Jimeng Seedream 5

Seedream 5 is the current top of the pack's image lineup, and it earns the spot with two things no earlier Seedream node has: internet search and a Pro model that thinks about your prompt before drawing. You'd reach for it when you want an image of something current - a product that launched last week, a real landmark as it exists now - or when you want the model to quietly improve your prompt instead of taking it literally.

Same ground rules as the rest of the pack: paid Volcano Ark API key, ByteDance servers, no local model. This is ByteDance's flagship image line, API-only by design, with the ComfyUI ergonomics bolted on.

How it works

The node is a dynamic combo, meaning the inputs reconfigure depending on which model_version you pick - doubao-seedream-5.0-pro or doubao-seedream-5.0-lite.

• Pro adds a thinking toggle (on by default) that has the model rewrite and optimize your prompt before generating. It's not optional cosmetics: the README notes thinking is required when Pro is used with reference images, so leave it on unless you have a reason.
• Lite is the workhorse variant. It gets the enable_web_search toggle, which lets the model fetch real-time information from the internet to ground the image - search for the actual product, the actual building, the current price tag in frame. It also gets enable_group_generation and max_images for producing a set of related images in one run.

Both variants accept reference images through the autogrowing images input and output at up to 2048×2048 by default (sizes up to 8192). There are hard pixel-count ceilings per variant the node checks, so giant canvases get refused rather than silently mangled.

The inputs that matter

• model_version - the Pro/Lite split, and it changes the whole input panel.
• prompt - still the main event, but with Pro's thinking toggle it's more like a brief.
• thinking (Pro) - prompt optimization, required with references.
• enable_web_search (Lite) - the internet search feature.
• enable_group_generation / max_images (Lite) - grouped, related-image output.
• images - reference inputs for both.
• seed / generation_count / watermark - the usual.

Outputs

• images - the generated image(s).
• response - the raw API response JSON.

Installing and what to expect

Standard install: ComfyUI Manager (search "ComfyUI Jimeng API") or git clone https://github.com/fkxianzhou/ComfyUI-Jimeng-API into custom_nodes, restart, SDK self-installs. The usual pack caveats: legacy UI only, and moderation on both prompts and reference images.

The realistic friction points: web search costs more and can be slow, so it's a toggle, not a default; group generation with reference images is where people accidentally blow through a batch, so check your limits; and the pixel-count caps are real - a 8192-wide canvas on the wrong variant gets rejected. None of that is friction enough to skip it. For anything current, consistent, or demanding a smarter prompt, this is the image node to reach for in the pack.