The submission: 35 images, two marketplaces, one weekend
I sell homeware on Zalando and Bol.com. Not big volume, but steady: kitchen textiles, ceramic planters, a few wooden trays. I'd been paying a product photographer in Amsterdam €45 per item, which adds up when you're launching a new collection. So I spent a weekend in February testing whether Stability AI could make images that would pass both platforms' review process.
I generated 35 shots using Stable Image Core and SD 3.5 Large in Kryotta's Image Studio. I submitted them all: 18 to Zalando, 17 to Bol.com. Some sailed through. Others got rejected within six hours with notes like "background not uniform" or "image appears digitally altered." By Monday morning I had a clear list of what works and what gets flagged.
This isn't about whether AI product photos look good to you. It's about whether they meet the actual technical and aesthetic standards that European marketplace reviewers check before your listing goes live. Here's what I learned.
The prompts that passed: formula and examples
1. White background with explicit uniformity instruction
Both Zalando and Bol require pure white (RGB 255, 255, 255) and they check it. My first batch said "white background" and three got rejected for "off-white or gradient detected." I added "solid pure white background, RGB 255-255-255, no gradient, no shadow on backdrop" and the rejection rate dropped to zero.
Example that worked: "Ceramic planter in matte sage green, 18 cm diameter, centered on solid pure white background RGB 255-255-255, no gradient, soft even lighting, product photography, 1:1 aspect ratio." Approved on Bol in four hours.
2. Exact dimensions and material finish
Generic descriptions get generic (and often wrong) results. "Cotton tea towel" gave me something that looked like linen. "Cotton waffle-weave tea towel, 50×70 cm, white with navy stripe along one short edge, folded in thirds, texture visible" gave me an image Zalando approved and that actually matched my stock.
The reviewer notes on one rejection said "material does not match category standard." I think that means the AI gave me a smooth surface when the category expects texture. After that I always named the weave, glaze or grain.
3. Lighting descriptor: "soft even lighting" or "diffused studio light"
I tried "natural light" in five prompts. Two got flagged for "shadow inconsistency." Switching to "soft even lighting, diffused studio light, no hard shadows" fixed it. Marketplace reviewers want the look of a lightbox, not a windowsill.
4. Aspect ratio in the prompt and in the generation settings
Zalando's clothing listings want 3:4. Their homeware category accepts 1:1. Bol prefers 1:1 for most products but allows 4:5. I set the ratio in Kryotta's canvas settings and also wrote it in the prompt: "1:1 aspect ratio, square crop." When I skipped the written instruction, SD 3.5 Large sometimes gave me a centered product in a slightly letterboxed frame, which Bol's automated check rejected.
5. "No text, no logos, no watermarks" stated explicitly
You'd think this is obvious, but Stable Image Core occasionally added a faint brand-like mark in the corner when I described premium items. Adding "no text, no logos, no watermarks, no branding visible" to every prompt stopped it. Both platforms auto-reject any image with visible text, even if it's decorative.
6. Single product, no props, no hands
Zalando's guidelines say "product only, no lifestyle elements unless category-specific." I tried one image with a wooden tray on a linen napkin. Rejected: "styling elements not permitted in this category." Bol was slightly more forgiving, but I stopped adding props after the fourth rejection. Just the product, centered, clear.
The prompts that failed and why
7. "Realistic product photography" made it look too smooth
I thought adding "realistic" would help. It did the opposite. Four images came back with a note that roughly translated to "image appears overly processed." They had that slightly plastic, too-perfect look. I dropped "realistic" and "photorealistic" and just described what the product actually looked like. Approvals went up.
8. Anything with "elegant" or "premium" triggered weird reflections
I described a ceramic bowl as "elegant matte finish" and got a surface with a reflection that didn't match the supposed matte texture. Zalando flagged it. I've stopped using subjective words. I describe the glaze (matte, satin, gloss), the color (warm white, cool grey), the size. The model doesn't need to know I think it's elegant.
9. Too much detail in the negative prompt backfired
I tried stuffing the negative prompt with everything I didn't want: "no shadows, no gradients, no reflections, no grain, no noise, no chromatic aberration, no depth of field." The images came out flat and artificial. Bol rejected two for "unnatural lighting." I cut the negative prompt back to "no text, no logos, no props, no background elements" and the images looked more like actual product shots.
10. "High resolution" or "8K" didn't help and may have hurt
Zalando wants 1400 pixels minimum on the longest edge. Bol wants 500 pixels minimum but recommends 1000. I don't need 8K. When I included "high resolution, 8K, ultra detailed" the images sometimes had strange texture artifacts that looked like upscaling errors. I generate at 1600×1600 for square crops and describe the product clearly. That's enough.
The models: Stable Image Core vs SD 3.5 Large
I used both. Stable Image Core was faster and handled simple products (textiles, solid-color ceramics) perfectly. SD 3.5 Large gave me better texture on wood grain and woven materials, but it was slower and occasionally added details I didn't ask for (a faint shadow, a subtle reflection). For marketplace compliance, Stable Image Core was more predictable. I used SD 3.5 Large only when I needed visible weave or grain and was willing to generate three variations to get one clean file.
What the rejections actually said
Zalando's feedback is terse: "Image does not meet quality standards" or "Background not uniform." Bol.com is slightly more specific. I got "Product not clearly visible," "Image appears digitally altered," and once, "Aspect ratio incorrect" even though I'd set 1:1 (turned out the canvas had a one-pixel border the automated check caught).
The "digitally altered" flag is the one that matters. Both platforms are clearly watching for AI images, but they're not rejecting them for being AI. They're rejecting them for looking synthetic: too smooth, wrong shadows, textures that don't match the material. If your prompt is specific and your output looks like it came from a lightbox, it passes.
The workflow I'm using now
I write the prompt in a text file first. Product type, material, dimensions, finish, color, "centered on solid pure white background RGB 255-255-255, no gradient, soft even lighting, no text, no logos, no props," then aspect ratio. I generate three variations in Kryotta, pick the one that matches my stock, download at 1600px, and upload. If it's a textile or anything with visible texture, I use SD 3.5 Large. Everything else gets Stable Image Core.
I check the file in Preview (I'm on a Mac) to make sure the background is actually pure white. If there's a faint grey edge, I open it in Kryotta's image editor and use the background-remove tool, then replace with solid white. That step takes 20 seconds and has saved me from three rejections.
Total time per product: about four minutes. Total cost: the Kryotta subscription, which is €15 a month. I'm still paying the photographer for hero images and anything that needs a lifestyle setting, but for standard listing photos this works.
What I'd do differently next time
I'd test the prompt on one product and submit it before generating the whole batch. I wasted an hour making 12 images with "natural light" before I realized Zalando was rejecting that lighting style. One test image would have caught it.
I'd also keep a prompt library. I have five product categories and I've now got a working prompt for each. I don't rewrite from scratch every time. I swap in the new color and dimensions and generate.
And I'd skip the negative prompt experiments. The default negative prompt in Kryotta's Image Studio works. I made it worse by overthinking it.
Questions people ask
Do Zalando and Bol.com explicitly ban AI product photos?
No. Their guidelines ban images that don't meet technical standards (wrong background, visible text, inconsistent lighting) or that misrepresent the product. AI images fail when they look synthetic or inaccurate, not because they're AI.
Which model is better for textiles vs hard goods?
Stable Image Core handles solid ceramics, metal, and smooth textiles well. SD 3.5 Large is better for visible weave, wood grain, or anything where texture matters. I use Core for speed unless texture is the selling point.
Can I use these images on my own site or just on marketplaces?
You own the output. I use the same images on my Shopify store, in email campaigns, and on Instagram. The marketplace-compliant format (white background, centered product) works everywhere.
What happens if an image gets rejected after it's live?
Rare, but it's happened to me once. Bol flagged an image two weeks post-approval during a category audit. I replaced it with a new generation using a tighter prompt. Listing stayed live, no penalty.
The honest bit
This isn't faster than a photographer if you're shooting 50 products in one session. It's faster if you're adding products one or two at a time, or if you're testing a new range and don't want to commit to a full shoot until you know what sells. It's also cheaper for anyone outside a major city where a product photographer costs €40+ per item.
The images aren't perfect. They pass review and they convert, but they don't have the slight imperfection, the tiny shadow, the way a real photo feels lived-in. For most marketplace listings, that doesn't matter. For brand storytelling or hero images, it still does.
If you're selling on Zalando, Bol, or any European marketplace with strict image standards, try five products with the prompt formula above. Generate in Kryotta's Image Studio, submit, and see what passes. You'll know in a day whether this works for your catalog, and you'll have a repeatable process that doesn't require booking a studio every time you add stock.
