Run Seedream 4.0 In ComfyUI! Nano Banana Vs Seedream 4.0, Multi-Image Reference, No Model Download
As AI search says, "AI never sleeps".
Just a few weeks ago, we talked about how to access Nano Banana inside of ComfyUI to access direct API pricing rather than paying through online platforms offering the model at an upcharge.
A few days later, ByteDance comes on to the scene with Seedream 4.0 to compete head to head with Nano Banana, and it does a pretty great job!
While I haven't had the opportunity to test the model truly extensively, it's results and direct comparisons against Nano Banana across the community speak for themselves, and it's native 4k resolution image generation at $0.03 per image is a major bang for your buck.
In this video, we'll show you how to access Seedream 4.0 in ComfyUI (essentially the same process as before with Nano), how to properly send reference images to the model (also the same as Nano), and perform some direct comparisons against Nano Banana to see who is king and how the models differ.
How to Run Nano Banana in Comfy
1.) First, boot up ComfyUI. Go to your manager and click "Update All". Once the update has completed, return to comfy and, in the manager, click "Restart". Wait for the restart to complete and reload the page.
2.) Download the workflow below and drop them into ComfyUI. Click on the Manager again then "Install Missing Custom Nodes" if you have any missing nodes. Once again, restart Comfy.
3.) Once you've returned, you'll need to sign into your Comfy account and load a balance to generate images. If you've already done this in the past, you're all set to generate! If you haven't, click on your profile in the top right and follow the steps to create your account and load your entry balance. At $0.01 per generation, $5 will get you 500 images!
Once your balance has loaded, you're all set!
Known Issues
A call out with this model:
Typically, when sending multiple images to a model, we will use the "Image Stitch" node. Just like Nano Banana, using this node will lead to undesired results. It is recommended to use the "Batch Images" node, though they function the same way. The attached workflows are already configured for this node.
If you want to pass your image generated from the API to further processes in your workflow, make sure to add the "Remove Alpha" node after your image generation Gemini node, then pass the image along.
Resources:
Seedream 4.0 ComfyUI Blog: https://blog.comfy.org/p/seedream-40-now-available-in-comfyui
No model downloads needed this time :)
Comments (4)
Facing the same issue you have at 9min30sec with the letters on the bottle being smeared. How do you overcome something like that ? Could you have an in painting workflow to modify the letters on the bottle and making them neat ? Flux tend to change the proportion of the bottle and the grade as well which is not good. If you had any suggestion this would be very appreciated.
So, this is a MASSIVE problem, and truthfully, in my opinion, one of the most neglected areas of generative AI for whatever reason. Truthfully, I can't say I have any recommendations on resolving this. What I've done in the past is typically use photoshop to actually fix any text and make it accurate. We are working on integrating a model that was largely neglected from a few months back, but is KILLER at text restoration. Our theory is that, with that model as a post-processing step, it should, at the very least, make text legible and reduce the amount of post editing needed read the text. Correcting text (especially labels and such for products) when using generative AI is definitely a case by case situation. If you're just changing the environment of a product for example while keeping the positioning of that product exactly the same, then you could overlay the original input image with the output image, mask the text from the input image, and paste it onto the generated image. That will give you clean text, but require relighting or color correction in photoshop or otherwise. I truly wish I could give you a better pointer here, but this is something that we and I think just about everyone struggles with when generating images with lots of text. The one thing is that there are models that are very good at text. Qwen and Flux 2 as good examples. If you're doing generations with products, it may make sense to lean heavier on models like the two listed that are capable of producing lots of text coherently. Most pre-Q3/Q4 2025 models will probably struggle. Finally, another option is to use Seedream or Nano to correct text, but that can be finicky and is not something I've tried myself yet. Hope this helps a little bit! Rest-assured, something we're looking into, and when there's a solution for it, we'll be on it :) Once again, welcome back!
@silentsnow hello... Coming back on this video (thank you for your great work !!)
Heyy!! Let's go! Welcome back Bert! :)