How to Make AI-Generated Images Look Less Generic: The Post-Processing Pipeline

Getting a "good enough" image out of an AI model is now routine. Getting an image that looks designed, not generated — that holds up next to real photography, real illustration, real design work — is a separate skill.

The "Why AI Images Look Generic" post covered the prompt-side fixes (specificity, film stocks, era references, materials, lighting direction). This post covers the workflow-side fixes: the iteration loop, the post-processing pipeline, the color grading, the upscaling, the composition refinement, and the "looks designed vs looks AI" final-mile check.

The rule: a great prompt gets you 70% of the way. The remaining 30% is iteration + post-processing. Most "AI-generated" tells come from the post-processing being skipped, not from the prompt being bad.

Why "looks AI" is a workflow problem, not a model problem

The current generation of image models (GPT Image 2, Nano Banana 2, Seedream 5.0 Lite, Midjourney v7) can produce images that are objectively indistinguishable from professional work — if you put in the iteration and post-processing. The "AI-generated tell" comes from a small set of identifiable patterns:

Default color grading. Models output slightly desaturated, slightly cool, slightly low-contrast images by default. Real photography has been color graded.
Default sharpness. Models output slightly soft images. Real photography has been sharpened.
Default grain. Models output clean images with no grain or film texture. Real photography has grain.
Default composition. Models default to centered subjects with shallow DoF. Real photography varies.
Default lighting. Models default to soft even lighting. Real photography has lighting that flatters the subject.
No post-production artifacts. Real photos have slight color casts, slight vignetting, slight noise. AI images do not.

The fix is a post-processing pipeline that addresses all 6. The pipeline is the same whether you use Lightroom, Photoshop, Capture One, or GIMP.

The post-processing pipeline (5 steps)

Step 1: Color correction (Lightroom / Camera Raw / Capture One)

The model gives you an image with a color cast — usually slightly cool, sometimes slightly warm, often slightly desaturated. Color correction fixes the cast and brings the colors to a neutral baseline.

The 4-step color correction:

White balance — set the white balance so that neutral grays are actually neutral. Use the white balance eyedropper on a known-neutral area (white background, gray card).
Exposure — adjust so the highlights are not blown out and the shadows are not crushed. Use the histogram.
Contrast — add 10-20% contrast. Models default to low contrast; real photography has more.
Saturation / vibrance — bump vibrance by 10-20% to bring back color depth. Saturation is more aggressive; use vibrance first.

Time: 30 seconds to 2 minutes per image. The difference between "AI" and "designed" is in this step.

Step 2: Color grading (Lightroom / Capture One / Photoshop)

Color correction is "make the colors accurate." Color grading is "make the colors feel like a specific look." This is where the image goes from "technically correct" to "designed."

The 5 color grading looks that work for AI output:

Warm film — slightly warm highlights, slightly cool shadows, soft contrast. Like Kodak Portra 400. The "lifestyle" look.
Cool cinematic — slightly desaturated, slightly cool, high contrast, deep shadows. Like a Denis Villeneuve film. The "moody" look.
High-key editorial — bright, even, low contrast, soft pastels. Like a magazine cover. The "editorial" look.
Vintage — slightly faded, slightly yellow, soft contrast, lifted blacks. Like 1970s film. The "retro" look.
Punchy commercial — high saturation, high contrast, deep blacks, crisp whites. Like a product ad. The "commercial" look.

Apply one of these looks as a starting point. Adjust per image. Save the look as a preset for consistency across a series.

Time: 1-3 minutes per image.

Step 3: Sharpening and clarity

AI images are slightly soft by default. Real photography has been sharpened. The fix:

Lightroom/Camera Raw: add 30-60% sharpening, mask the sharpening to edges only (so smooth surfaces stay smooth), add 10-20% clarity for midtone contrast.
Photoshop: use the "Unsharp Mask" or "Smart Sharpen" filter on a duplicated layer. Mask to edges.
Capture One: use the "Sharpening" panel, set amount to 80-120, threshold to 1-2, mask to edges.

The "AI sharpness tell": AI output has uniform sharpness everywhere. Real photography has differential sharpness — sharp on the subject, slightly soft on the background. The fix: mask the sharpening to the subject.

Time: 30 seconds to 1 minute per image.

Step 4: Grain and texture

Real photography has grain. AI images do not. The fix is to add grain back — but subtly, not heavily.

Lightroom: add 15-30% grain, size 25, roughness 50. Or add a film stock preset.
Photoshop: add a noise layer (Filter > Noise > Add Noise, 1-2%, Gaussian, Monochromatic). Set the layer to 10-20% opacity.
Capture One: add film grain in the "Film Grain" panel.

The "AI grain tell": no grain at all. Adding 15-30% grain is enough to break the "too clean" tell without making the image look noisy.

Time: 30 seconds per image.

Step 5: Vignette and final touches

A subtle vignette darkens the corners and draws the eye to the center. Real photography has subtle vignetting from lens optics. AI images do not.

Lightroom: add -10 to -20 vignette (subtle, just enough to darken the corners slightly).
Photoshop: add a soft black radial gradient on a new layer, set to 10-20% opacity, blend mode "Multiply."
Capture One: add a vignette in the "Vignette" panel, set to -10 to -20.

Other final touches:

Slight noise in the shadows (gives the image depth)
Slight color cast in the highlights or shadows (gives the image a "look")
Slight texture in the smooth areas (gives the image tactile feel)

Time: 30 seconds per image.

The iteration loop (the part most people skip)

The first generation is a draft. The accepted output is the result of 3-5 iterations. The iteration loop is:

Pass 1: Generate 4-8 candidates. Don't stop at the first image. Generate a batch. Pick the one closest to intent.

Pass 2: Refine the prompt based on the chosen candidate. Look at the chosen image. What works? What doesn't? Adjust the prompt:

"The skin tone is too warm" → add "skin tone accurate, not orange"
"The lighting is too flat" → add "dramatic side lighting, deep shadows"
"The composition is centered" → add "rule of thirds, subject on the left third"

Pass 3: Generate 4-8 variants of the refined prompt. Pick the closest to intent.

Pass 4: Add the post-processing pipeline. Take the accepted image through color correction, color grading, sharpening, grain, vignette. Compare to a "real" reference image (a stock photo in the same style).

Pass 5: Re-generate with the post-processed version as a reference. If the post-processed version has a specific look (say, a film grade) that you want to lock in, use it as a --sref (style reference) for future generations. The model locks onto the look.

The loop is: generate → critique → refine prompt → generate → post-process → use as reference for next batch. The first 5 images of a project take 2 hours. The 50th image takes 10 minutes.

The "looks designed vs looks AI" final-mile check

Before you ship an AI image, run the 6-point check:

Is the color grading applied? Compare the image to a stock photo in the same style. If the stock photo has more "look" (warm film, cool cinematic, etc.), your image needs more grading.
Is the image sharp on the subject? Tap on the subject's eye / detail. If it's not crisp, add more sharpening.
Does the image have grain? Look at a smooth area (a wall, a sky). If it's perfectly clean, add subtle grain.
Is the composition intentional? Centered subject? Soft DoF? Or is the composition working — rule of thirds, leading lines, foreground/background separation?
Is the lighting flattering the subject? Or is the lighting "default" — soft and even? Real photography has lighting that flatters.
Is the image slightly imperfect? Real photography has slight imperfections — a tilted horizon, a partial occlusion, a slightly off-color. AI output is too perfect. Add one imperfection deliberately (a slight crop, a slight tilt, a slight color cast).

The "imperfection" rule is the most counterintuitive: AI images look generic because they are too perfect. Adding a slight imperfection makes them look real.

The composition refinement (post-prompt)

The model defaults to centered subjects with shallow depth of field. Real photography varies. The fix is composition refinement after generation.

The 5 composition refinements:

Crop tighter. The model often leaves too much negative space. Crop in by 10-20% to tighten the composition.
Apply the rule of thirds. Crop so the subject is on a third, not centered. This is the single biggest composition fix.
Add leading lines. If the scene has natural lines (a road, a horizon, a wall), make sure they lead to the subject.
Vary the depth of field. Don't always shoot at f/1.4. Real photography uses f/4, f/5.6, f/8 for different subjects.
Vary the angle. Don't always shoot at eye level. Try low angle, high angle, Dutch angle, over-the-shoulder.

Time: 30 seconds to 2 minutes per image in Photoshop or Lightroom crop tool.

The upscaling pipeline (for high-resolution output)

If you need 4K output, you cannot just "generate at 4K" and call it done. Models have reliability cliffs above their sweet spot. The fix is generate at the sweet spot, then upscale.

The 3-step upscaling pipeline:

Generate at 2K (sweet spot for most models). GPT Image 2 high quality at 1024×1536, Nano Banana 2 at 2K, Seedream 5.0 Lite at 2K.
Upscale with an AI upscaler. Magnific (generative, adds detail), Topaz Gigapixel (faithful, traditional), Real-ESRGAN (free, open source), Let's Enhance (cloud-based).
Re-sharpen and re-grain. The upscale can soften the image. Apply the sharpening and grain pipeline from above.

The model pick for upscaling:

Magnific — best for creative upscaling, adds detail, can hallucinate
Topaz Gigapixel — best for faithful upscaling, doesn't hallucinate
Real-ESRGAN — best for free, open-source, good general-purpose
Photoshop "Preserve Details 2.0" — best for built-in, no extra tool, good for most cases

Time: 1-5 minutes per image.

The "real photograph" reference workflow

The most reliable way to break the "looks AI" tell is to compare your output to a real photograph in the same style. The workflow:

Find 3-5 reference photographs in the style you want. Pinterest, a stock photo site, a film still.
Generate the AI image with your prompt.
Post-process with the 5-step pipeline.
Compare to the references. What does the reference have that the AI image doesn't? More grain? More contrast? A different color cast? A different crop?
Adjust the post-processing to match the reference's "look" more closely.
Re-generate if needed. Use the post-processed version as a --sref for future generations.

This is the "calibration" step. Most "looks AI" tells come from the AI image missing a quality the reference has. The fix is to identify the missing quality and add it back in post-processing.

The post-processing time budget

For a single image:

Color correction: 1-2 minutes
Color grading: 1-3 minutes
Sharpening: 30 seconds
Grain: 30 seconds
Vignette + final touches: 30 seconds
Composition refinement: 1-2 minutes
Total: 5-10 minutes per image

For a batch of 50 images with consistent style:

Define preset once: 10 minutes
Apply to 50 images: 1-2 minutes each = 1-2 hours
Total: 1.5-2.5 hours for 50 images

This is faster than reshooting. And the result is more consistent than reshooting (every image has the same color grade, the same sharpening, the same grain).

The pre-flight checklist

Before you ship an AI image:

Color correction applied (white balance, exposure, contrast, vibrance)
Color grading applied (warm film / cool cinematic / editorial / vintage / commercial)
Sharpening applied (with masking to subject)
Grain added (15-30% in a smooth area)
Vignette applied (subtle, -10 to -20)
Composition refined (rule of thirds, tighter crop, leading lines)
Imperfection added (slight tilt, slight color cast, slight crop imperfection)
Upscaled to target size (if needed, with re-sharpening)
Compared to real reference (3-5 reference photos, identify what's missing)
Saved with all metadata (prompt, model, seed, post-processing steps)

Skip any of these and the image looks "AI-generated." Apply all 10 and the image looks designed.

The summary

The prompt gets you 70% of the way. The remaining 30% is the post-processing pipeline and the iteration loop.

5-step post-processing pipeline: color correction, color grading, sharpening, grain, vignette.
5-pass iteration loop: generate → critique → refine prompt → generate → post-process.
6-point final-mile check: color grading, sharpness, grain, composition, lighting, imperfection.
5 composition refinements: tighter crop, rule of thirds, leading lines, varied DoF, varied angle.
3-step upscaling pipeline: generate at 2K → AI upscale → re-sharpen and re-grain.
Reference workflow: 3-5 real photographs as the calibration target.
Time budget: 5-10 minutes per image, 1.5-2.5 hours for 50.

The model is not the bottleneck. The post-processing is. Add the pipeline, the loop, and the check, and AI images stop looking like AI images and start looking designed.