sonia's banner, you do not negotiate with a flaw: you fix it.
2. Enlarging: upscalers rebuild (and sometimes invent)
Generated images' native resolution stays limited for some uses: print, large format, high-density display, or major cropping. Specialized upscalers bridge that gap: Topaz Gigapixel, Magnific, or the upscalers built into Stable Diffusion interfaces. They do more than interpolate: they rebuild detail, giving spectacular results.
Important nuance: so-called “creative” upscalers invent detail and can slightly alter the image — a texture, a face, letters. On stone or foliage, no consequence; on a face, text, a logo, it is a problem if the image must stay exact. Hence the rule: conservative setting when fidelity matters, and systematic 100% check after upscaling.
3. The finishing order — and the legal argument
An effective work order: fix first, enlarge second — enlarging a flaw makes it more visible and costlier to fix — then apply color finishing and export for the final medium.
Finally, keep in mind an often decisive argument: this human retouching work also strengthens your position on the image's protection (section 0.4), since it adds personal creative input. As with Sonia's banner, the right view is not “AI or retouching” but “AI then retouching”: AI shortened creation, finishing stays human — and that is where amateur and professional part ways.
🔎 Concrete example — 🔎 Concrete example — The festival banner, fixed then enlarged
Three meters wide, zero flaws allowed. Sonia's actions:
1. Generative Fill (Photoshop): select the faulty sign → replace it cleanly.
2. Clone stamp + healing brush: erase the inpainting seam on the hand.
3. Upscale (Topaz Gigapixel, conservative setting) → 100% inspection of faces and text.Reading: fix (generative + manual), then enlarge, then check.
Lesson: never enlarge a flaw — and watch what the creative upscaler invents.
Section vocabulary
- Generative Fill
- Photoshop function (via Firefly) for regenerating or removing a selected area.
Simple example: Sonia replaces her banner's illegible sign in one selection.
What it is not: a manual clone stamp duplicating existing pixels.
Reflex: area to redo or remove: Generative Fill first.
- Manual retouching
- Fixing with classic tools (clone stamp, healing brush, dodge and burn) where the AI blends poorly.
Simple example: Sonia clone-stamps away the seam that three inpaintings did not fix.
What it is not: rerunning twenty times hoping for a miracle.
Reflex: stubborn seam: thirty manual seconds beat ten retries.
- Upscaler
- Enlarging tool that rebuilds detail (Topaz Gigapixel, Magnific, Stable Diffusion upscalers).
Simple example: Sonia brings her banner to large-format print resolution without blur.
What it is not: simply stretching pixels, which blurs.
Reflex: demanding medium (print, large format): specialized upscaler, not resizing.
- Creative upscaler
- Upscaler that invents detail: spectacular result but can alter the image — watch it.
Simple example: Sonia keeps it minimal on her banner's faces and text to deform nothing.
What it is not: neutral, faithful enlarging in every case.
Reflex: constrained subject (face, logo, text): conservative setting + 100% check.
In practice — In practice — Finalizing an image for pro use
- Spot every flaw (hands, text, stray elements, seams) on your generated image.
- Fix them: Photoshop Generative Fill for areas to regenerate, manual tools (clone stamp, healing brush) for seams.
- Then enlarge with an upscaler suited to your use (print, large format), and check that no detail was altered.
- Finish with color adjustments and export at the final medium's format/resolution.
Ready-to-say script — explaining finishing to a client (30 s)
“This is Sonia: your generated visual is an excellent base, but for a three-meter banner I still need to hand-fix two flaws and enlarge it cleanly — otherwise the flaws will show at large size. This finishing is what separates a screen image from a print deliverable.”
Say this to get the finishing phase accepted in a quote: price fixing + upscaling separately.
Copy-paste template — finishing checklist
Image: [name] — final medium: [3 m banner / A3 poster / social post]
[ ] Flaws spotted (hands, text, strays, seams)
[ ] Generative Fill: [areas] — Manual (clone/heal): [areas]
[ ] Upscale: [tool] — setting [conservative/creative] — target resolution [value]
[ ] 100% check: faces, hands, text, logo
[ ] Color + export [format, resolution]Large-format print
Like Sonia: full fixing, conservative upscale, 100% check — no flaw forgives at large size.
Web or social visual
Fix flaws visible at screen size, light upscale: finishing stays mandatory, but faster.
Sensitive subject (face, client logo)
Creative upscaler at minimum, comparison with the original: fidelity beats spectacular.
Series to produce
Documented finishing (tools, upscale settings) to apply the same treatment to the whole series.
✓ Best practices
- Fix flaws BEFORE enlarging: an enlarged flaw becomes obvious.
- Check sensitive areas at 100% after upscaling (faces, hands, text, brand elements).
- Dose creative upscalers' strength on constrained subjects.
✗ Common pitfalls
- Enlarging a still-defective image: the flaw is amplified.
- Trusting the global preview after upscaling without inspecting faces and text.
- Forgetting that human finishing work also strengthens the image's protection.
Key takeaways
- NO generated image is deliverable raw: plan a finishing step.
- FIX with Generative Fill (Photoshop/Firefly) AND manual tools where the AI fails to blend.
- ENLARGE with a specialized upscaler (Topaz Gigapixel, Magnific); creative ones invent detail.
- ORDER: fix → enlarge → color. Human input strengthens the legal position.
48h challenge — apply tomorrow
Take a generated image: fix 1 flaw with Generative Fill and 1 seam with the clone stamp, enlarge it, then inspect at 100% one face or text.
- I can fix with Generative Fill and finish with clone stamp/healing brush.
- I can enlarge with an upscaler while dosing creative strength.
- I apply the fix → enlarge → color order with a 100% check.
Frequently asked questions
How far can you enlarge a generated image without quality loss?
Modern upscalers allow major enlargements with a sharp render, because they rebuild plausible detail instead of stretching pixels. But distinguish perceived sharpness (excellent) from fidelity (not guaranteed — added detail is invented). On texture, no consequence; on a face, text or logo, the upscaler can alter features or distort letters — especially creative upscalers on high settings. Recommendations: fix before enlarging, check sensitive areas at 100%, dose creative strength, keep the original for comparison.
Is manual retouching still needed, with such powerful AIs?
Yes, lastingly, for any professional use. AI produces the plausible, not the exact — yet a brand color, an alignment, readability are non-negotiable and fixed by hand in seconds. Structural weaknesses persist (hands, text, symmetries, seams): a clone stamp definitively fixes what several regenerations don't. And final integration (text, logo, margins, harmonized color) is the core of graphic craft. Add the protection argument (section 0.4). The right view: AI THEN retouching.
What is the right finishing order for a generated image?