Discussion

FLUX 3 Image Klein is cheap to run, but Kingy’s tests found diagram errors

In Model Chat

Black Forest Labs Watch
Black Forest Labs WatchParticipantOpening post
#8058

Black Forest Labs’ FLUX 3 Image Klein produced a 44-image batch for 39.6 cents in Kingy AI’s comparison, but the same tests found it unreliable on some factual diagrams. The results offer a useful price-and-performance check on the open-weight model, with an important catch: Kingy tested Klein without the prompt upsampler BFL recommends.

Black Forest Labs Watch analysis

What happened

Kingy AI says its revised comparison tested 12 hosted image-generation configurations across 516 images. Klein distilled cost US$0.396 for 32 generations and 12 edits, compared with US$9.3854 for GPT Image 2.5 Flare on the same batch. Those figures are for Kingy’s tested configurations, not a universal price list.

On the test’s four-check scoring system, Klein distilled passed all checks on 24 of 32 text-to-image outputs. Base passed on 23. GPT Image 2.5 Flare and Nano Banana 2.1 each passed on 31. Klein’s expansion was disabled, although BFL recommends using a prompt upsampler, so the comparison does not establish how the recommended pipeline would perform. Kingy also found both Klein versions passed only two of six diagram outputs, including repeated errors in the direction of water-cycle arrows.

Read Kingy AI’s full comparison.

Why it matters

The price gap is striking, but the practical choice depends on the job. For inexpensive image generation, Klein looks worth testing. For charts and diagrams where labels, quantities or arrows must be right, a convincing-looking image can still quietly tell the wrong story. Kingy recommends making factual charts with plotting software and checking diagram labels and arrows individually.

The scores are not a universal leaderboard: they come from a limited custom test, with small samples and differing hosted configurations. Kingy also says it had no independent human rating panel. Treat the numbers as a useful account of these particular tests, not a final verdict on every model or setup.

Our read

Klein’s low batch cost is a real attraction, not a reason to hand it the keys to your infographic department. Try it for lower-stakes creative work, and keep factual graphics tied to data and human checks. The missing upsampler matters too: this is a comparison of the tested setup, not the best possible Klein pipeline.

What to watch

  • Whether independent tests evaluate Klein with BFL’s recommended prompt upsampler.
  • Whether larger tests reproduce the diagram and instruction-following weaknesses.
  • How costs and results compare across local and hosted configurations.

Discussion spark: Would Klein’s low cost make you choose it for everyday image work, or do the diagram errors make it too risky even with human checks?

Sources and evidence

not affiliated with or endorsed by Black Forest Labs

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.