Seedream 4.5 vs Z-Image Turbo: image quality, speed and text rendering

Use Z-Image Turbo to explore ideas quickly, then render the image you keep with Seedream 4.5. Z-Image Turbo is Alibaba's fast open-weight model; Seedream 4.5 is ByteDance's quality-first model. Both sit behind the same FastMetal endpoint, so switching between them means changing one model value.

At a glance

Figures are the developers' and benchmark sites' published numbers as of April 2026.

Seedream 4.5

ByteDance released Seedream 4.5 in December 2025. It generates at up to 2048×2048 and is built for cinematic output: lighting, composition and spatial layout are its strengths. It blends styles (dreamy, anime, photoreal) more freely than Z-Image Turbo does.

Text rendering is the other reason to pick it. Beyond English and Chinese it handles Japanese kana, Korean, Spanish and Arabic, which makes it the safer choice for posters, social cards and anything else where words have to be legible inside the image. Keep embedded text short: complex Chinese characters (kanji) are often simplified or garbled, so one short phrase per line works best.

It accepts up to 10 reference images, which keeps a character or a product consistent across a series, and it supports inpainting, outpainting and object replacement.

On the LM Arena text-to-image leaderboard it scored 1,147, inside the top 10, from more than 20,000 blind pairwise votes (April 2026).

The trade-off: it is slower than Z-Image Turbo and costs more per image.

Z-Image Turbo

Alibaba's Tongyi Lab released Z-Image Turbo on November 26, 2025, as open weights under Apache 2.0. It is a 6B-parameter model on a scalable single-stream diffusion transformer (S3-DiT), distilled with Decoupled-DMD so it needs only 8 sampling steps where typical diffusion models take 20 to 50.

That shows up as speed: about 2.3 seconds for a 1024×1024 image on an RTX 4090, and reports of under a second on data-center GPUs such as the H800. It ranked first among open-weight image models on Artificial Analysis, with more than twice the speed of the runner-up.

For its size, its photoreal portraits hold up well against much larger closed models. It renders Chinese and English text natively and supports image editing as well as text-to-image.

The limits: standard output is 1024×1024 (and that is the size the FastMetal route generates), it has less stylistic range than Seedream 4.5, and it does not render Japanese kanji reliably. For text in other scripts, use Seedream 4.5.

A two-step workflow

1. Draft with Z-Image Turbo. It is fast and cheap per image, so you can try many compositions and prompt variations for the same budget.
2. Once a direction works, send the refined prompt to Seedream 4.5 for the final render, especially if it needs text or a higher resolution.

Both are billed per image, deducted from your prepaid balance. Z-Image Turbo is the cheaper of the two. Current prices are on the Seedream 4.5 and Z-Image Turbo model pages and the pricing page.

Writing prompts

Both models do their best work from English prompts. A good prompt covers:

• Subject: what to draw (a woman in a traditional kimono)
• Style: ukiyo-e style, cinematic lighting, photorealistic
• Mood: season, time of day, palette (cherry blossoms, golden hour, warm tones)
• Quality terms: high detail, 4K, sharp focus at the end tend to help

For text inside the image, put the exact words in quotes in the prompt and keep them short.

FAQ

Q. Can I use both models with one API key?

Yes. One FastMetal key works for both; you only change the model parameter. Create a key from the dashboard.

Q. Can I use them commercially?

Z-Image Turbo's weights are Apache 2.0. Seedream 4.5 is a proprietary model. For use through FastMetal, also read the FastMetal terms.

Q. Does either model train on my prompts?

Our catalog records Seedream 4.5's route as not training on prompts. For Z-Image Turbo's route the provider has not stated a policy, so we list it as unknown. Each model page shows the current data policy.

Try it on FastMetal

# Seedream 4.5: quality first
curl https://api.fastmetal.ai/v1/images/generations \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bytedance-seedream-4.5",
    "prompt": "Mount Fuji with cherry blossoms, ukiyo-e style"
  }'

# Z-Image Turbo: speed first
curl https://api.fastmetal.ai/v1/images/generations \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-image-turbo",
    "prompt": "Mount Fuji with cherry blossoms, ukiyo-e style"
  }'

Every image and text model is in the model catalog.