Which image model should you use?

Aether runs seven image models and you pick one per design. The default, FLUX.2, is the right answer most of the time. The rest exist because a photoreal living room, a magazine-styled interior and a quick exploration are not best served by the same renderer — and because when one model keeps misreading a brief, a different one usually does not.

The choice changes how the room looks. It does not change the specification: dimensions, materials, lighting and the standards checks are written before any model runs, and the render is judged against them afterwards whichever one you chose.

The seven

Image models compared by strength, cost and speed
ModelWhat it is good atTokensDraftApprox.Photo redesign
FLUX.2DefaultBest material and lighting realism. Sticks closest to the brief.10516sYes
GeminiFollows the brief most literally. Strongest real-world design language.131212sYes
Recraft V3Most photographic — real depth of field and material detail.13119sYes
Ideogram V3Warm, staged, magazine-like. Slowest of the set.12921sYes
QwenBalanced alternative. A different eye when FLUX keeps missing.1179sYes
TurboFastest and cheapest. Flatter, and more prone to inventing objects.936sNo
Gemini ProThe best available, and priced like it — a design runs about 65 tokens.252522sYes

Tokens are charged per image, and a verified design renders up to three times to clear the quality bar — so a design costs the token price multiplied by the number of attempts it needed. Draft mode skips the reasoning and the judge, which is why it is cheaper. Times are the render itself, not the full design.

What to reach for

You are not sure
FLUX.2The default. Best material and lighting realism, and it stays closest to what you asked for.
It keeps adding things you did not ask for
GeminiFollows the brief most literally, which attacks that problem at the source rather than leaving the judge to catch it afterwards.
You want it to look photographed
Recraft V3Real depth of field and material detail — the most convincing as a photograph rather than a render.
You are exploring, not committing
TurboFastest and cheapest by a wide margin. Flatter, and more prone to inventing objects, which matters less when you are only testing a direction.
The others keep missing
QwenA genuinely different eye. Worth trying when a brief is not landing rather than rewording it a fifth time.

Why seven, and not one

Because the renderer is the part of this that is replaceable. Image models improve every few months and each has a different failing — one flattens materials, one invents furniture nobody asked for, one is slow but stages a room beautifully. Committing to a single one would mean inheriting its particular weakness permanently.

What does not change is the part in front of and behind the render: the specification that says what the room must contain, and the judge that checks whether the picture actually delivered it. A design scores out of 100, must clear 82 to be delivered, and is re-rendered if it does not. That is the same whichever of the seven drew it.