Guides

    8 image models that will letter a label you can read

    Type in the still is a generate of letters, not a caption overlay. Eight published still rows that actually letter: Ideogram V4 (6cr), Seedream 5 Pro (7cr), Recraft V4 (5cr), Recraft 4.1 (4cr), Qwen Image 3 (4cr / 2K), Qwen Image (2cr / 4K), GPT Image 2 (4cr / 4K), GPT Image 1.5 (2cr / 4K).

    Versely Team6 min read

    Type in the still is a generate of letters, not a caption overlay. Eight published still rows that actually letter: Ideogram V4 (6cr), Seedream 5 Pro (7cr), Recraft V4 (5cr), Recraft 4.1 (4cr), Qwen Image 3 (4cr / 2K), Qwen Image (2cr / 4K), GPT Image 2 (4cr / 4K), GPT Image 1.5 (2cr / 4K).

    A label you can read was generated as glyphs. A hook you can edit was burned later. Mixing those jobs is how a pack shot ships with a cousin brand, or how a legal line gets baked into pixels you cannot restyle. Run the eight from text-to-image when the letters are the picture. Run add a text overlay when the letters have to stay copy.

    Letters are a generate, not a caption

    Diffusion used to treat type as texture. Short display text is now a real job on a handful of stills rows: quote the string, describe the type as a brief, and the model draws letters. That is the label, the badge, the poster headline, the bottle face.

    It is not a subtitle. It is not a CTA you will change next Tuesday. Overlay is a later editing pass that stamps your words on a clip. If the wording might change, do not generate it into the still. If the wording is the object — a SKU name on a pack, a wordmark on a tote, a three-word poster — generate it.

    Cap the string. Under six words, expect a readable first take. Seven to twelve, budget a retry. Past that, generate empty space and overlay the real type. Zoom to 100 percent and read the letters aloud before you animate anything.

    The eight

    1. Ideogram V4 — 6 credits

    Ideogram V4 is the poster and logo row: crisp stills, accurate text rendering, fine detail. Six credits. No resolution listed in the catalog, so do not invent 4K. Use it when the words are the design — a badge, a thumbnail headline, a wordmark exploration. Quote the exact string. Keep it short. Ideogram is the graphic-design instinct on this list, not the photoreal pack photographer.

    2. Seedream 5 Pro — 7 credits

    Seedream 5.0 Pro is ByteDance's flagship stills row: deep-thinking prompt understanding, native text rendering in 14 languages, dense layouts. Seven credits. Use it when the label is not English, or when the type has to sit on a physical object — foil, glass, embroidery, a chalk board. Fourteen-language lettering is the reason this row is on the list. It is not a video model. Audio is null.

    3. Recraft V4 — 5 credits

    Recraft V4 is the design-quality stills row with typography support. Five credits. Recraft's job is layout-shaped graphics: icons, brand frames, merch comps, type that sits like a designer placed it. If Ideogram is the poster wall, Recraft is the brand kit sheet. Quote the mark. Do not ask it to invent a second slogan in the corner.

    4. Recraft 4.1 — 4 credits

    Recraft 4.1 is the cheaper Recraft stills row: high-quality generation, style control, consistency. Four credits. Same family, lower listed rate. Use it when you already know the Recraft look and you are iterating the lettering, not when you need V4's extra design punch. It is still a generate of letters. It is not overlay.

    5. Qwen Image 3 — 4 credits, 2K

    Qwen Image 3 is Qwen's third-generation stills row: strong prompt adherence, complex text in Chinese and English, prompt rewriting on by default. Four credits. Max listed resolution is 2K, not 4K. If the brief is a bilingual label or a dense English lockup, this is the Qwen current gen. Do not send it a 4K delivery spec it does not sell.

    6. Qwen Image — 2 credits, 4K

    Qwen Image is the budget 4K stills row at 2 credits (discounted). High-detail realistic stills. Use it when the label is short, Latin, and you need 4K without Qwen 3's 2K ceiling. If the type is complex or bilingual, spend the extra 2 credits on Image 3. Two credits is not a reason to skip the read-aloud check.

    7. GPT Image 2 — 4 credits, 4K

    GPT Image 2 is OpenAI's high-fidelity stills row: strong prompt adherence, up to 4K. Four credits, discounted. Quote the label in the prompt the way you would brief a retoucher. This row follows instructions; it does not know your brand kit. If the letters come back as a cousin spelling, regenerate — do not "fix it in the video model."

    8. GPT Image 1.5 — 2 credits, 4K

    GPT Image 1.5 is the cheap OpenAI stills row that is explicitly good at legible in-image text, plus image-to-image and edits. Two credits. 4K. Use it when the brief is long and specific, or when you already have a pack still and you need the letters corrected rather than a new scene. Edits are a stills job. They are not a caption pass.

    Quote the string, then stop

    The prompt pattern that actually letters:

    • Quote the exact text: the label reads "COLD PRESS".
    • Describe type like a brief, not a font menu: bold condensed sans, tight tracking, all caps.
    • One text element per clause, with a place in the frame.
    • Say what is not text, or the model will letter the wood grain.

    Do not ask a video row to typeset. Veo, Seedance, Flux — those are motion. They will smear a lockup you could have generated for 2–7 credits. Letter the still. Then, if the shot needs to move, image-to-video the plate.

    Overlay is the fourth door

    When the type must stay editable, skip the eight. Add a text overlay burns a line you wrote — a hook, a CTA, a badge — onto a clip. That is copy, not glyphs in the pixels. Legal disclaimers, licensed brand fonts, and anything you will restyle next week belong here.

    The two jobs can sit in one delivery: generate the pack with Ideogram or Seedream so the label on the bottle reads, then overlay the hook so the hook can change. Do not generate the hook into the bottle. Do not overlay the SKU name and call it a pack shot.

    FAQ

    Can I letter a label in a video model and screenshot it?

    No. Video rows are a bad typography engine. Generate the still on one of the eight, read it, then animate if you must.

    Which of the eight is 4K?

    Qwen Image, GPT Image 2, and GPT Image 1.5 list 4K. Qwen Image 3 lists 2K. Ideogram V4, Seedream 5 Pro, Recraft V4, and Recraft 4.1 do not list a max resolution — do not invent one.

    When do I use overlay instead of these rows?

    When the words have to stay copy: hooks, CTAs, legal, a line you will change. Overlay stamps text you wrote. These rows generate letters into the picture.

    Is Recraft 4.1 the same as Recraft V4?

    No. V4 is 5 credits with the stronger typography pitch. 4.1 is 4 credits on a separate slug. Same family, different listed rate. Pick on purpose.