Visual automation by generative models: what the trade-off requires clarifying

Note revised on 25 May 2026. Article originally published in January 2026 — full rewrite.

Generative models capable of producing images from a textual description have left the laboratory to become a category of tools usable in practice. For a Swiss SME regularly consuming visuals — social networks, product pages, illustrations for presentations, linguistic variants — this accessibility modifies the economics of visual production. It does not, however, exempt one from an explicit trade-off on the parameters that distinguish durable use from costly drift.

This note sets out what that trade-off requires clarifying: brand coherence, sovereignty over the data processed, the applicable legal framework, and the place that assisted production leaves to internal visual expertise.

One category of tools, several operational profiles

The market for visual-generation tools does not reduce to a single product. Several operational profiles coexist, with different trade-offs between accessibility, control, aesthetic quality and data sovereignty.

Tools oriented towards aesthetic rendering stand out by an immediate quality of produced images, often comparable to professional photography or illustration. Their ease of use is limited by their interface, and brand coherence requires explicit framing through detailed prompts. They are suited to uses where aesthetic quality prevails over granular control.

Tools integrated into conversational assistants offer maximum accessibility to non-technical staff. French understanding is generally good. Rendering quality is correct, without reaching the level of specialised tools. This category suits rapid prototyping, blog-article visuals, internal educational illustrations.

Tools integrated into a professional creative suite present a distinctive characteristic: their model is generally trained on content under verified licence, which reduces exposure to legal risks linked to copyright. Their integration into existing workflows — retouching, layout, export — constitutes a substantial operational advantage for design teams already working in that ecosystem.

Open models executed on controlled infrastructure offer maximum control — confidentiality of processed data, deep customisation through supplementary training, independence from a single provider. This control has a cost: it demands substantial technical competence and adapted hardware infrastructure.

The choice between these profiles depends less on the tool itself than on the constraints specific to the company: volume, coherence requirements, sensitivity of the data processed, internal competence available. No profile objectively dominates the others.

The central challenge: brand coherence

The main gap between experimentation and durable use of generative tools in visual production is not played out on the quality of the individual rendering. It is played out on visual coherence over time.

Each generation produces a slightly different result. Without an explicit framework, a company's visual communication loses coherence: colours drift, compositions vary without logic, the visual style is no longer identifiable. This drift is not seen on an isolated visual; it is revealed across the timeline of a social-media feed or across the whole material of a campaign.

Three practices distinguish organisations that maintain durable coherence.

The documentation of an explicit visual framework first — often called a brand prompt in the jargon of practice. This documentation describes the expected artistic style, the dominant chromatic palette, recurring elements, subjects to avoid, the general visual tone. It constitutes the reference brief for any generation, shared by all staff producing visuals.

Supplementary training of models next, for organisations with the technical capacity. On open models, it is possible to train an adaptation layer on the company's visual identity, which makes generations naturally coherent with the visual identity, without requiring heavy prompt formulation on each visual.

Systematic human validation completes the list. No automated generation goes without validation by an internal visual referent before publication. This validation absorbs anomalies — six-fingered hand, illegible text, spatial inconsistency — that generative models still produce regularly, and it preserves the overall coherence that documentation alone does not guarantee.

The applicable Swiss legal framework

The commercial use of images produced by generative models fits into a legal framework that deserves to be clarified, rather than ignored on the grounds that the subject is new.

Three elements of Swiss law apply directly.

The Copyright Act^[1] first. A work must result from an intellectual creation to benefit from Swiss copyright protection. A visual entirely produced by a generative model, without significant human creative intervention, may not benefit from this protection. Qualification depends on the assessment of human creative input in the production process, and this assessment is made on a case-by-case basis. This zone of uncertainty has a practical consequence: a company that produces its visuals exclusively by automated generation, without documented creative intervention, may find itself unable to oppose their re-use by third parties under copyright. This point deserves to be posed in the strategic arbitration of the practice.

The Federal Act against Unfair Competition^[2] next. The use of misleading generative visuals — false client testimonials, staged scenes presented as real when they are not, false professional validations — can constitute a sanctionable unfair practice. This qualification does not bear on the technology used, but on the misleading nature of the communication produced. It applies to generative visuals as to others.

Image rights complete the list. Producing a visual that reproduces the face of an identifiable person without their consent remains unlawful, whether the image is photographic or produced by a generative model. This rule probably constitutes the most immediate risk zone, because models can produce involuntary resemblances to real people that the user does not recognise as such.

For companies also active on the European market, the European Regulation on Artificial Intelligence^[3] is progressively introducing transparency obligations on the origin of content produced by models. These obligations deserve to be anticipated rather than discovered at the moment they become binding.

Sovereignty over processed data

For online visual-automation tools, the images or descriptions transmitted to the service are, by construction, processed at the provider's end. This characteristic calls for qualification for sensitive content.

Three risk zones deserve to be identified at the initial framing of the practice.

Personal data included in the briefs first. If the description supplied to the model includes information about identifiable people — clients, staff, prospects — this transmission constitutes a processing of personal data that falls within the scope of the Federal Act on Data Protection. The use of online tools for these briefs requires, at the very least, explicit analysis of the processing conditions at the provider.

Intellectual property of the visual elements transmitted next. When a tool takes an existing photograph as a reference to generate a variation, the original photograph may be protected by a third party's copyright. Its transmission to the provider is not legally neutral.

Strategic confidentiality completes the list. A visual describing a not-yet-launched product, a campaign in preparation, a visual identity in the midst of redesign, exposes the organisation if the provider's confidentiality is not assured by explicit contractual commitment.

For content where these risk zones are substantial, recourse to open models executed on controlled infrastructure constitutes an alternative that structurally resolves the question, at the cost of more demanding internal technical competence.

The place that practice leaves to internal visual expertise

Visual automation by generative models does not remove the visual-expertise function in an organisation. It shifts its content.

Repetitive production tasks — format variants, colour variants, linguistic adaptations of existing visuals, illustration visuals for standardised content — are absorbed by the tools, with acceptable quality provided framing. The time freed on these tasks deserves to be reinvested in higher-value activities.

Artistic direction takes on weight in the residual share of the work. Defining the visual coherence of a brand, arbitrating between stylistic directions, maintaining quality over time, now demands more structured attention — precisely because, without this attention, automated production would tend towards a poorly differentiated homogenisation.

Visual strategy also rises in value. Choosing where assisted production is appropriate and where human expertise remains structuring, framing internal practices to preserve coherence, integrating legal constraints into production flows, are subjects that are not resolved with an additional tool.

Quality supervision completes the list. The systematic control of visuals before publication — technical anomalies, stylistic drifts, legal risks — now constitutes an activity in its own right that deserves to be structured rather than dispersed.

For Swiss SMEs without internal visual expertise, this practice can be organised with an external referent — independent graphic designer or specialised agency — who assures artistic direction and quality supervision, while internal teams produce routine visuals with accessible tools.

The discipline that distinguishes durable use

Visual automation by generative models is neither the magical solution sold by some presentations, nor the systematic degradation of visual quality feared by others. It constitutes a real shift in required skills and trade-offs to pose, which demands to be treated seriously.

Three orientations distinguish the organisations that derive durable value from it.

Starting with a precise use case first, rather than attempting to automate the whole of visual production. Social-media visuals or blog-article illustrations generally constitute appropriate entry points.

Investment in the practice next: training at least one team member in prompt framing, documentation of the visual framework, setting up a library of proven prompts, sharing the practices that work.

Combination with manual retouching completes the list. The use that produces durable value consists of generating the first draft by model, then refining through the usual retouching tools to reach the expected professional standard. This combination recognises the tools' value without overestimating it.

For sectors where visual authenticity prevails — luxury, gastronomy, Swiss tourism — exclusive use of automated images presents a positioning risk that clients perceive, sometimes unconsciously. The combination of real photography and assisted production preserves the credibility of the visual discourse.

The trade-off between automation and human expertise is not posed in binary terms. It is posed in terms of proportion, discipline and framework. This qualification distinguishes the organisations consolidating their visual communication from those carried away by the apparent ease of the tools.

Sources

[1] Federal Act on Copyright and Related Rights (CopA), of 9 October 1992. www.fedlex.admin.ch/eli/cc/1993/1798_1798_1798/en [↩]

[2] Federal Act against Unfair Competition (UCA), of 19 December 1986. www.fedlex.admin.ch/eli/cc/1988/223_223_223/en [↩]

[3] Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence. eur-lex.europa.eu/eli/reg/2024/1689/oj/eng [↩]

Jérôme Deshaie is CEO of MCVA Consulting SA, a Swiss firm specialising in strategic consulting on artificial intelligence, based in Valais.