ByteDance Seedream 5

The odd one out in a video pack: an image generator. And it's a good one to have around, because it pairs perfectly with the Seedance video nodes in the same repo - generate a still with Seedream, feed it into an Image-to-Video node, and you've got a complete shot. Same BytePlus ARK API, same ByteDance closed-model story: Seedream is ByteDance's flagship image line, there are no weights to download, and every generation is a metered cloud call.

What makes this node mechanically different from its pack-mates is that it's synchronous. The Seedance nodes submit a task and poll for minutes; Seedream just POSTs to /images/generations and gets URLs back in the response. It downloads them into an IMAGE tensor batch, saves copies to ComfyUI/output/ as seedream_<timestamp>.png, and returns. No poll_interval, no max_wait - it's a one-shot HTTP request, which makes it feel instant by comparison.

What you actually set

• model - seedream-5-0-260128 (full) or seedream-5-0-lite-260128 (the cheaper lite variant).
• size_preset - the big one. Seventeen presets across 2K and 3K with ratios from 1:1 to 21:9, plus custom. Pick a preset and forget the raw dimensions.
• width / height - 512 to 4704, only read when size_preset is custom.
• sequential_image_generation - the genuinely unusual feature. Turn this on and the model can return a group of up to max_images related images in one request, like a quick character or style sheet.
• seed - 0 lets the server pick; set one for reproducibility.
• output_format - png or jpeg. watermark if you're feeling cooperative.
• fail_on_partial - when grouped generation returns fewer images than requested, error out instead of silently returning the short batch.

The optional image input is multi-reference: a batched IMAGE is sent as several reference images, which makes this node a legit image-editing tool - reference a character or a scene and generate variations around it. Or skip tensors entirely with image_url (multiple URLs, one per line).

Three outputs: images (an IMAGE batch - wire straight into the video nodes' first_frame), plus image_urls and image_paths as newline-joined STRINGs.

Install

Same as the rest of the pack - ComfyUI Manager, search comfyui-seedance-nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/theclueless-ai/comfyui-seedance-nodes
pip install -r ComfyUI/custom_nodes/comfyui-seedance-nodes/requirements.txt

Just requests, Pillow, numpy, opencv-python; no model files. Key via ARK_API_KEY env var or the api_key field, restart, and it appears under Seedance / Image Generation.

Where people get burned

• The 15-image budget. Reference images plus generated images must be ≤ 15. Feed in 14 references and ask for 5 images and the node clamps max_images down (or errors if there's no room at all). The console tells you what it did - check it.
• Custom sizes are easy to forget. The width/height inputs sit there looking active, but they're ignored unless size_preset is custom.
• The bill and the filter. This is a paid, metered call, and 3K multi-image runs add up fast. Your prompt and reference images leave the machine and pass through ByteDance's moderation - the model refuses at the source, and no node setting changes that.