Use DALL·E 3 to generate usable product photos (a 30-minute workflow)

Most AI product photos look obviously fake. The reason is rarely the model — it is the prompt. With a structured prompt, lighting keywords, and a 5-minute upscaling pass, DALL·E 3 can produce images that are usable for indie e-commerce. Here is the workflow I use to ship a small product line in 30 minutes.

What you'll learn

• The 4-part prompt structure that gives DALL·E 3 realistic product photos
• The lighting keywords that separate "AI fake" from "studio quality"
• How to control background and angle without Photoshop
• A 5-minute upscaling/cleanup pipeline

Why most AI product photos look fake

The model is not the problem. The prompt is. Three things cause the "AI fake" look:

1. Vague lighting. "Well lit" produces the model's default diffuse lighting, which reads as "rendered in a 3D engine."
2. No camera metadata. A real product photographer shoots at f/8, 100mm macro, ISO 100. DALL·E 3 does not have a "lens" by default, but you can describe the camera context in the prompt.
3. No background specification. The default is a seamless white or grey background, but a describable background reads as intentional and helps the model commit to a coherent scene.

Fix all three and the output changes dramatically.

The 4-part prompt structure

For every product photo, structure the prompt as:

1. Subject — the product, in plain language, with key materials called out.
2. Camera — focal length, aperture, distance. "Shot on 100mm macro at f/8, camera 60cm from the subject."
3. Lighting — at least one specific lighting term ("three-point softbox setup," "Rembrandt lighting," "golden hour side light," "overcast window light").
4. Background — concrete description, not "minimal." ("Light grey seamless paper, slight gradient to the corners," "white marble countertop with subtle veins," "raw oak table, soft shadow to the right.")

Example — skincare bottle on marble

A 50ml amber glass skincare bottle with a matte black dropper cap, label
facing camera, sitting on a white marble countertop with subtle grey
veins. Shot on 100mm macro at f/8, camera 50cm from the subject. Three-
point softbox lighting with the key light at 45 degrees camera-left. The
marble is slightly out of focus. --ar 1:1

The same prompt with vague lighting ("well lit, on a counter") gives you the AI-fake look. The structured version gives you something you can ship to an Etsy listing.

The lighting keywords that matter

Here are the lighting phrases I actually use, ranked by reliability:

1. "Three-point softbox lighting, key light at 45 degrees camera-left" — the most reliable, most "studio product" look.
2. "Overcast window light, soft shadows" — best for natural / organic products (food, candles, wood).
3. "Rembrandt lighting" — adds a small triangle of light on the cheek of a person, but works on bottles too.
4. "Golden hour side light" — for lifestyle / outdoor shots.
5. "Hard side light, deep shadows" — for dramatic editorial.

Background control

Three background styles cover 80% of e-commerce needs:

• Seamless paper ("light grey seamless paper, slight gradient") — for catalog shots.
• Surface ("white marble countertop," "raw oak table," "concrete slab") — for lifestyle / artisan.
• Context ("on a wooden shelf next to a folded linen towel," "in a sunlit bathroom near a green plant") — for hero / lifestyle shots.

To force a clean isolated cutout, specify: "subject in sharp focus, background completely out of focus, no other objects." Then in post-processing (see below), the background is easy to remove.

The 5-minute upscaling and cleanup pipeline

DALL·E 3 outputs are typically 1024×1024 or 1024×1792. For e-commerce you usually want 2048×2048 or higher, plus slight cleanup. Here is the 5-minute pipeline:

1. Upscale 2x with Real-ESRGAN (free, local) or Magnific AI (paid, higher quality). I use Real-ESRGAN for most product shots because the textures (bottle labels, fabric weaves) get sharp without looking over-processed.
2. Background removal (if needed) with rembg (free, local) or remove.bg. DALL·E 3 is usually good at clean cutouts if you said "no other objects" in the prompt, so this is often a 10-second step.
3. Color/tone correction in any photo editor. I usually do a 2% saturation boost and a +5 sharpening pass. AI images tend to be slightly desaturated.
4. Save as both PNG and a 70% JPEG for the web listing.

Tools I use

• Real-ESRGAN — local, free, batch mode. realesrgan-ncnn-vulkan -i input.png -o output.png -n realesrgan-x4plus.
• Magnific AI — paid, browser-based, better for skin and hair, overkill for products.
• rembg — rembg i input.png output.png. One line, very accurate.

Gotchas

1. DALL·E 3 cannot reliably do text on signs

This is the single biggest limitation. If your product has a label with text, DALL·E 3 will render plausible-looking but wrong text. For product labels with specific text, you have two options:

• Generate the bottle with a blank label, then add the text in Photoshop.
• Use a different model (FLUX, Ideogram) for the text, composite in Photoshop.

For a single product photo, the Photoshop route is 5 minutes. For a 50-SKU catalog, switch models.

2. Materials can drift

A "brushed aluminum" finish can come out as glossy plastic. A "raw oak" surface can come out as plastic-laminate. Be specific: "brushed 6061 aluminum, fine vertical grain, no fingerprints." The more material detail, the more stable the output.

3. Aspect ratios are limited

DALL·E 3 supports 1:1, 3:2, 2:3 natively. For other ratios, generate square and crop in post.

4. "Stock photography" feel comes from consistency

If you need 12 product photos that look like they belong in the same listing, generate all 12 in the same session, with the same prompt structure, on the same model version. Even slight prompt variation creates visual inconsistency.

A 30-minute end-to-end workflow

For a small product line (say, 6 SKUs):

1. 5 min — Write the master prompt template. Fill in the subject for each SKU.
2. 10 min — Generate 4 candidates per SKU in DALL·E 3 (24 generations total, ~10s each).
3. 5 min — Pick the best 6 (one per SKU), download.
4. 5 min — Run all 6 through Real-ESRGAN upscale + rembg.
5. 5 min — Light Photoshop pass: 2% saturation, +5 sharpening, save as PNG and JPEG.

Total: 30 minutes, 6 final images, no stock photo fees.

FAQ

Can DALL·E 3 put text on a sign correctly?

No. DALL·E 3 still hallucinates text on labels and signs. Use a blank label and add real text in Photoshop, or switch to Ideogram or FLUX for the text-heavy version.

Is this legal for commercial use?

Yes, with caveats. OpenAI's terms grant commercial use of generated images. However, you cannot trademark a purely AI-generated logo, and you may not be able to copyright the image in some jurisdictions. For product photos (not logos), commercial use is generally safe. Consult a lawyer for your specific case.

What if the product has unusual proportions?

Describe proportions in concrete terms ("a 30cm tall bottle, twice as tall as wide, with a 5cm wide cap"). The model is better with concrete numbers than with adjectives.

Can I use the same prompt for DALL·E 3 and Midjourney?

The prompt structure transfers, but Midjourney syntax (--ar, --v 6) does not. Strip Midjourney parameters and re-test lighting keywords — Midjourney is more sensitive to stylistic phrasing, DALL·E 3 is more sensitive to camera/lens metadata.

Why are my product photos all the same color palette?

The model defaults to warm, slightly desaturated tones. To shift the palette, add an explicit color in the prompt: "cool color palette, with a slight teal cast in the shadows" or "warm earth tones, ochre and burnt sienna."

Should I disclose AI generation on my e-commerce listing?

This depends on jurisdiction. The EU AI Act requires disclosure for some categories. The US FTC requires disclosure if the AI image misleads a reasonable consumer. To be safe, add "Image is AI-generated for illustration purposes" to product descriptions. Etsy has specific rules — check their current AI policy.