For Stable Diffusion product photography, decide the shot first (packshot, hero, lifestyle, flat lay or detail), then prompt the camera angle, framing, surface, background and light in plain photo terms, with one product and few props. Stable Diffusion cannot reproduce your exact product, label or logo, so for a real listing photograph the item and generate only the background with inpainting, or cut it out and place it on a generated scene.
Stable Diffusion product photography: what it can and cannot do
Stable Diffusion is good at the parts of a product photo around the product: the surface, the backdrop, the light, the props and the mood. It is not a way to photograph a specific item. The SDXL base 1.0 model card lists among its limitations that it does not achieve perfect photorealism, cannot render legible text, and struggles with compositionality, using ‘a red cube on top of a blue sphere’ as its example. The SD 1.5 card lists the same text and compositionality limits.
In practice that gives you two routes. For concept work, mood boards and ad ideas, prompt the whole scene, including a generic version of the product. For a listing or anything that must show the real item, start from a photo of the product and let the model generate only the background, so the shape, label and logo stay the ones you photographed.
Pick the shot before you write the prompt
Packshot: one product, centered, on a plain seamless background, lit evenly. This is the main image for a store listing, so keep the frame clean and leave a margin around the product.
Hero shot: the product large in the frame, often from a slightly low angle, with stronger light and a darker or colored backdrop. Leave empty space on one side if a headline will go there later.
Lifestyle or in-context shot: the product in use or in its setting, such as a mug on a breakfast table or a bottle on a bathroom shelf. Props tell the story, but each one competes with the product.
Flat lay: a top-down view of the product arranged on a surface with a few related objects. Detail shot: a close-up of a texture, material or feature, with a shallow depth of field.
Composition cues Stable Diffusion understands
Write the composition in the order a photographer sets it up: product, camera angle, framing, surface, background, light, then the final format. Plain photo terms work better than long lists of quality words.
Camera angle: ‘eye-level’, ‘three-quarter view’, ‘slightly low angle’ or ‘top-down overhead view’. Framing: ‘centered’, ‘product fills the frame’, ‘close-up’ or ‘wide shot with empty space on the left’. Surface and background: ‘seamless white backdrop’, ‘white marble countertop’, ‘concrete pedestal’, ‘pastel paper background’.
Keep the scene simple. Because the models struggle with compositionality, an exact arrangement such as ‘a bottle to the left of a box, both on a tray’ often comes out wrong. Use one product per image, two or three props at most, and describe position in simple terms. If a word keeps getting ignored, move it closer to the start of the prompt.
Leave room around the product for the crop. Google Merchant Center's best practice is a product that takes up no less than 75% and no more than 90% of the image, and Shopify recommends a consistent aspect ratio for main images shown side by side. A small margin in the generated image makes both easy to hit later.
Lighting and surfaces for product shots
Name one main light, its direction and its quality: ‘soft diffused light from the left’, ‘softbox lighting’, ‘hard sunlight with long shadows’. Soft light suits cosmetics, textiles and food; hard light gives graphic shadows for shoes, tools and drinks.
Glass, metal and glossy packaging show the light as reflections, so describe them: ‘rim light outlining the bottle’, ‘backlit glass’, ‘reflection on a glossy black surface’. For a matte product, ask for ‘soft shadow under the product’ so it does not look pasted on.
Keep the palette short. One background color and one accent color, such as ‘deep green background, brass accents’, keeps the attention on the product.
PhoneDiffusion generates on-device on iPhone or iPad, with optional cloud models. Download on the App Store →
Product photography prompt examples
Packshot: ‘product photography of a ceramic coffee mug, centered, eye-level, seamless white backdrop, soft even studio lighting, soft shadow under the mug, sharp focus’.
Hero shot: ‘product photography of a perfume bottle on a black glossy surface, slightly low angle, rim light outlining the glass, dark green background, empty space on the left’.
Lifestyle: ‘a canvas tote bag hanging on a wooden chair in a sunny apartment, morning window light from the right, plants in the background, shallow depth of field’.
Flat lay: ‘top-down overhead view of a leather wallet, a watch and sunglasses on a beige linen background, soft diffused light, neat spacing’.
Keep the seed fixed while you test one change, such as the angle or the light, so you can see what that word did; Stable Diffusion seeds explained shows how. For packaging text and logos, leave the surface blank and add the lettering in an editor, as the readable text guide explains.
Keep the real product exact: photograph it, then generate the background
Inpainting is the standard way to change only part of an image. In Diffusers, the mask's white pixels are filled in from the prompt and black pixels are kept. So mask the background, keep the product black, and prompt only the new scene, such as ‘on a marble countertop, soft morning light’.
Check the edges and the product itself afterwards. The Diffusers guide notes that inpaint-specific checkpoints blend more naturally but are more likely to change the unmasked area, and it offers apply_overlay() to force the unmasked pixels to stay the same, at the cost of a less natural transition. On desktop, a ControlNet with a canny edge image adds another way to hold the product's outline while the prompt changes the rest.
The other route is a cutout: remove the background from the product photo, keep the result as a PNG with transparency, and place it on a generated background in an editor. Add a soft shadow under the product so it sits on the surface.
Before you upload to a store
Check the marketplace's image rules. Google Merchant Center requires an unobstructed, clean view of the item, with no added text, promotional graphics, watermarks or borders. It asks for at least 500 × 500 pixels for all products from January 31, 2027, and recommends about 1500 × 1500 or more. Shopify says 2048 × 2048 usually displays best for square product images, and that PNG is the best file type for most product images, followed by JPEG.
Label AI images. Google Merchant Center requires that images created with generative AI contain metadata saying so, gives the IPTC DigitalSourceType property as an example, and asks you not to remove it. In the IPTC vocabulary, trainedAlgorithmicMedia means ‘Created using Generative AI’ and compositeWithTrainedAlgorithmicMedia means ‘Edited using Generative AI’, for example with inpainting.
Finally, compare the image with the real product. Colors, size, materials and included parts should match what the customer will receive, so a generated background is fine but a generated product detail is not.
PhoneDiffusion availability and article scope
PhoneDiffusion runs curated Core ML model packs (SD 1.5, SD 2.1, and SDXL) on supported iPhone and iPad devices with iOS 17 or later; local generation is private and works offline once a pack is downloaded. Text-to-image works for concept and mood shots, in square images (512 × 512 for SD 1.5 and SD 2.1 packs, 768 × 768 for SDXL packs), with the optional negative prompt for things to keep out.
For a real product, choose your photo in Inpainting, paint over the product and tap Invert, so only the background is regenerated; the app shows that only the red area will change, and Replacement strength sets how far it moves from the original. Background Removal cuts out the subject automatically into a PNG with a transparent background, with a Cleanup setting from Soft to Strong for the edges. 4× upscaling raises a finished square before you upload it. The app has no ControlNet input, and placing a cutout on a new background is a step for a separate editor.
Images saved or shared from the app are copies signed with Content Credentials (C2PA), titled ‘AI-generated image’ and marked with the trainedAlgorithmicMedia source type. Whether a marketplace reads that record is up to the marketplace, so check its rules. The optional cloud models on paid monthly or yearly plans, such as FLUX and Seedream, are text-to-image only: they send the text prompt online and cannot use your product photo. The ControlNet, apply_overlay and inpaint-checkpoint details above describe desktop tools.
Field notes
- Choose the shot first: packshot, hero, lifestyle, flat lay or detail.
- Prompt angle, framing, surface, background and one main light in plain photo terms.
- One product and two or three props; the models struggle with exact arrangements.
- Stable Diffusion cannot reproduce your real product, label or logo; generate the background around a photo instead.
- Check store rules: clean view, no added text, product at 75% to 90% of the frame (Google), and AI metadata kept.
Try it on your phone
- Pick the shot type and where the image will be used.
- Write the prompt in order: product, angle, framing, surface, background, light.
- Generate a few seeds, keep the best one fixed, and change one cue at a time.
- For a real product, photograph it and inpaint only the background, or cut it out and place it on a generated scene.
- Add any text in an editor, upscale, compare with the real product, and upload the exported file.
Sources and verification
Primary and maintainer documentation checked for this guide. Product-specific details were reviewed against the current PhoneDiffusion scope.
- Stable Diffusion v1.5 and SDXL base 1.0 model cards: limitationsSDXL base 1.0 lists ‘The model does not achieve perfect photorealism’, ‘The model cannot render legible text’ and struggles with compositionality, with the example ‘A red cube on top of a blue sphere’. The SD 1.5 card lists the same text and compositionality limits.
- Stable Diffusion v1.5 model card: limitationsModel card listing ‘The model cannot render legible text’, a lossy autoencoder, and mainly English training captions; uses a fixed CLIP ViT-L/14 text encoder and was trained at 512×512.
- Hugging Face Diffusers: Inpainting (masks, checkpoints and apply_overlay)Maintainer guide: white mask pixels are filled in from the prompt and black pixels are kept; inpaint-specific checkpoints blend more naturally but are more likely to change the unmasked area; VaeImageProcessor.apply_overlay() forces the unmasked area to stay the same.
- Hugging Face Diffusers: ControlNetMaintainer guide: a ControlNet steers a model with a structural control image such as canny edges or depth while keeping the text prompt; controlnet_conditioning_scale sets how strongly.
- Google Merchant Center Help: Image requirementsAt least 500 × 500 pixels for all products from January 31, 2027, about 1500 × 1500 or above recommended, no image over 64 megapixels or 16MB; an unobstructed, clean view of the item without added text, promotional graphics, watermarks or borders; frame the product at no less than 75% and no more than 90% of the image; images created with generative AI must contain metadata saying so, such as IPTC DigitalSourceType, and that metadata must not be removed.
- Shopify Help Center: Product media types and specificationsFor square product images, 2048 × 2048 px usually displays best; up to 5000 × 5000 px or 25 megapixels and under 20 MB; PNG is the best file type for most product images, followed by JPEG; use a consistent aspect ratio for main images shown side by side.
- IPTC NewsCodes: Digital Source Type vocabularytrainedAlgorithmicMedia, ‘Created using Generative AI’, is media created algorithmically by an AI model trained on captured content; compositeWithTrainedAlgorithmicMedia, ‘Edited using Generative AI’, covers augmentation such as inpainting or outpainting.