Chrome Extension
Merriam-Webster made "slop" its 2025 Word of the Year, defining it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence."
If your posters, menus and thumbnails look like everyone else's, the problem is almost never the model.
It is that you gave it a subject and nothing else. I went through the research on why AI images read as AI, pulled the documented prompting guidance from five providers, and built 25 side-by-side prompt pairs plus 6 repair prompts so you can see exactly what separates slop from work a client would pay for. Every pair uses real marketing formats with real specs.
Checkout 10+ AI Image Prompt generators.
Key Takeaways
What "AI Slop" Actually Means Now
The word stopped being internet slang in 2025.
Merriam-Webster named "slop" its 2025 Word of the Year, announced December 15. Their definition:
"digital content of low quality that is produced usually in quantity by means of artificial intelligence."
The American Dialect Society picked the same word for 2025. Simon Willison had popularised the framing back in May 2024 with a line I still think is the clearest version of it: slop is to AI content what spam is to email.
This matters commercially, because platforms started writing it into policy.
| Platform | Policy language | Date |
|---|---|---|
| YouTube | Demonetizes "inauthentic content," including AI work made with "generic or unoriginal templates giving the impression of mass production" | Jul 2025 |
| Meta | "simply stitching together clips or adding your watermark does not qualify as meaningful enhancement" | Jul 2025 |
| "AI modified" labels, applied by classifiers "even if the content doesn't have obvious markers" | Apr 2025 | |
| Spotify | Removed 75M+ spammy tracks in 12 months, naming "other forms of slop" | Sep 2025 |
| Etsy | "all items must incorporate a human touch" | Jun 2025 |
| Adobe Stock | Prohibits submitting "multiple versions from the same prompt or similar iterations" | Jun 2026 |
So this is no longer an aesthetic complaint. Generic output now gets labelled, deranked or demonetized.
Why AI Images Look Like AI
I wanted the actual research here rather than the usual "look at the hands" advice, and some of what I found genuinely surprised me.
The Tell You Are Not Looking For
The biggest study on this is Kamali et al., CHI 2025, with 50,444 participants and 749,828 observations. It found something I have not seen discussed anywhere:
People report using anatomical cues 61% of the time. Fingers, teeth, ears. But the most prevalent artifact class is functional implausibility, present in 58.7% of AI images, and it is the one people notice least.
Functional implausibility means objects that could not actually work. Slack guitar strings. A strap that merges into the fabric it is attached to. A chair leg that does not reach the floor. A menu board bracket bolted to nothing.
Nearly a third of images containing these landed in the 40 to 60% accuracy band, meaning viewers essentially could not tell.
The practical lesson: stop worrying about fingers and start checking whether the objects in your image could physically function.
Context Is Where The Tells Live
Two headline numbers look contradictory until you understand the difference.
Nightingale and Farid, PNAS 2022 found participants scored 48.2% identifying AI faces, which is chance. With training and per-trial feedback, they only got to 59%.
But the CHI 2025 study found 76% accuracy on AI images generally.
The difference is cropping. A tight face gives you nothing to catch. A full scene gives you shadows, reflections, text, physics and object relationships.
For marketing images, this cuts both ways. Your poster or menu is a full scene, so it has more surfaces where slop shows. It also means fixing those surfaces is the highest-leverage thing you can do.
The Physics Problem Is Documented
Hany Farid's work is the clearest source on why lighting goes wrong. In "Lighting (In)consistency of Paint by Text" he explains that generation "is not based on explicit geometric or physical models."
The model has no light source. It has a statistical impression of what lit things look like.
Bianchi et al., IEEE WIFS 2024 tested cast shadows and reflections across 150 images and concluded diffusion systems "are not generally designed to create physically consistent images."
That is why the fix is naming your light source explicitly rather than saying "good lighting."
Why Typography Breaks
This one has a specific documented cause. Liu et al., ACL 2023 showed that the text encoders these models use tokenize at the word or subword level, which means glyph-level information is simply absent from what the model receives.
The model never sees letterforms. It sees tokens.
Layer on top of that what a real typeface actually carries: kerning pair tables, sidebearings, a baseline grid, hinting, optical size axes. None of that exists in a generated image.
That is why generated type has inconsistent stroke weights on repeated letters, drifting baselines and invented ligatures.
My rule after learning this: generate the image, set the type in a real tool. More on that below.
The Default Aesthetic Is Trained In
There is a documented mechanism behind why everything looks the same. ImageReFL, May 2025 states it directly:
"Reward-based fine-tuning using models trained on human feedback improves alignment but often harms diversity, producing less varied outputs."
Models get tuned on aggregate human preference, which optimises toward the average pleasing image. That average is your slop baseline.
Midjourney is the only provider that documents a switch for this. Its Raw mode docs describe Standard mode as adding "its own creative touch" and Raw as "turning off this auto-pilot."
The 7 Layers That Separate Slop From Professional Work
Here is the framework I now run every marketing prompt through. A slop prompt has one layer. A professional prompt has all seven.
| Layer | Slop prompt | Professional prompt |
|---|---|---|
| 1. Subject | "a coffee shop poster" | Named subject, doing something specific, at a specific moment |
| 2. Purpose | absent | The commercial job the asset does and who reads it |
| 3. Light | absent or "good lighting" | Named setup, direction, quality, colour temperature |
| 4. Composition | absent | Grid position, negative space reserved, focal hierarchy |
| 5. Typography | "with text" | Exact wording in quotes, weight, position, size relationship |
| 6. Brand | absent | Hex codes bound to objects, palette ratio, tone |
| 7. Constraints | absent | What must not appear, what must not change |
Google's documentation backs the second layer specifically. From Google's image generation guide:
"Provide context and intent… 'Create a logo for a high-end, minimalist skincare brand' will yield better results than just 'Create a logo.'"
And Black Forest Labs is blunt about which layer matters most. From the FLUX prompting guide:
"Lighting has the greatest single impact on output quality… 'good lighting' is not enough."
25 Before-And-After Prompts For Marketing Materials
Every pair below is the same asset, prompted two ways. I wrote the slop version the way I used to prompt, so read that first and the difference lands harder.
Posters
1. Coffee shop promotional poster
Slop:
A coffee shop poster
Optimized:
Create a print-ready A2 promotional poster for an independent coffee roaster. Purpose: window display poster to drive weekday morning footfall, read from three metres away. Scene: a single ceramic pour-over cone mid-pour, shot on a worn walnut counter. Subject: dark coffee streaming into a glass carafe, steam visible against a dark background. Light: single soft window key from camera left at 45 degrees, deep falloff to near-black on the right, warm 3200K, one specular highlight on the carafe shoulder. Composition: subject occupying the lower-left third, upper-right 40 percent left as clean dark negative space for typography. Typography: the words "SLOW MORNINGS" in a bold condensed sans-serif, upper right, warm cream, set once. Beneath it in smaller regular weight: "Weekdays before 10". Colour: 60 percent near-black #14100D, 30 percent warm brown #6B4A32, 10 percent cream accent #F2E8DC. Constraints: no other text, no logos, no stock-photo treatment, no decorative clutter. Keep all text correctly spelled and legible at three metres. Format: 420x594mm at 300dpi with 3mm bleed, CMYK-safe colours only.
2. Gym membership drive poster
Slop:
Gym poster with a person working out
Optimized:
Create a print-ready 18x24 inch gym membership poster. Purpose: in-window recruitment poster for a January intake, targeting beginners rather than existing lifters. Scene: an empty functional training floor at early morning, rubber flooring and rig visible but unoccupied. Light: cool overhead industrial lighting at 5000K with a single warm practical from a window camera right, long shadows across the floor. Composition: wide empty floor across the lower two thirds, rig silhouette on the right edge, clean upper-left quadrant reserved for type. Typography: "START WHERE YOU ARE" in heavy uppercase condensed sans, upper left, off-white. Beneath in light weight: "No contract. First month free." Colour: 60 percent charcoal #1F2225, 30 percent cool grey #5A6169, 10 percent signal orange #FF5A1F on the accent rule only. Constraints: no people, no equipment brand names, no motivational stock imagery, no lens flare. Render each text line exactly once. Format: 18x24in at 300dpi, 0.25in bleed.
3. Music event poster
Slop:
Concert poster for a jazz night
Optimized:
Create a screen-print style jazz event poster. Purpose: paste-up street poster for an independent venue, must read at distance and survive being printed in two spot colours. Scene: flat graphic composition, no photographic elements. Subject: an abstract double bass silhouette built from overlapping geometric shapes, slightly off-centre left. Composition: subject on the left two fifths, type stacked right, generous margin on all four edges. Typography: "MIDNIGHT BRASS" in large condensed uppercase across the upper third. Below the illustration: "Thursday 14 November" then "Riverside Hall, 8pm". Colour: two-colour palette only, deep navy #14213D and warm mustard #FCA311 on off-white paper stock. Visible halftone grain. Constraints: no gradients, no drop shadows, no photographic texture. Render every text line exactly once, spelled correctly. Style: mid-century screen print, limited ink, flat colour separation.
4. Real estate open house poster
Slop:
Real estate poster for a house
Optimized:
Create a premium real estate open house poster. Purpose: window card for a high-end agency, signalling quality without shouting. Scene: a modern single-storey house exterior at golden hour, photographed from the front garden slightly off-axis. Light: low warm sun from camera right at roughly 5 degrees elevation, long soft shadows across the lawn, warm interior lights visible through glazing. Composition: house occupying the lower two thirds with vertical lines kept perfectly straight, clean sky across the top third for type. Typography: "OPEN HOUSE" in refined spaced uppercase serif, upper centre, warm white. Below it: "Saturday 11am to 2pm". Colour: muted natural palette, 60 percent warm neutrals, 30 percent sky blue-grey, 10 percent brass accent #B08D57. Constraints: no people, no cars, no agency logos, no oversaturated sky, no HDR glow. Keep vertical architectural lines true. Format: A2 portrait at 300dpi, CMYK-safe.
Restaurant Menus
Menus are where I see slop most often, because the text has to be perfect and there is a lot of it.
5. Cafe menu board
Slop:
A cafe menu with prices
Optimized:
Create a photorealistic printed cafe menu card standing upright on a wooden table. Purpose: table-top menu for an independent brunch cafe, warm and handmade rather than corporate. Scene: warm off-white card stock with visible cotton texture, softly lit from a window camera left. Typography, rendered exactly as written: Heading: "MORNING MENU" in spaced uppercase serif. Then five lines in a two-column layout with leader dots: "Sourdough Toast 6" "Baked Eggs 11" "Seasonal Granola 9" "Buttermilk Pancakes 10" "Flat White 4" Composition: heading in the top quarter, items in the middle half, generous margin at the base. Colour: warm cream stock #F5EFE4, near-black ink #1A1714, single terracotta rule #C05B3D beneath the heading. Constraints: every menu line rendered exactly once, correctly spelled, correctly aligned, prices right-aligned. No other text on the card. No logos. No decorative illustrations. Style: real photograph, shallow depth of field, natural window light.
6. Fine dining tasting menu
Slop:
Fancy restaurant menu
Optimized:
Create a photorealistic fine dining tasting menu card. Purpose: placed at each setting in a tasting-menu restaurant, should feel restrained and expensive. Scene: heavy uncoated card on a dark linen tablecloth, single overhead warm light creating a soft pool. Typography, exactly as written and centred: "AUTUMN TASTING" in small spaced uppercase at the top. Then five course lines in title case, each on its own line: "Celeriac, Hazelnut, Truffle" "Cured Trout, Buttermilk, Dill" "Aged Beef, Marrow, Onion" "Blackberry, Sorrel, Cream" "Petit Fours" Composition: extreme vertical centring with very wide margins, at least 25 percent of the card left empty on every side. Colour: warm ivory stock, single ink in deep charcoal. No accent colour. Constraints: no prices, no descriptions, no logos, no decorative flourishes, no border. Each line rendered exactly once with consistent letter spacing. Style: real photograph, low-key lighting, shallow depth of field.
7. Chalkboard specials menu
Slop:
Chalkboard menu for a restaurant
Optimized:
Create a photorealistic hand-lettered chalkboard specials menu mounted on a brick wall. Purpose: daily specials board inside a neighbourhood bistro, should look genuinely hand-drawn rather than digitally faked. Scene: a real slate board with visible chalk dust, faint ghosting from previous wipes, slightly uneven surface. Light: warm tungsten from above camera left, subtle sheen across the slate showing its texture. Typography, in white chalk, exactly as written: "TODAY'S SPECIALS" in hand-drawn uppercase at the top "Wild Mushroom Risotto 16" "Pan Roast Cod 19" "Sticky Toffee Pudding 8" Details: slight variation in stroke pressure between letters, minor imperfections, a small hand-drawn underline flourish beneath the heading only. Constraints: no other text, no chalk illustrations, no perfect uniformity in the lettering. Every price and dish rendered exactly once and correctly spelled. Style: real photograph, natural imperfection, visible chalk texture.
YouTube Thumbnails
YouTube's official spec is 3840x2160 at 16:9, under 2MB on mobile. YouTube itself recommends the rule of thirds.
8. Tutorial thumbnail
Slop:
YouTube thumbnail about AI with a surprised person
Optimized:
Create a 16:9 YouTube thumbnail for a software tutorial. Purpose: must be legible as a 320x180 pixel preview on mobile, so every element needs to survive an 8x downscale. Composition: subject on the right third per the rule of thirds, text block occupying the left two thirds, nothing important in the bottom-right corner where the duration stamp overlays. Subject: a person in their thirties, genuine surprised expression, upper body, looking down at a laptop screen that casts light upward onto their face. Light: strong screen glow from below at 6500K as the key, warm rim light from behind camera left separating them from the background. Typography: "IT ACTUALLY WORKED" in heavy condensed uppercase, three stacked lines on the left, bright yellow #FFD400 with a 6px dark outline for legibility against any background. Colour: 60 percent deep navy background #0B1B33, 30 percent subject and skin tones, 10 percent yellow type only. Constraints: render text exactly once, correctly spelled. No logos, no borders, no arrows, no red circles. Keep the face fully visible and unobstructed. Format: 3840x2160.
9. Listicle thumbnail
Slop:
Thumbnail for a top 10 video
Optimized:
Create a 16:9 YouTube thumbnail for a ranked list video. Purpose: competing in a crowded sidebar, needs one instantly readable number and one clear object. Composition: large numeral occupying the left 40 percent, single product object on the right, clean separation between them. Subject: a matte black wireless headphone floating at a slight angle, single strong specular highlight along the headband. Light: soft large key from upper left, deep shadow side retained for contrast, subtle rim light on the right edge. Typography: the numeral "10" rendered very large in heavy uppercase, off-white, occupying most of the left panel. Below it in much smaller type: "TESTED". Colour: 60 percent charcoal gradient background, 30 percent product, 10 percent electric cyan #00E5FF as rim accent only. Constraints: exactly two text elements and nothing else. No faces, no arrows, no logos, no starburst shapes. Both text elements rendered once and correctly. Format: 3840x2160.
10. Face-led vlog thumbnail
Slop:
Vlog thumbnail with my face
Optimized:
Create a 16:9 YouTube thumbnail with a strong single-subject face. Purpose: personality-led channel, the face is the brand, so the expression must read at thumbnail scale. Composition: face occupying the right half, cropped from mid-chest, eyes positioned on the upper-third line. Left half reserved for three short words. Subject: a person with a clear, readable expression of genuine disbelief, mouth slightly open, direct eye contact with camera. Light: bright soft key from front left at eye level for even facial illumination, plus a coloured rim from behind right. Typography: "I WAS WRONG" in bold uppercase, stacked left, off-white with a subtle dark shadow for separation. Colour: 60 percent deep purple background #2A1B47, 30 percent skin and clothing, 10 percent warm orange rim light #FF7A1F. Constraints: keep the eyes sharp and fully visible. No sunglasses, no hands covering the face, no additional text, no logos. Render the text exactly once. Format: 3840x2160.
Social Ads
Meta's current spec for feed images is 4:5 at 1440x1800, with primary text between 50 and 150 characters.
11. Product launch ad
Slop:
Instagram ad for a skincare product
Optimized:
Create a 4:5 social feed advertisement for a minimalist skincare brand. Purpose: cold-audience awareness ad, must communicate premium positioning in under two seconds of scroll. Scene: a single frosted glass serum bottle standing on a pale travertine ledge. Light: soft directional daylight from camera left at a low angle, producing a long soft shadow to the right and a single soft highlight down the bottle edge. Composition: product in the lower-right third, upper-left 40 percent left as clean negative space for type. Typography: "Less, but better." in a light-weight serif, upper left, warm charcoal. No other copy in the image. Colour: 60 percent warm sand #E8DCCB, 30 percent travertine texture, 10 percent deep clay accent #8A5A44. Constraints: no model, no hands, no water splashes, no ingredient callouts, no badges, no starbursts. No visible brand logo. Render the tagline exactly once. Format: 1440x1800.
12. Sale promotion ad
Slop:
Sale ad with 50% off
Optimized:
Create a 4:5 social feed sale advertisement for a clothing brand. Purpose: retargeting ad for warm audiences, discount must be the single dominant element. Scene: flat graphic composition on a solid ground, no photography. Composition: the offer centred in the upper two thirds at large scale, supporting line beneath, generous even margins. Typography: "50% OFF" in very heavy condensed uppercase, centred. Beneath it in much smaller regular weight: "Everything. Ends Sunday." Colour: 60 percent off-black #101010, 30 percent off-white type, 10 percent acid green #C8FF00 applied only to the percentage symbol. Constraints: exactly two text elements, each rendered once, correctly spelled. No products, no models, no confetti, no starbursts, no urgency icons, no countdown graphics. Format: 1440x1800.
13. Testimonial ad
Slop:
Customer review ad
Optimized:
Create a 4:5 social feed testimonial advertisement. Purpose: social proof ad, quote must be the hero and must feel like a real review rather than marketing copy. Scene: clean single-colour ground with a subtle paper texture, no photography. Composition: quote centred in the middle 60 percent, attribution line beneath it with clear separation, wide even margins on all four edges. Typography: the quote "I stopped looking after this one." in a medium-weight serif at large size, centred across three lines. Beneath it in small caps: "Sarah K., verified buyer". Colour: 60 percent warm off-white #F7F3EC, 30 percent deep ink type #1C1A17, 10 percent muted gold rule #B79355 as a short divider only. Constraints: exactly two text elements. No star icons, no quotation mark graphics, no product images, no logos, no photo of a person. Render each element once and spell correctly. Format: 1440x1800.
Packaging And Product
14. Coffee bag packaging
Slop:
Coffee packaging design
Optimized:
Create a photorealistic standing coffee pouch mockup. Purpose: shelf presence in a specialty grocery, must look premium at arm's length and legible at half a metre. Scene: a matte kraft foil pouch standing on a light concrete surface against a soft neutral backdrop. Light: soft large key from upper left, gentle gradient falloff across the pouch face, soft contact shadow at the base, no harsh speculars on the matte surface. Composition: pouch centred, occupying the middle 60 percent vertically, symmetrical framing. Typography on the pouch, exactly as written: "ORBIT" in bold spaced uppercase across the upper third "Single Origin" in small caps beneath it "Ethiopia Guji 250g" in small type near the base Colour: 60 percent natural kraft #C9A87C, 30 percent deep charcoal print #22201D, 10 percent burnt orange accent #D45D2E on a single thin rule. Constraints: all pouch text rendered exactly once, correctly spelled, correctly aligned to the pouch surface curvature. No barcodes, no certification badges, no additional text. Style: real photograph, commercial packaging photography.
15. Cosmetic product hero
Slop:
Product photo of a cosmetic bottle
Optimized:
Create a photorealistic hero product photograph for an e-commerce listing. Purpose: primary listing image, must show the product accurately with no ambiguity about shape or finish. Scene: a seamless pale grey sweep, product centred and upright. Subject: a frosted glass dropper bottle with a matte black cap, slight three-quarter rotation. Light: large softbox key from upper left producing a soft gradient down the bottle, white bounce card camera right lifting the shadow side, controlled specular highlight along the left edge only. Composition: product occupying 70 percent of the frame height, centred, with even margin all round. Constraints: no props, no water droplets, no ingredients scattered around, no text, no logos, no reflection floor. Keep the bottle silhouette clean and the cap edges sharp. Style: real photograph, commercial product photography, focus stacked so the entire product is sharp.
Signage And Print Collateral
16. Trade show banner
Slop:
Trade show banner for a tech company
Optimized:
Create a roll-up trade show banner design, 850x2000mm portrait. Purpose: read from six metres in a crowded hall, one message only. Composition: message occupying the upper 40 percent at eye height, supporting line in the middle, clean lower third since that area is blocked by people standing in front of it. Typography: "SHIP IN DAYS, NOT QUARTERS" in heavy condensed uppercase across three lines. Beneath in medium weight: "Stand D42". Colour: 60 percent deep indigo #1B1F3B, 30 percent off-white type, 10 percent electric lime #B8FF3C on a single underline rule. Details: flat graphic treatment, no photography, subtle diagonal texture at 5 percent opacity in the background only. Constraints: nothing critical below 600mm from the base. Cap height minimum 60mm on the headline for six-metre legibility. No logos, no icons, no additional copy. Render each line exactly once. Format: 850x2000mm at 150dpi with 30mm bleed.
17. Business card
Slop:
Business card design
Optimized:
Create a photorealistic business card mockup, front face only. Purpose: architecture practice, should signal restraint and craft. Scene: a single card resting at a slight angle on a raw concrete surface, shot from directly above. Light: soft even daylight from camera left, very subtle shadow under the card edge showing its thickness. Typography, exactly as written and left-aligned in the lower half: "MERIDIAN" in spaced uppercase at the top of the block "Architecture & Interiors" in small italic beneath "hello@meridian.studio" in small regular at the base Colour: uncoated warm white stock #F4F1EA, single ink in deep charcoal #2B2A28. No accent colour. Details: visible uncoated paper texture, slightly deckled edge, 420gsm thickness visible at the corner. Constraints: no logo mark, no phone number, no address, no QR code, no additional text. Render every line exactly once. Style: real photograph, top-down product photography.
18. Tri-fold brochure cover
Slop:
Brochure cover for a wellness company
Optimized:
Create a tri-fold brochure cover panel, 99x210mm. Purpose: picked up from a clinic reception desk, must earn the open in one glance. Scene: flat design with a single duotone photographic element in the lower half. Subject: a pair of hands resting open in a lap, cropped so no face is visible, rendered as a warm duotone. Composition: type in the upper 45 percent, image bleeding off the bottom edge, clean horizontal division between them. Typography: "Room to breathe." in a light serif, sentence case, upper left. Small caps beneath: "Six week programme". Colour: 60 percent warm off-white #F6F2EC, 30 percent duotone image in sage #7C8B72, 10 percent deep forest #2F3D2C on type. Constraints: no clip art, no icons, no logos, no stock-photo smiling faces, no gradients. Render each text element once. Format: 99x210mm at 300dpi with 3mm bleed, safe margin 3mm inside trim.
Digital And Web
19. Email header banner
Slop:
Email header image
Optimized:
Create an email header banner, 1200x400 pixels. Purpose: top of a monthly newsletter, must render well in dark mode and at 600px scaled width. Scene: flat abstract composition, no photography. Subject: a simple horizontal arrangement of overlapping translucent geometric shapes, weighted to the right. Composition: shapes across the right 45 percent, type left-aligned in the left 45 percent, generous vertical centring. Typography: "The Monthly Build" in medium-weight sans-serif, left aligned, single line. Small caps beneath: "Issue 14". Colour: 60 percent deep slate #1E2430, 30 percent muted teal shapes #3E7C77, 10 percent warm amber #E9A23B on one shape only. Constraints: minimum 4.5:1 contrast between type and background for accessibility. No photographic elements, no logos, no buttons, no additional text. Render each line exactly once. Format: 1200x400 pixels.
20. Blog featured image
Slop:
Blog header image about productivity
Optimized:
Create a 16:9 blog featured image. Purpose: article header for a piece on focused work, appears both full-width and as a 400px card thumbnail. Scene: a top-down desk composition, deliberately sparse. Subject: a single closed notebook, one pen laid at an angle, and an empty ceramic cup, arranged asymmetrically on a warm oak surface. Light: soft directional daylight from camera left, soft shadows extending right, no harsh contrast. Composition: objects clustered in the lower-left third, upper right two thirds intentionally empty for headline overlay later. Colour: 60 percent warm wood tones, 30 percent off-white paper, 10 percent single muted blue on the pen only. Constraints: no laptop, no phone, no plants, no coffee art, no text, no logos. Keep the composition sparse with at least half the frame empty. Style: real photograph, natural light, subtle imperfection in object placement.
Retail And Promotional
21. Instagram carousel cover
Slop:
Carousel cover about marketing tips
Optimized:
Create a 4:5 Instagram carousel cover slide. Purpose: first slide of an educational carousel, must stop the scroll and promise a specific payoff. Scene: flat design on a solid warm ground with subtle paper grain, no photography. Composition: headline occupying the centre 55 percent across three lines, small swipe cue at the base, wide even margins. Typography: "5 mistakes killing your open rates" in a bold serif, sentence case, centred. At the bottom in small caps: "Swipe". Colour: 60 percent warm cream #F4EEE2, 30 percent deep ink type #1B1815, 10 percent burnt orange rule #C4541F beneath the headline only. Constraints: exactly two text elements, each rendered once and spelled correctly. No icons, no arrows, no photos, no logos, no profile picture. Format: 1440x1800.
22. Billboard
Slop:
Billboard ad for an insurance company
Optimized:
Create a 48-sheet billboard design, 6096x3048mm landscape. Purpose: read in under three seconds by a driver at 50kph, so one idea and nothing else. Composition: message occupying the left 60 percent, single simple graphic element on the right, nothing within 200mm of any edge. Typography: "Cover that actually covers." in heavy sans-serif across two lines, left aligned. Nothing else. Subject: one simple flat graphic of an open umbrella, minimal detail, right side. Colour: 60 percent deep navy #16233A, 30 percent off-white type, 10 percent warm yellow #FFC145 on the umbrella only. Constraints: maximum seven words total in the image. No phone number, no website, no small print, no logos, no photography, no gradients. Cap height minimum 400mm. Format: 6096x3048mm at 30dpi viewing-distance resolution.
23. Event flyer
Slop:
Flyer for a food festival
Optimized:
Create an A5 street flyer for a weekend food market. Purpose: handed out on the street, read in five seconds, kept or binned immediately. Scene: flat graphic composition with one duotone photographic element. Subject: an overhead crop of hands passing a paper-wrapped parcel across a stall counter, rendered as a warm duotone, no faces visible. Composition: image bleeding off the top edge across the upper 45 percent, type block in the lower half, clear horizontal division. Typography: "STREET FEAST" in condensed uppercase. Beneath: "Saturday & Sunday, 11am to 8pm" then "Canal Yard, Unit 4". Colour: 60 percent warm off-white #F7F1E6, 30 percent duotone in deep plum #4A2545, 10 percent acid yellow #E8DB2F on one rule. Constraints: three text elements only, each rendered once and spelled correctly. No prices, no vendor logos, no QR code, no clip art. Format: 148x210mm at 300dpi with 3mm bleed.
24. Product label
Slop:
Label design for a hot sauce bottle
Optimized:
Create a photorealistic hot sauce bottle with a wrapped paper label. Purpose: shelf product in a specialty food store, artisanal rather than mass-market. Scene: a clear glass bottle on a dark slate surface against a soft dark backdrop. Light: single hard key from camera right creating a bright specular line down the glass edge, deep shadow to the left, subtle rim from behind to define the bottle silhouette. Typography on the label, exactly as written: "EMBER" in bold condensed uppercase across the centre "Smoked Habanero" in small italic beneath "150ml" in small type at the base Details: uncoated paper label with visible fibre texture, slightly imperfect application with one soft crease. Colour: 60 percent deep charcoal scene, 30 percent warm cream label #EFE3CE, 10 percent burnt red #A32C1E on the type. Constraints: all label text rendered exactly once, correctly spelled, wrapping naturally around the bottle curvature. No barcode, no nutrition panel, no certification badges. Style: real photograph, dramatic product lighting.
25. Gift voucher
Slop:
Gift card design for a salon
Optimized:
Create a photorealistic printed gift voucher card lying flat. Purpose: sold at reception in a hair salon, should feel like a considered gift rather than a receipt. Scene: a single card on a pale marble surface, shot from directly above. Light: soft even daylight from camera left, very subtle shadow along the card edge indicating thickness. Typography, exactly as written and centred: "GIFT VOUCHER" in small spaced uppercase at the top "Fifty Pounds" in a large light serif in the centre "Valid twelve months from purchase" in very small type at the base Colour: 60 percent soft blush stock #EDDCD6, 30 percent deep charcoal type #262220, 10 percent brushed brass foil #B08D57 on the top line only. Details: 350gsm uncoated stock with visible texture, foil line catching the light slightly. Constraints: three text elements only, each rendered once and spelled correctly. No salon logo, no barcode, no terms paragraph, no decorative flourishes. Style: real photograph, top-down, soft natural light.
The Editing Layer: Fixing Slop You Already Made
I reach for these before regenerating, because a repair is usually cheaper than another roll of the dice. They use [attached image] since they operate on something you upload.
26. Strip the plastic look
Edit [attached image] to remove the artificial smoothness. Changes: restore realistic skin texture including visible pores and fine lines. Reintroduce natural micro-contrast in fabric weave and surface materials. Add subtle asymmetry to lighting so one side falls off more than the other. Constraints: keep the composition, framing, pose and colour palette exactly as they are. Do not change the subject's identity, proportions or expression.
27. Fix the oversaturation
Colour grade [attached image] to a professional standard. Changes: reduce global saturation by roughly 20 percent, pull the highlights back from clipping, lift the blacks very slightly for a filmic rolloff, and add a subtle cool tint to the shadows while keeping skin tones neutral. Constraints: do not alter composition, content or sharpness. Keep all colours within CMYK-printable gamut. Do not introduce a heavy orange and teal split.
28. Replace generated text with clean space
Edit [attached image] to remove all text. Changes: remove every piece of rendered text and fill the space naturally with the surrounding surface, texture and lighting. Constraints: keep the composition, subject, lighting and colour exactly as they are. Leave the cleared areas clean and even so real type can be set over them later.
That third one is the single most useful edit in this whole article. Generate the layout with text removed, then set real type in Figma, Canva or InDesign. It fixes the one problem the model structurally cannot solve.
29. Add brand colour accuracy
Edit [attached image] to match brand colours exactly. Changes: shift the primary accent colour to #C05B3D and the secondary to #2F3D2C, applying them only to the elements that currently carry accent colour. Constraints: keep all other colours, the composition, lighting and subject unchanged. Do not shift skin tones or neutral greys.
30. Rescue a flat composition
Edit [attached image] to improve the lighting direction. Changes: relight so the key light comes from camera left at roughly 45 degrees, creating visible falloff and a defined shadow side. Add a subtle rim light on the opposite edge to separate the subject from the background. Constraints: keep the subject, pose, composition, framing and colour palette identical. Do not change the background content.
31. Remove the AI giveaway objects
Review [attached image] and correct any physically implausible details. Changes: fix objects that could not function as depicted, including slack or floating straps, hardware attached to nothing, shadows falling in inconsistent directions, and reflections that do not match the scene. Constraints: keep the overall composition, subject, lighting direction and colour palette unchanged. Make only the minimum corrections needed for physical plausibility.
Format Specs You Should Be Prompting With
Half of looking professional is just using the right dimensions, and I get these wrong more often than I would like. These are the current official numbers.
| Asset | Spec | Source |
|---|---|---|
| YouTube thumbnail | 3840x2160, 16:9, under 2MB mobile | YouTube Help |
| YouTube Shorts | 2160x3840, 9:16 | YouTube Help |
| Meta / Instagram feed | 1440x1800, 4:5, 30MB max | Meta Ads Guide |
| Stories and Reels | 9:16 | Meta |
| Print resolution | 300dpi, from 150lpi x 2 | PrintNinja |
| Standard bleed | 3mm / 0.125in per edge | Print convention |
| Large format bleed | 0.25in per edge | Print convention |
| Poster A2 | 420x594mm | ISO 216 |
| Poster A1 | 594x841mm | ISO 216 |
| US movie one-sheet | 27x40in | Industry standard |
| Text contrast minimum | 4.5:1 normal, 3:1 large | W3C WCAG 2.2 |
| Minimum print body type | 8pt | Print convention |
| Ink coverage limit | 300% coated, 240% newsprint | Cummings Printing |
The Lighting Vocabulary That Actually Works
Naming a real lighting setup did more for my output than any quality adjective ever has. Here are the ones I keep going back to.
| Setup | What it produces | Best for |
|---|---|---|
| Three-point | Key plus fill plus backlight, clean separation | Corporate, interview, general commercial |
| Rembrandt | Illuminated triangle on the shadow cheek | Dramatic portraits |
| Butterfly | Symmetrical shadow under the nose | Beauty, glamour |
| Loop | Small nose shadow looping onto the cheek | Default flattering portrait |
| Split | Face divided exactly in half | Hard, cinematic |
| Short | Key on the side away from camera | Slims the face |
| Broad | Key on the side toward camera | Widens the face |
| Rim / kicker | Backlit edge separating subject from background | Thumbnails, product |
| Softbox key plus bounce | Even gradient, controlled speculars | Product, packaging |
| Backlight at 45 degrees | Translucency and texture | Food, drinks |
Golden hour is roughly +6 to −4 degrees of sun elevation. Blue hour is −4 to −6. Naming those beats writing "beautiful lighting" every time.
What The Providers Themselves Say
I found it useful that the model makers openly document this problem, and that they contradict each other on how to fix it.
OpenAI's image prompting guide is the most direct. Its own example prompts include instructions to avoid:
"clip art, stock photography, gradients, shadows, decorative elements, or anything that feels generic or overdesigned."
And it tells you to specify "no generic stock-photo treatment" outright.
But here is where they disagree, and it matters for which tool you use:
| Question | OpenAI | Midjourney | FLUX | Ideogram | |
|---|---|---|---|---|---|
| Prompt length | Structured and detailed | "Be hyper-specific" | "Short and simple prompts typically generate the best images" | "Start short. Add only what changes the image" | Under ~150 words |
| Negative prompts | In-prompt only | Rephrase positively | Yes, --no parameter | Not supported | Rephrase positively |
| Camera specs | "cues for appearance, not a guarantee" | Documented and encouraged | Not documented | Name the body, lens and film stock | Not documented |
| Naming a typeface | Describe properties | Describe properties | Not documented | Describe properties | "not currently possible" |
I learned this the hard way: a prompt I tuned for FLUX underperformed badly on Midjourney. FLUX explicitly wants "Shot on Fujifilm X-T5, 35mm f/1.4." Midjourney documents no lens vocabulary at all.
One documented technique worth stealing regardless of tool, from Google's guide:
"Use 'semantic negative prompts': Instead of saying 'no cars,' describe the intended scene positively: 'an empty, deserted street with no signs of traffic.'"
What People Are Building And Arguing About
I went through X to see how working creators are handling this, and the debate has settled on one word: art direction.
The clearest articulation of the divide, and it applies directly to marketing work:
https://x.com/LeonMandala/status/2097443516523659689
His point is that "no aesthetics and no art direction is for everyone," meaning the failure predates AI and AI just made it faster.
The distinction people keep defending, that realistic output is not automatically slop when there is direction behind it:
https://x.com/Ablurman/status/2097350221617307830
The best practical example I found of specification beating generation. Note the level of visual-style detail in the prompt, right down to naming a camera era:
https://x.com/SimplyAnnisa/status/2097914538544992645
Her framing is the thesis of this entire article: "AI slop is easy. Making AI feel like a real memory is the hard part."
On deliberate imperfection, which is the counterintuitive lever most people miss:
https://x.com/Strength04_X/status/2096498079717621933
The observation that it works because there is "no music, no perfect shots" is worth sitting with. Polish is often what marks something as generated.
And the backlash is real, which is the commercial risk if you get this wrong:
https://x.com/AkuDennis/status/2097922373169406378
There is also tooling emerging specifically for this. An anti-AI-slop design skill for Claude and Codex is circulating in developer tool roundups:
https://x.com/csaba_kissi/status/2097939657015394482
My Honest Take
I want to be straight about what actually moved the needle for me, because it was not what I expected.
What made the biggest difference:
What did not help as much as I expected:
The honest limitation: none of this makes AI good at typography. The tokenization problem is structural, not a prompting skill issue. Every genuinely professional piece I have produced involved generating the visual and setting the type separately.
Frequently Asked Questions (FAQs)
What exactly is AI slop?
Merriam-Webster, which made "slop" its 2025 Word of the Year, defines it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence." In practical terms for marketing work, it means output with no art direction: no specified lighting, no composition plan, no brand system and no constraints.
Why do my AI images look fake even when the quality is high?
Usually lighting and physics rather than resolution. Research from Hany Farid shows generation "is not based on explicit geometric or physical models," so shadows and reflections often do not agree with each other. Naming your light source direction, quality and colour temperature fixes most of it.
Why is text in AI images always slightly wrong?
Because the text encoders tokenize at word or subword level, so glyph-level information never reaches the model. Generated type has no kerning tables, no baseline grid and no hinting. The reliable fix is to generate the image without text and set real type in a design tool afterwards.
How long should an image prompt be?
It depends entirely on the model, and the providers openly disagree. Midjourney documents that "short and simple prompts typically generate the best images." Google says "be hyper-specific." Ideogram caps its recommendation at around 150 words. FLUX advises starting short and adding only what changes the image.
Do negative prompts work?
Only on some models. Midjourney has a --no parameter. FLUX states plainly that most of its models do not support negative prompts. Google and Ideogram both recommend rephrasing positively instead, describing the scene you want rather than what you want removed.
Will using AI images hurt my reach on social platforms?
It can if the output is generic. YouTube's July 2025 policy update demonetizes "inauthentic content," which it defines partly as work made with "generic or unoriginal templates giving the impression of mass production." Meta and Pinterest have similar policies. Directed, original work is treated differently from mass-produced output.
Final Thoughts
The thing I keep coming back to is that slop is a specification failure, not a technology failure. The same model that produces the left column above produces the right column, and the only variable is how much direction you gave it.
If you take one habit from this piece, make it the seven layers. Subject, purpose, light, composition, typography, brand, constraints. A prompt missing five of those will look like everyone else's.
And if you take one workflow change, make it this: generate the image with no text, then set the type properly. That single split solved more of my problems than every other tip combined.
The full prompt library from this article, plus everything else I use, is on promptslove.com.
