A visual experiment examines how AI can explore light atmosphere materials and framing without replacing architectural control or art direction.
An AI-generated variation may suggest exactly the light a scene needs while inventing a handrail. That is the useful tension: it speeds up visual questions but cannot answer for the architecture. Art direction begins by separating a valuable intuition from a change the project does not permit.
This experiment follows one reference through rain, changing light and alternative cameras. We are interested in the decisions, not a gallery of effects. Each test asks what the image gains, what it loses and what would need rebuilding in 3D before it could faithfully represent the design.
From the finished image to the image as a hypothesis
Generative retouching is often explained through before and after comparisons. Textures gain detail, vegetation changes, or a digital appearance becomes softer. Yet more detail does not guarantee a more interesting image. The decision organising the scene may remain exactly where it was.
Bertrand Benoit’s published comparisons in AI finishing provide examples of work on image finishing. [3] At article viewing scale, their changes do not support an unequivocal artistic improvement across every example. This observation defines our reading of the images; it is not a general assessment of the software used.
A more productive question concerns using a render to imagine images we have not yet decided to produce. Side light may suggest another relationship with materiality. A close view of an entrance may turn a façade into an experience of arrival. A dark night may give meaning to a single window.
A 3D model still provides a scene in which cameras and lighting can be defined and an image recalculated. [5] A generative tool receiving a flat file works with what that file shows and what it infers. It can offer a convincing idea without demonstrating that the idea corresponds to the building.
One reference and two ways of imagining after the rain
The reference is an image of Bell Bell Apartments published in JVA Renders’ portfolio. It offers a recognisable starting point: a brick façade, two balconies, an entrance between planted walls, and a palm in the foreground. We use our own case to document transformations from the same input without turning the selection into a commercial comparison between studios.
The first two tests distinguish working approaches. Text generation describes scene elements but leaves their relationships open. It can invent plausible architecture that is no longer the same building. Image editing receives visible relationships as a reference, although it can reinterpret them too.
The second test better retains the overall composition and changes the situation: overcast sky, wet ground and interior light. Its contribution is to make a different moment imaginable. Proportions and details require review even when the whole looks familiar. This distinction underpins the following series: a proposal can guide an image without being ready to replace a production render.
One framing in eight lighting states
Each variant starts directly from the original. The instructions requested the same camera, architecture and dry paving, changing only the lighting situation. The eight names describe artistic states, not verified solar times at the site.
What changes when the light changes
First and last light share a warm vocabulary. That does not mean both should remain in the final selection. Producing eight alternatives makes sense for exploration; publishing eight similar images can weaken the story. If two serve the same editorial purpose, the stronger expression of that intention should be selected.
Morning presents the building relatively openly. Midday gives more weight to balcony thickness and the shadow beneath the roof. This reading could explain protection or depth, although it would need lighting consistent with the model before becoming a final image.
Backlight changes the hierarchy more clearly. Vegetation gains presence and the building loses information. The same loss can support a landscape image and weaken one intended to explain finishes. Assessing the variant requires deciding its purpose first.
Blue hour shifts attention towards a lit room. Night takes the idea further, accepting that much of the architecture remains barely visible. A night scene does not need every surface illuminated: it can be organised around what stays lit and what disappears.
The series retains broad features but does not establish exact preservation. Some daylight variants still show a small light beneath the balcony despite instructions to switch it off. This is a specific departure from the brief. We also lack orientation, date and solar trajectory to associate these lighting states with real hours. The experiment proposes atmospheres that can subsequently be checked against the 3D scene.
Proposing a camera means proposing an observer
Changing framing alters the viewer’s relationship with the project. We explored this through three independent instructions: approach the entrance, look from a low position, and seek a frontal view. The results distinguish a useful intention from a successfully executed transformation.
The threshold reduces the presence of sky and palm. The entrance stops being one component within a façade and becomes the main subject. It can guide an arrival camera: an image about entering, waiting or crossing a boundary. The proposal is interesting because attention changes, even though its surfaces and relationships remain unverified.
The low viewpoint expands the sky and emphasises the roof’s underside. It suggests a bodily relationship with the architecture. It also exposes an artefact: the mark in the original becomes a small grey symbol. Apparent coherence across the scene does not guarantee continuity in every element.
The third test does not provide a clear frontal view. It keeps an oblique view close to the original. This result identifies an instruction limit: requesting a different camera does not mean obtaining it. A research series should retain responses that fail to fulfil the hypothesis too.
Cropping an existing image selects information already visible. Moving a camera may reveal surfaces and relationships absent from the reference. View synthesis research, such as Stable Virtual Camera, specifically considers input images and target cameras. [4] We did not use that model in these tests. Its approach helps distinguish a proposed view from reliable project documentation: hidden areas still require information or verification.
From retouching to concept art
CommonPoint’s published process includes visual research, asset generation, a 3D scene and postproduction. It also includes lighting and camera tests. [2] This case offers a view of AI use that starts before finishing.
The board brings together people, animals, fabrics and urban scenes. Each fragment can contribute something different: scale, density, a gesture or a relationship between bodies and architecture. Visual research starts constructing a world before deciding which image will show it.
The final image organises dark nearby planes, figures, fabrics and stone masses against a luminous landscape. Its interest lies in how it directs the eye and makes a place imaginable. Elements do more than add detail: they establish routes, distances and forms of inhabitation.
The board and scene do not demonstrate an automatic conversion between them. They are published references from a broader process. We include them for their ability to show a search for visual language, without attributing our series’ experimental control to them or comparing them as if they had received the same input.
Using AI as an artistic editor
A tool can produce many alternatives; art direction must give them a purpose. An approaching view can open a sequence. Midday can explain shade and thickness. A lit room can close the story with a scene of use. Selection becomes meaningful when images perform different jobs.
| Test | Editorial question | Decision it informs |
|---|---|---|
| Side light | Which surfaces and edges should stand out | Light direction in the 3D scene |
| Midday | How mass and protection are perceived | Depth and extent of shadows |
| Backlight | How prominent the landscape should be | Balance between façade and silhouette |
| Blue hour and night | Where interior life appears | Localised light and the sequence ending |
| Approach | How arrival is experienced | Entrance camera |
| Low or lateral viewpoint | How the body relates to the building | Camera position to validate in 3D |
Asking for a “more artistic” image leaves too many decisions undefined. It is more precise to ask for shadow that reveals a balcony’s thickness, a landscape that leads the eye, or a single room that suggests occupation. The brief describes an intention that can then be discussed through concrete results.
Exploration allows extreme proposals: a night hiding most of the façade, backlight giving the tree prominence, or a camera so close that the building’s overall image disappears. Their usefulness depends on what they reveal. Some will become production references; others will show which directions to discard.
From exploration to an image we can deliver
A variant can contain a good lighting idea and an incorrect railing. Professional judgement separates the two and transfers the intention into an environment where geometry can be controlled. Taking advantage of a proposal does not require accepting every change in its output.
Camera proposals particularly need checking. A generated approach may suggest a position that does not exist in the model, or reveal an invented surface. Reconstructing it in 3D can change the composition. This check belongs within image direction.
Two questions should be reviewed separately. Architectural continuity requires checking openings, junctions, proportions and visible elements. Narrative intention requires assessing hierarchy, atmosphere and function within the series. A technically correct image can say little; an evocative proposal may need substantial reconstruction before representing the project.
The experiment does not measure time savings, audience preference or objective quality improvement. It documents alternatives that support decisions and discussion. AI expands the testing field; editorial value emerges in choosing which idea deserves to become an image and with what degree of control.
The exit rule: keep the visual intention that helps the project, then recheck geometry, materials and camera in a controlled environment. A convincing image is not yet a responsible one.
Method and sources
The article contains sixteen images: one original CGI, thirteen in-house generative tests and two external CommonPoint references. The thirteen tests comprise one text generation, one rainy edit, eight lighting states and three camera hypotheses. The eleven new variants start independently from the original without chaining outputs.
One output was produced per condition and all were retained, including the unresolved frontal view. No subsequent corrections were applied. We used ChatGPT’s image generation and editing tool; its response did not provide an exact model version or a seed. The available technical archive documents the test conditions and results, but does not preserve the prompts verbatim or a complete provenance log.
The reading is qualitative and informed by image provenance. There were no participants or blind test. Dimensions, reconstructed cameras and solar lighting were not verified. Generative images are labelled experimental and external references retain their credits. Research and revision closed on 2 October 2026.
- JVA Renders. Bell Bell Apartments and reference portfolio, page 8. https://jvarenders.com/projects/bell-bell-apartments/ · https://jvarenders.com/wp-content/uploads/2022/01/jva_portfolio_english.pdf
- CommonPoint. Published AI visualisation process, asset research and Silk Road case. https://thecommonpoint.com/ai
- Bertrand Benoit. AI finishing, 25 March 2024. https://bertrand-benoit.com/blog/ai-finishing/
- Jinghao Zhou and colleagues. Stable Virtual Camera Generative View Synthesis with Diffusion Models, 2025. https://arxiv.org/abs/2503.14489 · https://huggingface.co/stabilityai/stable-virtual-camera
- Blender. Official manual, introduction to rendering. https://docs.blender.org/manual/en/latest/render/introduction.html
RENDERS


