Skip to content
trending

Benchmarks

Seedream 5.0 Flash vs. Recraft: Image and Editing Test

Seedream 5.0 Flash scored 8.0/10 for image generation and 9.2/10 for editing. See how it compares with Recraft V4.1 Flash.

By 6 min read

👁 115 reads

ByteDance’s Seedream 5.0 Flash has been available on fal.ai since September 23, 2026, for both image generation and editing. In a test on October 20, it scored 8.0/10 for generation and 9.2/10 for editing, the highest editing score in this series. The cheaper Recraft V4.1 Flash scored 7.35/10 for text-to-image generation. The test used Polish prompts; translated task descriptions below are translations of the original Polish test material, not new English-language tests.

At a glance

  • Seedream: 10 text-to-image generations and 8 edits, starting from a FLUX 3 Image input.
  • Recraft: the same 10 generation prompts, at a cost of $0.007 per image.
  • In the Polish-language tests, Seedream misspelled the Polish currency abbreviation in a menu and on a poster.
  • Seedream performed better at editing: the sign text and the price on a tag were correct.

What is Seedream Flash?

Flash is the ByteDance image model variant tested here. I’m not applying these results to Seedream 5.0 Pro, covered in the Seedream hub. A shared model-family name does not mean the variants have the same price, speed or text-rendering behavior. This review covers the specific fal.ai deployment and the prompts recorded for the test.

Recraft V4.1 Flash is a lower-cost reference point for generating images from scratch. I did not test its editing capabilities here. The models are compared on text-to-image generation, while the editing results are Seedream’s alone. This distinction avoids attributing a result to Recraft for a task it did not perform.

Image generation cost and speed

In the test’s cost record, Seedream costs $0.027 per image and the same amount per edit. Recraft costs $0.007 per image. These are the rates used in the cost breakdown for this series.

Ten Seedream images cost $0.270 in total, and eight edits cost $0.216. Ten Recraft images cost $0.070. Median times were 16.1 seconds, 17.8 seconds and 6.5 seconds, respectively. The measurement includes submission, queueing and file download. Recraft’s low price came with the shortest generation time, but also a lower average score.

Test access and scope

I generated images from a Polish fal.ai account on October 20, 2026, and the service was working at that time. These results apply to access through that API and to the settings recorded for the test. The measured time and cost should not be assumed to apply to other apps.

Text-to-image test results

Each model received the same prompts, with one attempt per prompt. I assessed image accuracy, text and composition. The portrait and hands were AI-generated. The table gives a concise view of the results; the gallery below shows both strong results and mistakes. The task descriptions are translations of the Polish prompts used in the test; the models were tested in Polish, not English.

Task (translated from Polish) Seedream Recraft
Kraków at night after rain 6.5 5
An apartment-block kitchen 8 7.5
A sign in Polish 10 8.5
A product on a table 9 9
A portrait with natural-looking skin 9 7
Hands making pierogi 9 8
A Polish milk bar menu 5.5 8.5
Morskie Oko at dawn 6 6
A concert poster 8 9
Chinese and English text 9 5
Flawless result
Flawless. 10/10.
Unnecessary comma after “Fresh”
Unnecessary comma after “Fresh.” 8.5/10.
The Polish currency abbreviation is misspelled in every price
The Polish currency abbreviation is misspelled in every price. 5.5/10.
Correct text, but the last line is misaligned
Correct text, but the last line is misaligned. 8.5/10.
The Polish currency abbreviation is misspelled
The Polish currency abbreviation is misspelled. 8/10.
Flawless small text
Flawless small text. 9/10.

Seedream rendered the bakery sign flawlessly, but misspelled the Polish currency abbreviation in every menu price. The poster had the same kind of error. Recraft got the menu text right, though the last line was misaligned, and its poster had no spelling error despite the small text. A lower average does not mean a worse result on every task.

Two matching towers, not St. Mary’s Basilica
Two matching towers, not St. Mary’s Basilica. 6.5/10.
An ordinary Gothic street
An ordinary Gothic street. 5/10.
Correct label, but the framing is cropped
Correct label, but the framing is cropped. 9/10.
Correct label
Correct label. 9/10.
Both lines are correct
Both lines are correct. 9/10.
Chinese characters are incomplete
Chinese characters are incomplete. 5/10.

For the Kraków scene, Seedream showed two matching towers instead of St. Mary’s Basilica, while Recraft produced an ordinary Gothic street. Neither image made the location recognizable. The product with a label did better, with both models scoring 9/10. A neon sign with Chinese and English text revealed another difference: Recraft’s Chinese characters were incomplete.

Seedream editing: the highest average in this series

All eight editing tasks started from a FLUX 3 Image source image, not from an earlier Seedream output. The average score was 9.19/10, rounded to 9.2, the highest in the editing test. Changing a sign, moving a product, removing a kettle and changing a night scene to a winter day each scored 9.5/10.

Flawless text, slightly shifted framing
Flawless text, slightly shifted framing. 9.5/10.
Sharp label, best advertisement
Sharp label, best advertisement. 9.5/10.
Same face, close framing
Same face, close framing. 9/10.
Clean result, espresso machine
Clean result, espresso machine. 9.5/10.
Pouring from a kettle by the stove, face preserved
Pouring from a kettle by the stove, face preserved. 9/10.
1
First line is readable; second is distorted. 8.5/10.
Winter day, everything in place
Winter day, everything in place. 9.5/10.
Correct Polish price, but the tag covers the top of the label
The Polish price is correct, but the tag covers the top of the label. 9/10.

The price result shows the difference most clearly: generation from scratch misspelled the Polish currency abbreviation, while editing added the correct Polish price. The result was not perfect, though, because the tag covered the top of the label. With three reference images, the second line of the label was distorted. A high average score does not remove the need to check text and detail placement.

When the scene changed, the generated face was preserved, and the jar in a meadow had a sharp label. Removing the kettle produced a clean result with an espresso machine. For a full comparison of the same editing task with another model, see the AI photo editing test.

Strengths

  • Seedream: preserved the label when moving a product and cleanly removed an object.
  • Seedream: rendered Polish prices correctly in the editing test.
  • Recraft: low cost and fast generation, plus a strong poster result.

Weaknesses

  • Seedream: misspelled the Polish currency abbreviation in prices generated from scratch.
  • Both models: misinterpreted a recognizable location.
  • Recraft: incomplete Chinese text and a different hair color in the portrait.

Who should use each model?

In this series, Seedream Flash is a strong candidate for editing existing images, especially product shots and interiors. Recraft can be useful for inexpensive composition sketches, provided you check the output. For a Polish price list, choose based on whether the finished graphic reads correctly, not just on the model’s average score.

Frequently asked questions

Does Seedream always get the Polish currency abbreviation wrong?

No. The error appeared in the generated menu and poster, while the price in the tested edit was correct.

Was Recraft better at the poster?

Yes. It scored 9/10, compared with 8/10 for Seedream. That was one task, despite Recraft’s lower average.

Do these results apply to Seedream Pro?

No. The test covers Flash on fal.ai. The Pro variant needs to be evaluated separately.

For more context, see the roundup of new fal.ai models from October 2026. For a broader overview of available tools, see the AI image generator comparison.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources

Kinetyra

// read next

More on this from QUANTSLANT

Piotr Olszewski

Editor and publisher · Quantslant

I edit and publish Quantslant. AI tools help with research, drafts and translation; choosing topics, running the tests, checking facts and the decision to publish are mine, and so is the responsibility for what appears here.

read nextguidesChatGPT Course: 8 Practical Exercises for Real TasksBuild safer, more reliable ChatGPT habits with eight practical exercises for writing, research, spreadsheets and reusable prompts.

Most read

  1. 1Best AI Video Generator Apps for Your Phone (2026)
  2. 2Claude Opus 5.5 Motion Graphics: 60s Showreel from One Prompt
  3. 3Wan 2.2 on RTX 4090: timings, VRAM and four test clips
// site map

This page in other languages5

Quantslant in numbers 1,427posts in archive26news from the last 7 days18video models in the observatory
🎬 AI Video Benchmarks: Higgsfield DoP vs OpenAI Sora • Higgsfield vs Runway Gen-3 • AI Video Credit Calculator
× ‹ enlarge ›