Text Rendering in FLUX.2 (What Works, What Doesn't)

TL;DR: Black Forest Labs names FLUX.2 [flex] as its typography-specialized variant, and its own launch post says complex typography and legible fine text "now work reliably in production." What it does not publish is a character limit, or any claim about non-Latin scripts. Ideogram publishes both. Verified against BFL's own docs and API spec on September 3, 2026.

If your actual question is why AI-generated text comes out garbled at all, for models generally, our diagnostic on garbled AI image text covers the model-versus-prompting split across vendors. If you're specifically on GPT Image 2, our GPT Image 2 prompt generator has that model's own text-in-image rules. This page is FLUX.2 only.

Does FLUX.2 actually render text well now?

Yes, according to Black Forest Labs' own claim, and it is a stronger claim than most vendors make.

Its FLUX.2 launch announcement lists Text Rendering as one of the model's named capabilities: "Complex typography, infographics, memes and UI mockups with legible fine text now work reliably in production." The same page separately claims, under "Output Versatility", that "FLUX.2 is capable of generating highly detailed, photoreal images along with infographics with complex typography, all at resolutions up to 4MP". Both are vendor marketing copy from BFL's own site, not an independent benchmark, but the framing itself is the story: text rendering has flipped from AI's classic failure mode to a feature vendors compete to claim, and FLUX.2 is one of the models making that claim explicitly.

Which FLUX.2 variant should you use for text?

flux-2-flex, by name, from Black Forest Labs' own model comparison.

Its "Compare FLUX.2 Models" table describes each variant's specialty in one line, and flex's is unambiguous: "Specialized for typography." Best for text rendering and preserving small details. Pro, max and klein all sit under "Standard" for the same "Controls" row in BFL's model-choice table; only flex is called out for typography specifically. The pricing page repeats it in three words: FLUX.2 [flex] is positioned for "Fine-grained control, typography".

BFL's own MCP troubleshooting guide, written for people hitting quality problems inside a chat client, gives the same steer as a one-line fix under "Image quality issues": ask for FLUX.2 [flex] specifically when the issue is typography or readable text, and ask for FLUX.2 [max] instead when the issue is a hero shot or final asset. Two different problems, two different named variants.

There's a documented mechanism behind why flex specifically: its steps field runs 1 to 50, and BFL's own launch post shows a side-by-side of the same prompt at 6, 20 and 50 steps, captioned as "trading off typography accuracy and latency." No other FLUX.2 endpoint exposes that dial at all. Pro, max and klein all take a prompt and dimensions, with no way to trade latency for typography accuracy.

What is the exact method for prompting FLUX.2 to render text?

Three steps, stated directly in BFL's prompting guide as its own procedure for readable text: "Enclose in Quotation Marks", "Describe Placement", and "Specify Font Style".

Put concretely: wrap the literal words in quotes so FLUX.2 reads them as a rendering instruction rather than a description of the scene ("COFFEE SHOP" or "Est. 1952" are BFL's own examples); state where the text sits relative to everything else in the frame; and name a font's character rather than a typeface. Its "Text Rendering Tips" section adds four more: front-load the text description early in the prompt, describe colour and effect together ("red neon letters", "gold serif lettering", "chalk on blackboard"), use hex codes for brand-precise colour, and "Keep text short", since the same tip goes on to say "long strings are harder to render accurately".

[SCENE], with the text "[EXACT WORDS]" [PLACEMENT], in [FONT
CHARACTER: bold industrial sans-serif / elegant serif / handwritten
script], in [COLOR OR HEX].

FILLED: A weathered shopfront at dusk, with the text "OPEN LATE"
above the door in hand-painted gold script, warm interior light
glowing behind the glass.

FILLED: A minimalist product label on a matte black tube, with the
text "FIELDWORK No. 4" centred on the front panel in a condensed
sans-serif, white ink, color #F4F1EA.

What do you do when the text comes out wrong on the first try?

Regenerate with one change at a time, not a full rewrite. That's the general iteration loop BFL's own prompting guide recommends for FLUX as a whole, and it applies directly to a failed word of text.

Its own framing: "Strong prompts usually come from iteration, not from trying to write the perfect prompt on the first attempt." The practical loop it gives is three steps: "Start with a simple version", "Check what FLUX got right and wrong", and "Adjust one important detail at a time". Its closing advice generalizes past text specifically: "If the image is close but not there yet, tweak the subject, framing, lighting, or style before rewriting everything."

Applied to a misspelled or garbled word, that loop looks like this in practice:

ATTEMPT 1: A café chalkboard reading "TODAY'S SPECIAL: SOUP"
[if the apostrophe or colon comes out wrong, change ONE thing]
ATTEMPT 2: A café chalkboard reading "TODAYS SPECIAL SOUP" (no
punctuation, since punctuation inside quoted text is where a render
most often slips)
[if the word count is still the problem, shorten before anything else]
ATTEMPT 3: A café chalkboard reading "SOUP TODAY"

Shortening the string is the single highest-leverage change available, since it's the one limit BFL states directly: "Keep text short" because "long strings are harder to render accurately". If a two-word sign still fails after two or three regenerations on the same endpoint, that's the signal to switch to flux-2-flex rather than keep rewording, since flex is the one variant with a documented steps dial that trades latency for exactly this kind of accuracy.

Can you name an exact font, like Helvetica or Futura?

BFL doesn't say no. It just never tells you to say yes.

Its own guidance is to "Specify font character": serif = traditional/formal, sans-serif = modern, script = elegant, display = bold/impactful. Nowhere in BFL's prompting guide, style reference, or typography use-case page does it instruct you to name an actual typeface, and nowhere does it forbid one either. It's simply absent from the documented method.

What does FLUX.2 not document about text?

Two things, and both matter before you commit a design brief to it.

No character or word ceiling. BFL's only stated limit is qualitative: "Keep text short" because "long strings are harder to render accurately", with no number attached, unlike the hard pixel minimum of 64 it publishes for width and height. There's nothing to cite as a hard cutoff, so treat a headline as the reliable case and a paragraph as a proofreading job.

No claim about non-Latin scripts or foreign-language text rendering. Searching BFL's full documentation for any comparative claim about script or language support in rendered text returns nothing. Its one language-adjacent statement is about prompting, not rendered text: "FLUX can be prompted in multiple languages." and, separately, "FLUX understands a wide range of languages and responds to them with the same level of quality." That is immediately followed by its own hedge that "English prompts tend to produce the most precise results, as the majority of FLUX's training data is in English." That is a claim about how well FLUX parses your instructions, not about how accurately it spells non-Latin characters inside the image, and BFL simply does not make the second claim anywhere we could find.

How does that compare to what Ideogram publishes for the same question?

Directly, and Ideogram is far more explicit about its own limits than BFL is about FLUX.2's.

Ideogram's text-and-typography documentation states outright: "Foreign language support is limited." The same page adds: "Non-Latin scripts often produce unpredictable results." Its warning box goes further still: "For the time being, text that you would like to be written using a non-Latin alphabet or accented Latin characters may have some difficulty being generated correctly, if at all." Ideogram also publishes a length warning in the same shape as BFL's: "The longer the text you want to include, the higher the chance of spelling errors, distortions, or incomplete words."

The pattern worth noticing isn't which vendor is better. It's that Ideogram commits to a specific, checkable limitation in writing, and BFL, for FLUX.2, does not. Absence of a published limit is not evidence the limit doesn't exist; it just means you're testing it yourself instead of reading it off a page.

Where does FLUX.2 actually rank against GPT Image 2?

Not on any chart BFL publishes. The closest available evidence comes from a competitor, and it's worth reading carefully rather than at face value.

Ideogram's own Ideogram 4.0 technical post runs a blind designer-preference arena: "Pairwise ELO from 4,366 graphic-designer votes across nine pipelines. Voters were not told which model produced each image. Higher is better." Reading the chart's own labelled values directly:

FLUX.2 Pro sits sixth of nine, and the open-weight FLUX.2 [dev] sits last. Two honest caveats before you read too much into that. First, this is Ideogram's own arena, about general design-generation preference across "nine pipelines", not a text-rendering-only benchmark. Designers were judging whole images, not grading spelling accuracy in isolation. Second, it's a competitor's chart, and the fact that it still ranks Ideogram's own model second rather than first is the reason we treat it as credible rather than self-serving.

The same technical post carries a second, narrower chart that is closer to the actual question, titled "Parameter efficiency on text rendering (X-Omni EN OCR)", for open-weight models only. Its own caption states "Ideogram 4.0 (9.3B) sits alone in the small-but-strong corner, ahead of every other open-weight release model." Reading the plotted markers against that chart's own axis, FLUX.2 [dev] (32B) sits well up the scale, clustered with Hunyuan Image v3 (80B) and clearly ahead of Qwen-Image, some way behind Ideogram 4.0's own mark but nowhere near the bottom of the field either. Treat that as a rough read of a competitor's own chart, not a number BFL or Ideogram states in prose.

What actually breaks, in our own experience?

BFL publishes no limitations list for FLUX.2 text rendering specifically, so everything past this point is our own observation, not vendor documentation.

Long strings degrade before short ones do. This tracks BFL's own hedge exactly. A two-word sign is a different problem from a paragraph of body copy, and we'd plan to typeset a paragraph afterward rather than generate it directly, on any model.

Dense, small text in a busy scene is the hardest case. A single word on a clean background is the easy case everywhere. A five-word sign crowded into a wide establishing shot, competing with a dozen other visual elements, is where legibility drops first. Describe placement and reduce surrounding visual complexity if a sign keeps coming back wrong.

Untested non-Latin scripts are a real risk, not a documented one. BFL's silence on this specific question, next to its own hedge that English prompting is more precise because training data skews English, is enough reason to test your actual script before a client deliverable, the same way you would with any model that hasn't published a claim either way.

The same word drifts across a multi-shot series unless you lock it down. A logo or product name that renders correctly once can come out spelled slightly differently on the next generation, because pro and max both rewrite a short prompt by default before it reaches the model. The fix isn't a text-specific setting; it's the same determinism lock that matters for any repeated brief on FLUX.2, disable_pup on pro and max or prompt_upsampling: false on flex, so the exact quoted string reaches the model unchanged every time rather than being quietly reworded. Our companion piece on FLUX.2 photorealism formulas covers that lock in more depth for a full shoot list, not just a single word.

Prompt Architects does not generate FLUX.2 images or check the spelling inside them. We generate the prompt: the exact wording, the quotation marks, the placement instruction, and store it once so the next poster, label or UI mockup starts from the version that already worked. Image prompt generation is available from the Pro plan upward; current pricing is on the pricing page, and the free plan runs 5 prompt enhancements a day, forever.