Seedream 4 is very impressive. On the image editing side, Seedream 4 scores extraordinarily highly placing 2 points higher even above Google's recently released NanoBanana on my GenAI Image Editor Comparison leaderboard.
An important callout is that it is able to handle significantly larger native resolutions without having to downscale before making edits. Unlike many other image-to-image GenAI models, which cap out around 1 MP, Seedream can handle images up to 4K resolution, and this means less overall image degradation.
It also has pretty decent text-to-image ability as well though gpt-image-1 is still the heavyweight in that category when strictly considering prompt adherence.
BoredPositron · 40m ago
Seeddream works with a fixed patch size in transformer rope and vae. While it has impressive prompt and image understanding the generated images are full of small patches of high frequency artifacts. The real kicker with nano banana is that the model works in pixel space without degrading input images besides the resolution constraints. None of these hype blog goes into the technicalities that really matter for production or "replacing Photoshop" it's just novelty here, novelty there. Now ignore the shortcomings and look at Darth Vader in a pink tutu.
ThrowawayTestr · 3h ago
Anything like nano banana in terms of quality that can run locally?
vunderba · 1h ago
Nano Banana is still at the top of the class but there are two fairly decent image editor style models that you can run locally - Qwen Edit 20b [1] and Kontext Dev [2].
https://genai-showdown.specr.net/image-editing
An important callout is that it is able to handle significantly larger native resolutions without having to downscale before making edits. Unlike many other image-to-image GenAI models, which cap out around 1 MP, Seedream can handle images up to 4K resolution, and this means less overall image degradation.
It also has pretty decent text-to-image ability as well though gpt-image-1 is still the heavyweight in that category when strictly considering prompt adherence.
[1] https://docs.comfy.org/tutorials/image/qwen/qwen-image-edit
[2] https://docs.comfy.org/tutorials/flux/flux-1-kontext-dev