ai-image-generators
The video argues that AI image generation follows a difficulty-and-value pyramid. Model choice barely matters at the easy, low-value bottom. At the top, where a real object or photo must be reproduced faithfully, accuracy and restraint decide, and in the September 2026 tests GPT-2.5 won clearly, with Midjourney last.
Bernlieger frames generative AI as a pyramid. At the bottom, generic prompts like a dune in the desert are easy for every model and carry almost no value, because there are a million right answers. Each added detail narrows the correct outputs, which raises both difficulty and perceived value.
He tests four models (GPT 2.5, Nano Banana 2, Grok Imagine 2, Midjourney V8) on eight scenarios that climb the pyramid. Each scenario uses the same prompt and source for every model, and many sources are deliberately poor iPhone shots. He ranks the models per scenario and opens with a disclaimer that, at September 2026 speed, the results could be entirely different in a few months.
At the low end (stock-style Alps convertible, family beach photo, impossible-photography thumbnails) the rankings are taste-dependent and scatter. Grok wins the Alps and is last on the beach, and GPT does the reverse. The faults the video flags are physical plausibility errors: driving side, footprints, finger anatomy, a third child, guardrails that change. Stock imagery tolerates these because the result does not have to be exact.
Once a real-world object enters the prompt, fidelity outranks beauty. Grok gives the best-looking watch ad but alters the dial. GPT reproduces the watch exactly. Midjourney gives stunning scenery but fails the main idea of keeping the real object's visual integrity, so it ranks last on both watch tests.
The portrait test is the exception. Nano Banana is the only model that produces a professional-looking result from a bad selfie, because it is willing to change the hair, lighting and contrast, though it drifts furthest from the source. The same willingness to alter hurts it on the preservation-style edits.
At the top of the pyramid (people removal, summer-to-winter, interior retouching) the winning behaviour is judgment and restraint: change what was asked and nothing else. GPT leads all three. It knows what to remove and what to keep, leaves the sky and building untouched, and delivers an interior edit the speaker calls astounding. Grok over-removes, Nano Banana changes pose, crop and sky, and Midjourney fails or needs a manual tool. The speaker says this makes him nervous about his job as a hospitality photographer.
Scoring 4/3/2/1 points per scenario, GPT-2.5 is the clear winner, Grok Imagine and Nano Banana sit close together in the middle, and Midjourney is last.
The speaker's usage recommendation follows the pyramid. At the bottom any model works, and Midjourney's cinematic, less-AI look is an advantage. Higher up, accuracy matters and Midjourney cannot deliver it, so Grok and especially GPT-2.5 are the picks. GPT-2.5 offers pixel-perfect edits and the most consistency, while Nano Banana is smarter than Midjourney but lacks consistency and accuracy.
A final clock test (an analog wall clock showing exactly 9:15) supports the accuracy ranking. Midjourney looks cinematic but shows the wrong time, Grok draws three hands, and only GPT-2.5 consistently shows the right time. The speaker says correct time on a clock or watch is still something AI really struggles with.
The rubric changes with the pyramid level. At the bottom the judge is aesthetics, such as light direction and realism. At the top it is fidelity and not touching what wasn't requested, so the best-looking output can be the wrong deliverable for client work.
Nano Banana's willingness to reinterpret the image is both its win and its loss. It wins the selfie-to-portrait task and chooses a more pleasing sky for the winter edit, but it alters pose, crop and building details in tasks that call for preservation. Whether that trait is a strength depends on whether the task wants transformation or preservation.
Scope judgment in edits is its own capability. In the museum cleanup GPT removed the security entrance check and kept everything else, while Grok removed literally everything and invented content. The speaker calls GPT's discernment next level.
Errors are weighed by how fixable they are. A scratched dial, melted watch text or '200 SAR instead of bar' can be repaired in post, but a changed building or lost scene integrity cannot. This is why GPT, which changed nothing in the glacier scenery, outranks the prettier Midjourney and Nano Banana images.
By the per-scenario rankings, GPT-2.5 never placed below third, while Grok went from first (Alps, watch) to last (beach, cleanup). The case for GPT rests on consistency across tasks more than on winning every one.
Capability moves fast. The speaker says Grok's winter-edit result is far better than it was half a year earlier, when it kept changing the building on a real client job.
Midjourney's prompt-based cleanup could not do the museum task at all, but its manual paint-over removal tool worked quite well, though still worse than GPT.
Apple's built-in iPhone cleanup left a messy floor, far worse than the other tools. The speaker attributes the gap to local versus cloud processing.
The input still matters. The speaker notes that with more effort on the original selfie (hair, cleaner lighting) AI could go a long way, so the poor sources used in the tests understate what careful inputs would yield.
«they can hardly be wrong. There are million different right answers because there are million different sand dunes»
— 00:17
«Now, when we add specific details to the prompt, the difficulty as well as the perceived value of that image increase»
— 00:24
«it's September 2026 and these things move very fast»
— 01:01
«It fails the main idea here, which is incorporating a real-life object accurately and keeping its visual integrity.»
— 11:41
«To know what to remove and what not is absolutely next level. I'm really impressed.»
— 15:56
«So, that's local versus cloud, I guess.»
— 17:03
«Stuff like that really makes me nervous about my job as a hospitality photographer.»
— 18:19
«Midjourney cannot be accurate, sadly, which is where Grok and especially GPT-2.5 really shine up here.»
— 23:11
«Midjourney, in my opinion, produced the most cinematic-looking image, but it's just not 9:15.»
— 23:56
«having the right time on a clock or watch is still something that AI really struggles.»
— 24:16
Reception
Only two low-engagement comments: one neutral-to-helpful viewer-added timestamp list and one confused, mildly negative remark, so reception is mixed and sparse.
A practitioner-led, scenario-based ranking that places GPT-2.5 first for fidelity-critical editing, Grok and Nano Banana in the middle, and Midjourney last outside low-stakes cinematic generation. The speaker states that the specific rankings are dated to September 2026.

25:27