Superside Seedream V4.5 Edit

ByteDance's Seedream family is API-only - there are no weights to download, so the only way it gets inside your ComfyUI graph is through a wrapper like this one. Superside Seedream V4.5 Edit calls fal-ai/bytedance/seedream/v4.5/edit and gives you the model's headline trick: multi-reference editing with up to ten images, a big resolution ceiling, and multi-image output. It's the pack's "throw the whole shoot at it" edit node.

What it does

You give it a prompt plus at least one reference image, and it edits per your instruction while holding the rest of the frame together. The up-to-ten-images surface is what sets it apart: for a product or character that appears across several shots, feeding multiple references gives the model a much richer idea of what to preserve than any single frame can. The README frames V4.5 as the broader, higher-resolution sibling of V5 Pro - up to auto_4K output versus V5's 2K-ish ceiling.

The inputs that matter

• prompt + image_1 + api_key - the required trio. Then image_2…image_10 to fill out the reference set.
• size_mode - preset or custom. Preset picks from image_size; custom uses width/height.
• image_size - from auto_2K (default) up through auto_4K, plus square/portrait/landscape presets.
• width/height - custom mode only, and mind the tooltip: the API requires min 1920 on these. Set them below that and the call can fail, so stick to presets unless you have a specific target.
• num_images and max_images - both 1–8. Two knobs here: how many to request and the max the endpoint will return. Leave both at 1 unless you're deliberately batching variations.
• seed - -1 = random; set a fixed seed to re-roll the same edit.
• enable_safety_checker - on by default.
• sync_mode - off by default. This node uses fal's queued execution path internally (it polls until the job completes, because multi-minute generations drop long-lived connections), so leave sync_mode alone unless you know the endpoint needs it.

Outputs

• images (IMAGE) - the edited result(s) straight back into the graph as a normal image tensor.
• info (STRING) - the result URL.

Installing it

Part of comfyui-superside-nodes - clone, install requirements, restart:

cd ComfyUI/custom_nodes
git clone https://github.com/Superside/comfyui-superside-nodes.git
cd comfyui-superside-nodes
pip install -r requirements.txt

Paste your fal api_key into the node. No model downloads; every call is metered and your references leave the machine for ByteDance's servers.

The honest take

The multi-reference editing is genuinely useful, but it's also where people overpay without realizing it. More reference images usually means a more grounded result, but ten isn't automatically better than three - the model still has to reconcile them, and extra images are extra upload time and cost. Start with one or two, add more only when the edit keeps losing something you care about. And remember the whole family's standing caveat: Seedream is closed and filtered at the source, so anything its safety checker refuses is refused, full stop - no local patch can undo that. Use it for the capability you can't get locally, not as a default for edits your own GPU could handle.