Seedream 4.0

vunderba 10 hours ago | parent | next [-]

It's kinda crazy that Google's NanoBanana [1] which was just announced a few short weeks ago has already been upstaged in a lot of ways.

Seedream 4.0 can also handle much higher resolutions. While most image-to-image GenAI models cap out at around 1 megapixel and require downscaling, Seedream supports images up to full 4K so theoretically you should be able to make noticeably cleaner edits.

[1] https://developers.googleblog.com/en/introducing-gemini-2-5-...

	▲	BoredPositron 4 hours ago \| parent [-]
		Nano Banana has a significant architectural advantage. Seedream and Qwen exhibit high frequency artifacts throughout their images due to fixed patch sizes in their transformer, RoPE, and VAE components. (16x16 for Qwen, 4x4 for seedream) These models cannot produce clean or more important high fidelity outputs. Nano Banana operates directly in pixel space that means if you are working with input images your images don't get degradged because of the VAE roundtrip and your t2i are not riddled with high frequency noise. That's the advantage of nano banana. Resolution is not an indicator for quality.

▲

sumedh 12 hours ago | parent | prev | next [-]

Its not available in Australia, US might not be on the supported list as well.

	▲	vunderba 10 hours ago \| parent \| next [-]
		Accessing it directly from Bytedance might be an issue, but it's been available via Fal.AI [1] since it was initially announced - that's where I've been doing all my comparative testing (against Nano Banana, Kontex, etc.) https://fal.ai/models/fal-ai/bytedance/seedream/v4/text-to-i...
	▲	adt 10 hours ago \| parent \| prev [-]
		Plenty of options. https://poe.com/Seedream-4.0

▲

candseven 21 minutes ago | parent | prev [-]

[dead]