AI and generated media

# AI product image generation with SKU-level safeguards

Generate product imagery without inventing features, colors, packaging, or claims that customers will mistake for the real item.

Published August 13, 2026

## Key takeaways

* Treat product identity, labels, materials, colors, and included accessories as facts that generation must not improvise.
* Use deterministic compositing when exact approved product pixels matter more than creative scene variation.
* Measure rejection and customer-correction outcomes alongside creative throughput and rendering cost.

AI can create inexpensive product-scene variants, but ecommerce imagery is also a factual interface. A beautiful output that changes a control, texture, scale, or accessory can create returns and regulatory risk.

## In this guide

1. [Separate creative scene freedom from SKU truth](#ai-product-image-generation-section-1)
2. [Ground edits in approved product references](#ai-product-image-generation-section-2)
3. [Choose compositing when exact pixels must survive](#ai-product-image-generation-section-3)
4. [Use masks to define the intended edit region](#ai-product-image-generation-section-4)
5. [Review identity, materials, color, and accessories](#ai-product-image-generation-section-5)
6. [Reject candidates that change a material product fact](#ai-product-image-generation-section-6)
7. [Invalidate derivatives when the source SKU changes](#ai-product-image-generation-section-7)

## Separate creative scene freedom from SKU truth

List the product properties that must remain exact: shape, proportions, packaging, labels, color, material, included accessories, and legally relevant claims. Then name the areas where invention is allowed, such as a background surface, shadow, or decorative context. Review criteria should mirror that split.

A plausible image can still be commercially false. Tie every request to an approved SKU version and publication channel, and define whether the output is a concept, internal draft, or customer-facing asset. Keep a deterministic photography or compositing fallback for protected products.

## Ground edits in approved product references

Use controlled reference photography and authoritative product data, not marketplace screenshots or filenames. Preserve source resolution, color treatment, and checksum. If multiple views or masks are required, label their semantic roles and record which approved product version they depict.

Do not send confidential launch products or unnecessary customer data to a provider without an approved data path. Review provider region, retention, model terms, and access. Prompts should describe permitted scene changes without inviting the model to redesign protected product attributes.

## Choose compositing when exact pixels must survive

Use exact compositing when the product, text, regulatory mark, or packaging cannot change. Background removal and deterministic layout can often place approved source pixels into a new scene more reliably than generation. Reserve inpainting for bounded regions where review can detect unwanted changes.

Compare generation with the deterministic baseline on correction time and publishability, not visual novelty. A generated shadow or reflection may still contradict material or scale. Product owners should approve the protected-fact checklist before a creative reviewer judges aesthetics.

## Use masks to define the intended edit region

The example uses two uploads with semantic `image` and `mask` roles for documented inpainting. Keep the mask aligned with the source and verify its dimensions and intended polarity. The prompt requests a background change, but a reviewer must still confirm that the product pixels remained acceptable.

Store source, mask, prompt version, model, candidate, SKU version, and review state. Create delivery derivatives only from the approved candidate. A failed export should retry independently, and a late result for an obsolete SKU image must not replace a newer approved asset.

Propose a background edit from two uploads: product and mask

```
{
  "steps": {
    ":original": { "robot": "/upload/handle" },
    "scene_candidate": {
      "use": [
        { "name": ":original", "as": "image" },
        { "name": ":original", "as": "mask" }
      ],
      "robot": "/image/generate",
      "model": "stability-ai/stable-diffusion-inpainting",
      "prompt": "Replace only the masked background with a neutral studio set. Keep the product, packaging, labels, colors, and geometry unchanged.",
      "format": "png"
    }
  }
}
```

## Review identity, materials, color, and accessories

Compare candidate and source for silhouette, dimensions, labels, spelling, logos, color, finish, texture, openings, controls, included items, and reflections. Review every target crop and compressed output. Use objective color or geometry checks where possible, but retain qualified visual review.

Record rejection reasons per product category because failure patterns differ for transparent items, jewelry, textiles, reflective packaging, and products with small text. Acceptance rate without SKU-fact errors and correction time can hide candidates that are attractive but unusable.

## Reject candidates that change a material product fact

Reject rather than retouch candidates that add features, alter quantity, change packaging, obscure required warnings, or imply an unsupported environment or use. Keep the approved source and choose compositing, reshooting, or a different creative concept when model freedom conflicts with product truth.

Verify rights, talent consent, trademark treatment, disclosure, and claim substantiation for the intended territory and channel. Protect reviewers and credentials, sanitize errors, and provide a correction and takedown path for published assets.

## Invalidate derivatives when the source SKU changes

Version the SKU reference, masks, prompt, model settings, review rubric, finishing Steps, and approved outputs. When packaging or product facts change, mark dependent generated media stale and block republication until it is reviewed against the new source.

Monitor rejection categories, correction time, stale-result rejection, provider errors, cost per accepted SKU, and later replacement. Keep prior approved assets available during rollout, and make callbacks and publication idempotent so completion order cannot select product truth.

## Technical details worth knowing

* Task boundary: AI product-image generation creates or edits merchandising visuals from a brief and approved product references. Product-image generation synthesizes or modifies a scene; product photography records an item, and deterministic compositing preserves approved pixels.
* Input contract: Ground generation in approved product references and structured facts, lock protected labels and geometry where possible, and separate lifestyle freedom from SKU truth. The sample expects two uploaded files—a product image and mask—tagged through semantic field names or the as directives. Input preparation must be evaluated with the model because preprocessing can remove evidence as well as noise.
* Output contract: Produce candidates linked to SKU, approved references, prompt and model provenance, generated-area disclosure, review state, and final channel renditions. A valid response does not prove that a predicted label, region, or generated pixel is correct.
* Method choice: Prefer deterministic background replacement for exact SKU depiction and reserve generation for surroundings or concepts that can be reviewed against the approved product. Model names alone do not describe the training data, thresholds, latency, licensing, or failure behavior of a deployed system.
* Evaluation: Review SKU identity, geometry, color, materials, labels, included accessories, shadows, brand rules, channel crops, rejection rate, and conversion complaints. Aggregate scores should be segmented by content type so common easy examples do not hide failures on important edge cases.
* Failure and safety: If the output changes a material product fact, reject it and use approved photography or a deterministic composite rather than “fixing” it with more automation. Require product and legal review for regulated claims, people, trademarks, packaging, or material appearance, and never imply included features absent from the SKU.
* Operations: Version product facts, references, prompts, and approvals; monitor rejection reasons and customer reports; invalidate derivatives when the source SKU changes.

## A practical approach

1. 1\
   Write the decision, output schema, and rejection criteria for AI product-image generation.
2. 2\
   Build a representative AI product-image generation evaluation set and preserve each source, preprocessing choice, and provenance record.
3. 3\
   Benchmark the complete workflow on representative evidence and compare the result with predefined task-specific acceptance criteria.
4. 4\
   Release AI product-image generation behind explicit review and fallback paths, then monitor the operating signals that determine whether it remains useful.

A four-stage media workflow

## When Transloadit is useful

Use /image/generate for supported generation tasks, /image/bgremove for approved source cutouts, and /image/resize for channel-specific derivatives after product review.

## Architecture boundary

Transloadit can generate images, remove backgrounds, and perform deterministic resize, composition, and export steps, but it does not know whether a generated product depiction is legally or commercially accurate.

## Frequently asked questions

### When is compositing safer than product-image generation?

Use compositing when exact product pixels, packaging, labels, colors, or regulatory details must survive. Generation is suitable only for bounded, reviewable invention.

### Does an inpainting mask guarantee that unmasked pixels stay unchanged?

Treat the mask and prompt as requested constraints, not proof. Compare protected product details with the approved source and reject candidates that alter material facts.

### What should happen after a product or package revision?

Invalidate dependent generated assets by SKU version, retain the prior history, and review or regenerate against the new authoritative reference before publication.

### What should product review measure besides visual appeal?

Check identity, geometry, labels, spelling, material, color, accessories, claims, target crops, correction time, and rejection reasons for each product category.

## Build the workflow

Move from the concept to a tested Assembly with Robot documentation and working demos.

### Relevant Robots

* [/image/generate](/docs/robots/image-generate.md)
* [/image/bgremove](/docs/robots/image-bgremove.md)
* [/image/resize](/docs/robots/image-resize.md)
* [Read the API documentation](/docs.md)
* [Explore working demos](/demos.md)
* [Create a free workspace](/c/signup/)

AI and generated media

## Continue with related guides

* [How to give AI agents safe media-processing capabilities](/guides/transloadit-for-ai-agents.md)\
  Give AI agents constrained media capabilities through stable Templates, structured inputs, and observable Assembly results.
* [Two ways to describe images with AI](/guides/describe-images-with-ai.md)\
  Compare direct vision-model descriptions with a preprocessing pipeline that creates cleaner, bounded analysis inputs.
* [Creating 3D models from images: a practical pipeline](/guides/create-3d-models-from-images.md)\
  Plan a photo-to-3D capture and processing workflow without confusing image preparation with 3D reconstruction.
* [AI image analysis: tasks, architecture, and safeguards](/guides/ai-image-analysis.md)\
  Understand classification, detection, OCR, captioning, embeddings, and moderation as different image-analysis tasks.
* [AI image processors inside production media pipelines](/guides/ai-image-processing-pipelines.md)\
  Combine generative or analytical image services with deterministic validation and transformations.
* [How to extract text from documents and images at scale](/guides/extract-text-from-documents-and-images.md)\
  Recognise text across PDFs, scans, and photographs, and keep the result attached to the file it came from.