Images · 9 Oct 2026 · Prompt Punks

How AI image models have improved

From early Midjourney experiments to GPT Image and Nano Banana: a visual look at detail, editing and creative control.

Three glass panels show a blue cup progressing from a fragmented silhouette to a fully resolved ceramic form. — Conceptual AI-generated editorial illustration—not an example from the models compared below.

The early promise of AI imagery was possibility: type a few words and see something unexpected. The more useful development is control. Can you preserve a subject, change a specific detail and carry an idea through several revisions?

This visual tour looks at early Midjourney models, GPT Image and Google's Nano Banana. It is not a league table. The examples come from different people, prompts and publication dates; they illustrate changes in capability, not a fair head-to-head test.

How to read the examples. Every image below is externally hosted and credited. Select “Load examples” to contact its host. These are published demonstrations, not generations made or independently benchmarked by Prompt Punks.

Early Midjourney: the idea appears before the detail settles

Midjourney's legacy documentation dates V1, V2 and V3 to 2022. It describes V1 and V2 as painterly with low coherence, and V3 as more coherent. Those early models were compelling tools for exploring atmosphere, even when an image needed interpretation rather than close inspection.

Cape Town generated with Midjourney V1 in Michael Howe-Ely's published comparison. — The earliest version in this selection.

Cape Town generated with Midjourney V3 in Michael Howe-Ely's published comparison. — Compare the organisation of buildings and landscape.

Cape Town generated with Midjourney V6 in Michael Howe-Ely's published comparison. — A later version, not a claim about today's latest model.

Prompt: “Cape Town”. Examples: Michael Howe-Ely, published 14 January 2024. These are later runs of different model versions, not images documented as generated in 2022. Seeds, settings and selection can affect the outcome.

Look for the relationships between objects, not just detail. Does a building remain a building when you zoom in? Do foreground and background form a plausible space? A convincing generated city is still not a reliable photograph or geographic record.

Hands: a familiar test of structure

Hands make a useful comparison because they need more than convincing texture: fingers, joints and contact with an object must fit together. In this historical experiment, RBC Trends used the same prompt, “close-up hands playing piano”, across Midjourney V1–V5. Each image below shows the source's four-output grid rather than one chosen favourite. Inspect the variation within each version as well as the differences between versions; a polished surface does not guarantee correct anatomy.

Hands playing piano generated with Midjourney V1 in RBC Trends' published experiment. — Start with whether the shapes read as hands and fingers at all.

Hands playing piano generated with Midjourney V4 in RBC Trends' published experiment. — Inspect finger separation, bends and their relationship to the keys.

Hands playing piano generated with Midjourney V5 in RBC Trends' published experiment. — More detail does not remove the need to check anatomy and contact.

Selected examples: RBC Trends / Ivan Zvyagin, original V1–V5 experiment. Images credited by the source to Midjourney. Open an image to inspect it at full size; the source includes V2 and V3 as well. Settings, seeds and selection may differ; this is not an independently controlled Prompt Punks benchmark or a claim that hands are now solved.

GPT Image: a specific change, without starting over

OpenAI's GPT Image 1.5 cookbook demonstrates a different kind of task: taking an existing scene and changing its conditions. In the example below, a billboard scene is edited into a snowy evening. Compare the layout, product and lettering as well as the added weather. Source: OpenAI's published prompting guide.

GPT Image 1.5 cookbook example showing a generated shampoo billboard scene before its weather edit. — The generated scene used as the next edit's input.

The cookbook's billboard scene edited into a snowy winter evening. — A weather and lighting change applied to the scene.

Examples: OpenAI. Model: GPT Image 1.5. This is a documented version-specific demonstration, not a claim that 1.5 is the latest GPT image model.

For a designer, that suggests a practical workflow: approve an image, identify exactly what needs to change, then inspect what the edit altered unintentionally. An attractive result is not enough if the product label, proportions or composition drift.

Nano Banana: changing the treatment while keeping the subject

Google's original Nano Banana showcase demonstrates image editing and subject preservation. Its cat-to-game-character example makes the intention easy to read: translate a recognisable subject into a different visual language. Source: Google's September 2025 showcase.

Google's Nano Banana example pairs an orange-and-white cat photograph with its transformation into a pixel-art game character. — Input and output published together by Google. Original Nano Banana showcase; not a Nano Banana Pro or Nano Banana 2 test.

“Nano Banana” now describes a family, not one timeless model. Google's Nano Banana 2 announcement describes a later generation. Keep the actual version attached to each example: otherwise an old screenshot can quietly become a misleading claim about a newer release.

The improvement that matters is usable control

Sharper images are welcome. But the creative question is whether the output survives the brief. Can you maintain identity, reserve space for a headline and make a correction without losing the qualities you approved?

Structure: inspect hands, edges, repeated objects and spatial relationships.
Instruction: compare the result with the actual brief, not your first impression.
Consistency: put related outputs side by side.
Accuracy: check every product detail and every word.
Finish: judge the image at its intended crop and delivery size.

A fair studio test would use the same brief, declared references, comparable formats and a recorded number of attempts. Show failures as well as favourites. These published examples are a starting point for that conversation, not a substitute for it.

Better tools expand the options. They do not decide which option is worth using. That remains the job of the person directing the work.

Put the control to work

Read Consistent campaign images with AI for a practical next step, or talk to Prompt Punks about your next visual brief.

Sources and example labels checked 9 October 2026. Historical and provider-selected demonstrations; no paid endorsements. External image availability can change.