DIGIT Seedream Image

ByteDance's Seedream 5.0 is a closed image model - there are no weights to download, so if you want it in a ComfyUI graph you have to call it. DIGIT Seedream Image is one of the cleanest ways to do that: it talks to fal.ai, drops the result back as a normal IMAGE tensor, and slots into the rest of your workflow like any generator node. The KB's ByteDance panel sums up the whole situation: Seedream is an API product, full stop, so this is the only door.

It's a two-in-one node. Connect nothing and it's text-to-image. Connect any of the ten image1–image10 inputs and it switches to edit mode automatically - feed it a reference image and a prompt and it revises or composes. One node, no mode toggle to get wrong.

The inputs that matter

• model - seedream-5.0-pro (default, better quality, max ~2K) or seedream-5.0-lite (cheaper, up to 4K, can return multiple images per call).
• image_size - auto_2K by default. The tooltip is the map: auto_1K is pro-only, auto_3K/auto_4K are lite-only, and custom uses custom_width/custom_height (multiples of 16, pro capped at 2048² pixels, lite up to 4096²).
• num_images and batch_count - the sneaky cost multiplier. num_images is how many separate generations go into one API request; batch_count is how many parallel fal jobs to fire (up to 128). Total output is num_images × batch_count, and you pay per image.
• seed - this one is a trap if you expect reproducibility. The tooltip says it plainly: the Seedream 5 API has no seed input, so this field only controls re-runs. 0 regenerates on every queue; a fixed value keeps ComfyUI from caching the result when nothing else changed. It does not pin the actual output.
• enable_safety_checker - fal-side moderation, on by default.
• output_format - jpeg/png, pro only; lite picks its own format.

Outputs: image (the batch) and text (whatever the model returned alongside, usually empty). Wire image to a preview, a saver, or downstream nodes.

Setup

It's a fal node, so it only needs FAL_KEY, not the whole GCP rig that most of the DIGIT pack requires:

export FAL_KEY=your_fal_key

Install via ComfyUI Manager (search comfyui-digit) or the usual clone-and-pip:

cd ComfyUI/custom_nodes
git clone https://github.com/thedepartmentofexternalservices/comfyui-digit.git
cd comfyui-digit && pip install -r requirements.txt

Where people get burned

Missing FAL_KEY raises before anything happens - export it before starting ComfyUI, not after. Content-policy rejections (422s from fal, which flag things like likenesses of real people) are not retried, because retrying a moderation block is pointless; the node gives you the error and moves on. And if you pick auto_1K while the model is set to lite, it errors out rather than silently downgrading - swap the model or the size. All of this is in the source, so none of it is a surprise once you know it's there. It's a tidy, well-built node for the one thing it does, and for Seedream that's the whole story.