A balcony view in Penang, rebuilt as a single imagined day. Three photographs were merged into one wide panorama, then the sky above it was rebuilt out of many separate photographic prints. The place stays whole. The time of day does not, and the joins are meant to show.
The first decision mattered most, and it came before any collage was attempted: the three source photographs were merged into one wide panorama. That merge became the base to return to whenever detail, density or variation started to pull the view apart.
It was never meant to be the finished picture. It was the map of the place: towers in their zones, the sea at the right height, roofs connecting, the hill on the right. Once that held still, the sky above it was free to break time apart.
| Source role | What it locked |
|---|---|
| Left daylight | Cool-blue sky, the left tower cluster, red roofs, and the first sea read. |
| Middle bridge | The central horizon and sea hinge connecting the two times of day. |
| Right sunset | The orange horizon, wooded hill silhouette, sunset light, and right-side depth. |
The reference was David Hockney's Pearblossom Hwy., used to learn how photographic collage behaves rather than as a picture to copy. It is useful because its joins have a hierarchy. The sky is a dense ceiling of small separate photographs. The ground is broken by fewer, larger prints of specific things. The whole stays readable even though the light and the viewpoint shift from piece to piece.
Penang kept those rules and threw the subject matter out. No roads, signs, text or desert motifs, and nothing from the land allowed to drift into a sky-only piece. The reference set how the picture should behave, not what it should look like, so the result still reads as Penang.
The scorecard was the most useful tool in the project. It turned quick visual reactions into criteria that survived across many rounds, conversations, contact sheets and restarts.
The split of work was explicit. Human judgement set the quality bar and overruled the scores when they were wrong. The AI system turned that judgement into a stable way to compare candidates, and into tighter prompts, review notes and the next round of experiments. The scoring was marked harder when the early collage scores proved too generous, because a high collage score means nothing while the picture is still too smooth, too uniform, or not holding together.
| Criterion | What it protected |
|---|---|
| Still one place | The city, sea, hills, towers and roofs have to read as a single view. |
| Day into sunset | The left stays cool, the centre carries the change, the right stays warm. |
| Enough sky pieces | The sky needs enough separate photographs to look built rather than blended. |
| Prints, not overlays | Each piece looks like a photograph with its own exposure and crop, not an effect laid on top. |
| No two alike | Neighbouring pieces should not repeat the same clouds, weather or tone. |
| Big against small | Large anchors and smaller sky prints together give the picture a change of scale. |
| Joins, not a grid | The joins stay visible without turning into a quilt or a rigid grid. |
| Clean | No dark mud, floating objects, foreign motifs, hard frames or destructive edges. |
These are not finalists in one style. Each was kept because it solved a particular problem better than the others, so together they worked as a reference library. A new prompt could take its structure from one, its sky from another, its atmosphere from a third, and its limit on how far to push the collage from a fourth.
Several routes never reached the final artwork, and each one still shaped it. They were worth the time because they changed the method rather than simply being thrown away.
| Route | Why it failed | What it taught the next attempt |
|---|---|---|
| Sliced-panorama overlays | Looked like an effect applied on top of one smooth image. | Individual pieces need their own photographic content, not merely a crop boundary. |
| Hard outlines and card-like frames | Made the image feel decorated rather than assembled. | Preserve soft internal edge falloff and keep the outer crop clean. |
| Many overlapping thin rectangles | Read as a quilt or technical grid rather than a sequence of photographs. | Use mostly square pieces with clear roles and limited overlap. |
| Uniform land tiling | Destroyed subject hierarchy and made the city feel mechanically patterned. | Keep land slower: a few larger, object-led anchors with support pieces. |
| Dark real-photo insertions | Produced muddy patches that pulled attention away from the overall view. | Reject dark crops early and match each source crop to its surrounding tonal zone. |
| Broad manual sky replacements | Source skies were too pale and incompatible at a large scale. | Place real crops only inside established cells, matched for local colour and exposure. |
| Land-only regeneration | Flattened the work back into a conventional smooth panorama. | Treat the sky as the active collage register and protect the land spine. |
| High-resolution restoration alone | Improved polish but reduced the evidence of photographic assembly. | Resolution work must preserve, not erase, the joiner language. |
The selected artwork is not the most polished thing the project produced. It is the one that carries the whole decision history, because every rejected route narrowed what the next attempt was allowed to do.
Long visual sessions exposed a practical problem. Opening or attaching dozens of full-resolution renders makes a chat session heavy and fragile, and the attachment link of a generated image is not an archive. So storage was treated as part of the creative system rather than as a backup chore.
That helped in two ways. A crash or a restart no longer lost the work, because every candidate had a stable, meaningful path and could be found again. And light previews cut the memory load while still allowing fast side-by-side review. Keeping the creative history out of the chat window is what made long sessions survivable.
The final artwork is presented on a clean white field with the title set beneath it, and prepared as a single large-format master.
One_Window_All_Day_Penang_WALL_MASTER_12000px_150ppi.jpgThe artwork was generated at a much smaller size, so the 12,000 px file is a good enlargement, not new detail. It holds up at the distance you would stand from a wall, but a print bureau proof is still the right check before committing to anything at that scale.
The picture was made with AI image generation, run in a loop, and the composition that became the final artwork is a generated render. What the machine did not do is decide anything: the panorama that fixed the geography, the criteria for a good result, the decision to mark those criteria harder when they proved too generous, and the repeated rejection of routes that scored well but looked wrong.
The work went best when three things stayed in balance: a human eye, a fixed panorama underneath, and a scorecard that kept its memory between rounds. That combination made it possible to push hard without losing either the place or the standard.