Nano Banana 2 vs Imagen 4 vs Ideogram V4 for a label

Update, 8 October 2026: Imagen 4 is retired on Versely. Google shut down the Imagen 4.0 endpoints on 17 August 2026 (Gemini API deprecations), and Versely no longer offers it. Where this post picks Imagen 4, use Nano Banana 2 for the same still. What Google retired, and when.

Ideogram V4 at 5 credits letters the mark. Imagen 4 (retired on Versely) was a photoreal still, text-only input — it would invent a cousin label. Nano Banana 2 at 4 credits (discounted from 7) is an edit of a still you already have. Overlay is the fourth door when the type must stay editable.

This is not a general image-model bake-off. Texture, faces, and arena Elo are a different test. The job here is a label: a short string that has to be the string, on a mark, a poster, a bottle face, a quote card. Pretty grain that misspells the brand is a fail. Read the output aloud at 100 percent zoom before you talk about light.

Three rows, one job

A label job has three honest inputs. Pick the row that matches the input you actually have. Mixing them is how you pay for a photoreal cousin of a mark you already approved.

You have a quoted string and no plate. The words are the design — a logo lockup, a poster headline, three words on a pack face that does not exist yet. That is Ideogram V4: text-to-image, 5 credits, built for lettering. Quote the copy. Proofread like a print vendor.

You have a paragraph and you want a photograph of an object. That was Imagen 4, retired on Versely: text-to-image only, 4 credits, 4K, photoreal. It would not take your existing pack as an image input. Describe a bottle and you get a bottle. The label on it will be a plausible cousin, not the mark legal signed. Use a photoreal text-to-image row, now Nano Banana 2, when the still does not need readable proprietary type.

You have a still someone already approved. That is Nano Banana 2: 4 credits (discounted 50% from 7), text-to-image plus image-to-image plus edit-image. Point at the object. Change one attribute. Do not throw the approved file away for a new text-to-image lottery.

You will change the line next week, or the font is licensed, or the line is legal. None of the three. Stamp real type after the picture exists: add a text overlay. Overlay draws the string once. A generate redraws it every time.

Run all three still rows from text-to-image. Run Nano Banana edits from the AI photo editor. Do not run this test on a video model. Video credits are a terrible typesetter.

What each row actually does

Ideogram V4 — 5 credits, text-to-image. Catalog description: posters and logos with accurate text rendering. audio is null; this is a still. Quote the exact string. Unquoted copy gets paraphrased into a cousin slogan. Check spelling, hierarchy, and survival at the real size. If you have to zoom to know what it says, it does not say it. Ideogram is the row when the mark is the picture — a lockup, a wordmark, a poster where type is the subject. It is not a photoreal product photographer that happens to spell. "This glass bottle, this bar, and also the word VERSELY in the house font" is two jobs. Split them.

Imagen 4 (retired) — 4 credits, 4K, text-only. Google's photoreal still row until it was retired. Categories: text-to-image, nothing else. No image-to-image. No edit. You cannot hand it the pack and say "keep this label." You can hand it a paragraph and get a beautiful object that is not your object. Imagen 4 was for a scene or a material study where the type is incidental (use Nano Banana 2 for that now) — then do not require that incidental type to be the brand. Soften background signage; do not proofread a ghost word on a 4K billboard the model invented. If the label is the product, a photoreal text-only row is the wrong door no matter how good the glass looks.

Nano Banana 2 — 4 credits (discounted from 7), edit the file you have. Google's lightweight iteration row. Categories: text-to-image, image-to-image, edit-image. You can run it from text. Once a still is approved, treat that as a reason not to. "Make the cap matte black, keep the label" is an edit. "A cinematic bottle on wet marble" is a new still. The Nano Banana 2 edit page is that rule. For a label: one string change per call. "Change SUMMER to AUTUMN, keep the mark" can work. "Restyle the whole pack and fix the kerning" is how you get a third bottle. Type on Banana 2 is competent, not the reason you opened it. The reason is the plate.

None of these rows is a licensed typesetter. Past about six words, expect retries. Past a paragraph, stop. Overlay.

Overlay when the type has to stay editable

Generated type is pixels. Pixels do not have a font file. If next Tuesday's brief is a different price, a different language, or a legal line the lawyer will rewrite, you do not want to regenerate the plate to change three characters.

Add text overlay to video is the later step: one authored line, burned at top, center, or bottom, the same glyphs on every frame. It is not transcription. It is not a second Ideogram pass.

The hybrid that actually ships:

1. Generate or photograph a clean plate — empty upper third, or a real pack whose type is already correct and will not be redrawn.
2. If the mark has to live in the scene (perspective on a shop sign, print on a garment), Ideogram the still, proofread, then image-to-video from that file. Do not ask the video model to typeset.
3. If the mark has to stay editable, overlay on the clean plate after picture lock.

Do not stack two typesetters. An Imagen bottle that invented "VERSLEY" plus an overlay that says "VERSELY" is two labels in a fight. Generate empty. Stamp once.

GPT Image 2 versus Nano Banana 2 for legible type is the scene-with-words test on a different pair. Stay here when the decision is Ideogram vs Imagen vs Banana 2 vs overlay.

A one-hour bake-off

Same quoted string. Same aspect. Read aloud. Ignore everything else until spelling passes.

Prompt A — mark as design. Logo lockup on a cream field, the word "NORTHMADE" in bold sans, small line "EST. 2019" under it, no other lettering. Run Ideogram V4 and Nano Banana 2. Keep the one you can read. A photoreal row will often win on paper texture and lose on the word. That is a fail.

Prompt B — pack in a scene. Photoreal glass bottle on a wet bar, 85mm, label facing camera. Run Nano Banana 2 as a text-to-image. Do not require the label to be a real brand. Run Ideogram only if the brief was the label, not the bar. If you already have a pack photo, skip both and edit on Nano Banana 2: replace the flavour line with "Yuzu", keep the mark and the glass.

Prompt C — the line will change. Generate or shoot a clean plate. Overlay Yuzu · 250ml on the finished clip. Change it to Yuzu · 500ml without touching the model. That is the overlay door, and it is the only one of the four that does not spend a generate to edit a string.

Budget the hour as six stills, not sixty. Two Ideogram, two Banana 2 generates, two Banana 2 edits. The winner is the file a stranger can read, plus a path to change the line without a new lottery.

FAQ

Which of the three spells the brand most reliably?

Ideogram V4 is the lettering row at 5 credits. Proofread anyway. Imagen 4 was photoreal and text-only (it is retired on Versely); it would invent a cousin. Nano Banana 2 is an editor that can keep a label you already have. Run the quoted string. Keep the one you can read.

Could Imagen 4 take my existing pack shot as input?

No, and it is retired on Versely. It listed text-to-image only. A 4K photoreal still from a paragraph is a new object. If the brief is "this file, that one word changed," you are on Nano Banana 2 at 4 credits (discounted from 7).

When is overlay the right first choice, not the fallback?

When the type must stay editable, licensed, legal, or dense. Overlay is a later step on a locked picture, not a way to rescue a misspelled generate. Clean plate, then text overlay.

Should I generate the label in video and pause on a good frame?

No. Video redraws letterforms every frame. A pause is how you find the typo, not how you prevent it. Typeset in a still row, or stamp overlay after the clip exists. Video credits are the wrong meter for a spelling test.