FL Fal Seedream Edit

ByteDance's Seedream v4, the image-editing side of a lab that runs a deliberately two-track strategy: flagship generation products - Seedream, Seedance - stay closed and API-only forever, while a steady stream of open components (SDXL-Lightning, PuLID, Depth Anything, and more) ships free underneath. Seedream itself is squarely in the "you will never self-host this" column, so Fal.ai's hosted endpoint is the only door in, and this node is that door.

The mechanism

Composites or edits up to ten reference images against one shared prompt (image_1 through image_10, all optional), auto-scaling inputs to meet Seedream's minimum dimension requirements while preserving aspect ratio if auto_scale_to_minimum is left on - it defaults to true, and you generally want it left there, since turning it off means undersized inputs can just fail the API call instead of being quietly upscaled first. All ten reference slots are optional, so unlike FL_Fal_GPTImage2_Edit in this same pack, you can leave every image input empty and use this purely as a text-to-image node against the prompt alone.

Inputs and outputs that matter

prompt - the default sample is literally "Dress the model in the clothes and shoes," which tells you the kind of compositing task this is tuned for: garment or product try-on style edits. image_size gives six fixed presets (square_hd through landscape_16_9), num_images (1-4) sets outputs per call, and seed, sync_mode, max_retries round out the required fields. On the optional side, use_custom_resolution unlocks custom_width/custom_height (both 256-4096px) if none of the six presets fit, and retry_indefinitely offers the same "keep hammering on transient failure" behavior as this pack's other Fal nodes.

Outputs: images, image_urls, seed (returned as a string - useful for logging exactly what Seedream used, since with num_images above 1 the outputs in a batch may not all share one), and status_msg.

Installing it

ComfyUI Manager: search Fill-Nodes, install, restart. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/filliptm/ComfyUI_Fill-Nodes, then restart.

Where people get burned

image_size and use_custom_resolution are effectively mutually exclusive: set custom_width/custom_height but forget to toggle use_custom_resolution on, and the node falls back to whatever preset is selected in image_size - your custom dimensions get silently ignored.

And because this is a closed API model with no local equivalent to fall back on for comparison, if an edit doesn't land right, the fix space is entirely in prompt and reference-image choice. There's no CFG scale or sampler to blame it on - Seedream's own instruction-following is the whole mechanism. If you're trying to decide between this and the Fal Kontext or Gemini Image Edit nodes elsewhere in this pack for a compositing task, Seedream's ten-reference ceiling is the highest of the three, which matters if you're combining several distinct product or garment shots into one scene rather than editing a single subject.