Brand-safe generation for commercial work

The legal status of training generative models on copyrighted images is unsettled and probably will be for years. There are active lawsuits against Stability, Midjourney, Runway, and others by groups of artists and stock photo companies. The pragmatic stance for commercial users is: assume there is some legal risk, even if the cases get resolved in the model labs' favor, and minimize your exposure where you can.

The lowest-risk path is to use Adobe Firefly, which is the only frontier-quality model trained exclusively on licensed dataLicensed dataTraining data used with the permission of and usually payment to its rights holders, the basis for models marketed as commercially clean. (Adobe Stock plus public domain content). Adobe offers full IP indemnificationIndemnificationA vendor's promise to cover a customer's legal costs if the product triggers a lawsuit, offered by some licensed-data model providers as protection against copyright claims on generated output. to commercial customers, if a third party sues you for copyright infringement based on a Firefly-generated image, Adobe defends the case. This indemnification is valuable enough that many enterprise customers use Firefly even when its raw output quality is lower than FLUX or Imagen, because the legal certainty matters more than the marginal quality.

The next-lowest-risk path is to use models with documented training data sources that you can audit (Stable Diffusion 3.5, with the LAIONLAIONA German nonprofit that assembled the LAION-5B dataset of 5 billion image-text pairs used to train Stable Diffusion.-5B-derived training sets), combined with prompts that do not invoke specific copyrighted styles or characters. Avoid prompts like 'in the style of [named living artist]', those create the strongest legal exposure even if the underlying training is legal.

The highest-risk path is to use a closed model with undisclosed training data on prompts that invoke specific protected content. Midjourney, in particular, has been the subject of multiple lawsuits over its alleged training on copyrighted artist work, and using it commercially for stylistically-specific work involves accepting that risk.

Generating images that look like specific real people without their permission is a legal and ethical minefield. Public figures have different rights than private individuals, the rules vary by jurisdiction, and the best practice for commercial work is: do not do it without explicit permission, even for public figures, even in jurisdictions where it might be technically legal. The downside risk (lawsuit, brand damage, regulatory action) is much larger than the upside.

If your commercial workflow needs to depict identifiable people (actors, models, executives), the right path is to license their likeness through normal industry channels and then either use them as the trained subject of a custom model (with their permission and an appropriate contract) or composite them into AI-generated environments using traditional photo techniques. The line you do not want to cross is: training a model on a person's likeness without their permission and using the outputs commercially.

For brand work, the generated images need to comply with the brand's visual guidelines, color palette, typography, visual tone, prohibited content categories. This is hard to enforce automatically. The two practical approaches are: train a brand LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. that biases the model toward brand-compliant aesthetics (effective for color palette and visual tone, less effective for typography and prohibited content), and run a manual review pass on every generation before it ships (slow but reliable). Most enterprise brand programs use both, automated bias plus human review.

The growing category of 'brand AI' products (Typeface, Jasper Studio, Adobe GenStudio) is built around this problem. They wrap base models in brand-aware workflow tooling, the brand uploads its guidelines, the tool ingests them, and the generation interface enforces compliance through a combination of prompt engineeringPrompt engineeringThe craft of writing and structuring prompts to reliably get the result you want., LoRA application, and review queues. This is a category that did not exist three years ago and is now a meaningful business segment.

As regulatory pressure increases, commercial users will increasingly need to attach provenanceProvenanceA verifiable record of how a piece of media was made and by whom, increasingly required so that synthetic content can be traced. metadata to AI-generated content. The C2PAC2PA (Content Credentials)An open provenance standard, backed by Adobe, OpenAI, and Microsoft, that attaches cryptographically signed metadata proving a file's origin, though the metadata strips easily when a file is re-encoded. standard (Coalition for Content Provenance and Authenticity) is the dominant technical approach. Most major model providers and platforms have committed to it. For commercial work, the practical guidance is: enable C2PA watermarkingWatermarkingHiding a detectable signal inside generated media so its origin can later be identified. For audio and AI media it remains an arms race rather than a guarantee. wherever your tool supports it, and assume that your customers (especially in regulated industries like journalism, advertising, and healthcare) will increasingly require provenance metadata as a condition of use.

KEY TAKEAWAYS

1. Quality in generative imagery is multidimensional. The eight dimensions are photorealism, stylistic distinctiveness, prompt adherencePrompt adherenceHow faithfully a model follows the specifics of your prompt, meaning the objects, counts, colors, spatial relationships, and text you asked for. High adherence means you get what you described, not just something in the right neighborhood., compositional control, text rendering, coherence, speed, and cost. No model is best on all of them.

2. The right model for a use case is the one with the best profile on the dimensions that matter most for that use case. Asking 'which model is best' without specifying use case is the wrong question.

3. Different use cases (product photography, editorial illustration, brand campaigns, concept art, social content, architectural viz) have very different quality requirements. Match models to use cases deliberately.

4. Brand-safe commercial generation requires attention to IP exposure, real person depictions, brand guideline compliance, and provenance/watermarking. Adobe Firefly has the strongest IP position for risk-sensitive enterprises.

5. The category of 'brand AI' products that wrap generation in brand-aware workflow is the fastest-growing B2B segment in generative imagery. Custom-trained brand models plus enforced workflows are how enterprises will use this technology at scale.