Imagen vs. gpt-image-1: AI Image Models in a Real-World Test

All models hang off the same MCP tool here; I only switch the provider and model:

• Imagen 4 (Google, imagen-4.0-generate-001): text→image only. Controlled via aspect_ratio; at 16:9 it returns a 1408×768 image.
• gpt-image-1 (OpenAI): text→image and image edit (images.edit), so it can process reference images. Fixed output sizes; at 16:9 a 1536×1024 image.
• gemini-2.5-flash-image (Google): the image-capable Gemini model, which also takes reference images. My tool switches to it automatically as soon as a reference image is supplied and provider gemini is selected.

The decisive difference is already here: two of the three accept an image as input – Imagen does not.

The task: a tablet showing the Home Assistant logo on its screen. The two image-capable models (gpt-image-1 and gemini-2.5-flash-image) were given the real logo as a reference image – exactly the same file. Imagen got only a text description, because it can't take an image as input at all.

Clear ranking: gpt-image-1 > gemini-2.5-flash-image > Imagen. Both image-capable models clearly beat Imagen – logically so, because this is an architectural limit, not a slip-up: without image input Imagen has to reconstruct the logo from words and reliably fails on brand marks. gpt-image-1 matches the original most precisely; gemini-2.5-flash-image gets close but interprets the logo more freely. As soon as a brand, logo or specific product has to look exact, there's no way around a reference image.

An icon is one thing, text is another. Image models are notorious for mangling type. So here's the harder test: our own devmaker.net logo including the wordmark as a reference – again the two image-capable models got the real logo, Imagen only the description.

This is exactly where the reference image pays off most. gpt-image-1 is the only model that reliably nails symbol and lettering together. gemini-2.5-flash-image gets close but stumbles in the typical way on text – a single missing letter is enough to break a logo. And Imagen shows why “just describe the logo” isn't a solution: it can type the letters, but it can't invent the brand identity. For anything involving text – wordmarks, labels, UI strings – gpt-image-1 with a reference is currently the only reliable choice.

Same prompt, deliberately without any brightness instruction: “developer workspace at night, moody atmosphere”. The result flipped my expectation.

Lesson from practice: both models can come out too dark – but gpt-image-1 interprets mood words far more aggressively. Even if you want dark hero images (like our terminal theme here), you still have to state the brightness explicitly: “well-lit subject, clearly visible, balanced exposure”. Otherwise all you end up seeing is the brightest element in the frame.

For conceptual heroes (agents, pipelines, “X vs. Y”) there's no photographable subject. Same prompt to both: a glowing “agent loop” core, explicitly no text, no letters.

Both deliver usable results here and – importantly – neither smuggles in text. The “no text” pattern works for both, but it's no guarantee: as soon as the subject contains type (buttons, diagram labels, a keyboard), both happily produce mangled pseudo-letters – see the dropped letter in the wordmark above. Tendency: Imagen leans towards a crisp stock look, gpt-image-1 towards illustration.

Imagen controls the format via aspect_ratio (16:9 → 1408×768), gpt-image-1 only knows fixed sizes (16:9 → 1536×1024). gemini-2.5-flash-image simply ignored my 16:9 setting in both reference tests – once square (1024×1024), once extremely wide (2048×512). Nasty when you're filling fixed hero slots; so double-check before cropping.

The pitfall that cost me a failed generation: if you switch the provider but forget the model, the resolver keeps pulling the default model from the settings – and that's an Imagen model:

A quick note so the setup stays honest: the production Wagtail site runs on a netcup server. These experiments, however, run in my dev environment locally on a small mini PC in the homelab – that's where the MCP tool issuing the image calls lives. The image models themselves run in the cloud anyway; the mini PC only orchestrates. For exactly this kind of always-on dev box I use this one:

• Exact cost & latency: both vary with load and size – without a clean measurement setup any numbers would be guesses. Anecdotally: gpt-image-1 a touch slower, Imagen 4 brisk.
• Midjourney / Stable Diffusion: deliberately left out – I cared about the models that hang directly off my tool.
• Other formats: I tested 16:9 heroes (apart from the gemini outliers). With square/portrait the result may differ.

• Imagen 4 is my default for abstract, bright heroes with a sharp, clean look – fast and fuss-free, but without reference images.
• gpt-image-1 is what I reach for as soon as a reference image is involved (logo, product, brand) or a mascot needs to stay consistent across several images. Best logo and, above all, text/wordmark fidelity. The price: prompt brightness explicitly, otherwise it goes too dark.
• gemini-2.5-flash-image is the fast reference-image alternative from the Google camp – close, but freer in its interpretation, prone to dropping a letter in text, and with a mind of its own on format.