Understand the technology
Start with a source. Define the transformation.
For May and November 2027. This introductory lesson is shared SL and HL content.
01 | BEGIN WITH AN IMAGE
A designer has a rough drawing of a building and wants to explore how it could look as a realistic scene. Instead of describing the entire building from scratch, the designer supplies the drawing as an input.
Image-to-image translation generates a transformed image guided by an input image. The input can communicate structure, content or appearance that the system should use when producing the result.
The 2027 booklet gives examples including sketches transformed into realistic renderings and black-and-white images converted into colour. Visionary Studios could investigate these approaches when existing visual material is a useful starting point.
02 | WHAT DOES TRANSLATION MEAN HERE?
In language translation, a message is expressed in another language. In image-to-image translation, content is transformed between visual forms or domains, such as line drawings and realistic-looking images.
A domain is a type or distribution of data, such as sketches or colour photographs. The source domain provides the input; the target domain describes the intended kind of output.
The analogy has limits. A transformed image does not necessarily preserve every detail or convey an objectively correct interpretation. Some information may be changed or invented, so the result must be checked against the task.
03 | INPUT, GUIDANCE AND OUTPUT
A compatible system processes the input image using learned relationships. Some systems also accept a text prompt, a region mask or settings that influence the transformation. The supported inputs depend on the model and application.
A sketch, photograph or other supported image.
The requested target appearance and supported controls.
A transformed image that needs review.
Supplying an image during generation is not automatically training the model on that image. Ordinary inference can use the image as a condition while keeping learned parameters fixed.
04 | TEXT-TO-IMAGE AND IMAGE-TO-IMAGE
Text-to-image generation uses textual guidance to produce an image. Image-to-image translation includes a source image that guides the transformation. A system may combine both, using an image for layout and text for the intended style.
The important distinction is the role of the supplied visual material. A sketch gives spatial information that could be difficult to communicate precisely in words alone.
Do not assume that every application’s reference-image feature performs the same operation. Some use references primarily for style or identity, while others transform the content of the supplied image. Investigate what the chosen tool actually supports.
05 | SKETCH TO REALISTIC RENDERING
A sketch can indicate the shape of a building, its windows and its position in a scene. Translation can add plausible materials, shadows and background detail to produce a realistic-looking rendering.
For concept exploration, these additions can help a designer compare possible appearances quickly. However, the sketch may not specify the material of a wall or the depth of a doorway, so the model must make choices.
A photorealistic appearance is not evidence of engineering accuracy. If the building is intended to represent a real design, the studio must check dimensions and structural features separately rather than treating the generated image as a technical drawing.
06 | COLOURISATION ADDS PLAUSIBLE COLOUR
Colourisation adds colour to a greyscale image. The input contains brightness and structural information but usually does not contain enough information to identify the original colours uniquely.
A light grey jacket could originally have been one of several colours. The model can produce a plausible choice based on learned patterns and any supplied guidance, but that does not prove what colour the jacket really was.
For Visionary Studios, plausible colour may be useful for creative work. Exact product or brand colours may require references and controlled editing.
07 | APPEARANCE AND STRUCTURE ARE DIFFERENT
Appearance includes colour, texture, lighting and style. Structure includes shapes, boundaries and the arrangement of objects.
A transformation might aim to change appearance while retaining structure. For example, a street sketch could become a painted scene while keeping the main building positions. Another task might deliberately permit structural changes.
The studio should state which properties matter before generating. “Make it more realistic” is incomplete if the designer also needs the exact number and placement of windows to remain unchanged.
08 | STATE WHAT MAY CHANGE AND WHAT MUST STAY
| Task | May change | Should be checked for preservation |
|---|---|---|
| Building sketch to realistic concept | Materials, lighting and fine surface detail | Required silhouette, window arrangement and layout |
| Greyscale campaign image to colour | Colour treatment | Subject identity and important geometry |
| Photograph to illustration | Artistic treatment and texture | Required subject features and composition |
| Campaign scene variation | Agreed background or atmosphere | Essential product and character features |
These are proposed requirements, not guaranteed capabilities. Some transformations may need manual correction or a different workflow.
09 | ONE SOURCE CAN HAVE MANY VALID OUTPUTS
A line drawing of a house does not fully determine its colours, materials, weather or time of day. Multiple outputs can follow the same basic structure while making different appearance choices.
This ambiguity can support creative exploration. It can also create problems when the task requires factual fidelity. The designer must distinguish an open creative decision from a property that must match a reference.
Random inputs and generation settings can add further variation. A visually convincing output is only one possible interpretation of the source, not necessarily a recovered original.
10 | TRANSFORMATION CONTROLS VARY
Some diffusion-based image-to-image tools provide a control often described as strength or denoising strength. In many such workflows, increasing it permits greater departure from the source by changing how much noise is introduced or how the denoising process is started.
The exact scale, direction and behaviour depend on the implementation. A value in one tool is not automatically equivalent to the same value in another. Other model types may use different controls entirely.
Greater change can support exploration but may lose important features. Less change can retain more of the source while limiting the requested transformation. Neither setting guarantees exact preservation.
11 | A REGION MASK IS A DIFFERENT KIND OF GUIDANCE
A region mask identifies an area for a particular operation. For example, an editing workflow may use a mask to request a change to the background while retaining the main subject.
This differs from a semantic segmentation map, which assigns categories such as sky, water or person to regions. A simple edit mask may only indicate selected and unselected areas.
Mask conventions and the treatment of unselected areas depend on the tool. A designer should verify the exported result rather than assume every unselected pixel must remain identical. Masked editing is a related workflow, not a requirement of all image-to-image translation.
12 | TRANSLATION IS NOT AUTOMATICALLY A FILTER
A conventional image filter follows a defined transformation, such as adjusting brightness. A learned generative transformation can introduce new detail or reinterpret ambiguous content using patterns from training.
Both can alter an image, but their behaviour and risks differ. If exact product geometry must stay fixed, a predictable editing operation may be preferable to a model that invents details.
Some applications combine conventional editing and generative processing. Evaluate the actual operation rather than assuming that every image transformation uses generative AI or that generative editing is always the better choice.
13 | PAIRS AND UNPAIRED COLLECTIONS: A TRAINING PREVIEW
Some translation approaches learn from paired data : corresponding source and target examples, such as a sketch and a matching photograph. The pairing provides a direct relationship for learning.
Other approaches use unpaired data : collections from the two domains without a matching target for every source. Additional learning constraints are needed to obtain useful relationships; a collection of unrelated images does not itself guarantee faithful translation.
These are supporting ideas for understanding models, not a requirement to study a particular architecture here. Visionary Studios using an existing system would not necessarily need to assemble a new training dataset.
14 | REVIEW FIDELITY AS WELL AS VISUAL QUALITY
Fidelity means how well the output preserves properties that should remain faithful to the input or reference. In this lesson, always state which properties you mean.
A building rendering could be attractive yet move a doorway. A product illustration could look polished yet alter a logo or the shape of a handle. Those are important failures if the brief requires accuracy.
Review source and output together. Check essential geometry, identity, layout and any text or brand details, then assess the requested style. A high score for attractiveness should not hide a critical preservation failure.
15 | COMPARE THE COMPLETE WORKFLOW
To investigate a transformation, record the source image, prompt, model version and relevant settings. Use several outputs rather than judging only the most successful candidate.
Compare requested change and required preservation separately. Also record preparation, generation, review and editing time. A transformation that produces attractive candidates quickly may still require expensive repair.
If comparing settings, change a relevant factor while keeping others as consistent as possible. Seeds may help control one source of variation in a compatible system, but do not guarantee identical results across tools or versions.
16 | USE SOURCE IMAGES RESPONSIBLY
Before using client material in a service, Visionary Studios should investigate permissions, confidentiality and data handling. The right to view an image does not by itself establish permission for every transformation or upload.
A person’s appearance or a product’s features can be changed in misleading ways. The studio should consider whether disclosure and review are appropriate for the intended use.
These are evaluation questions, not a conclusion that a particular service or output is legally acceptable. The case study’s data, bias and transparency concerns remain relevant even when generation starts with a supplied image.
17 | APPLY IT TO VISIONARY STUDIOS
Image-to-image translation could help the studio turn early sketches into candidate renderings or explore new treatments of existing visual assets. Its usefulness depends on whether it preserves what the project requires.
The booklet does not report actual productivity results. Describe these as proposed uses and identify the evidence needed to justify adoption.
18 | WRITE A TRANSFORMATION BRIEF
Choose either a building sketch, a greyscale campaign image or a photograph to become an illustration. In your book, state the source, target appearance, three properties that must be preserved and two properties that may change.
Then propose a review method and explain one limit of the available source information. No image upload or generator account is needed.
Reveal a possible response
Source: a building sketch. Target: a realistic-looking concept at sunset. Preserve the silhouette, doorway location and required window arrangement. Allow materials and lighting to vary. Review source and output side by side and record any moved or invented features. The sketch does not uniquely specify wall materials, so a generated surface is a proposed interpretation.
Compare several candidates and include correction time when judging the workflow.
Read alongside the image-to-image examples in the official 2027 case study. The detailed classroom briefs here are teaching extensions.
CHECK YOUR UNDERSTANDING | WHAT CHANGES AND WHAT STAYS?