Jimeng Seedream 4

Seedream 4 is where the pack's image nodes start getting interesting. It keeps the basic prompt-to-image flow, then adds two things Seedream 3 never had: reference images you can feed in to steer identity and style, and group generation - one run that produces a set of related images instead of a single picture. If you're doing character consistency or "here's my product, show it in three scenes," this is the node.

It's cloud-only, like the whole pack. Paid Volcano Ark key, ByteDance's servers, no local weights. The README also mentions the 4.5 model is supported here, so this node covers a bit more than its version number suggests.

How it works

Mechanically it's the same request/response loop as Seedream 3, with the reference-image machinery bolted on. You feed images into an autogrowing images input; the node base64-encodes them and sends them along as references, and the API uses them to keep the output consistent with your source. Group generation tells the API to return a set of related images in one call rather than N independent samples - same seed, same prompt family, coherent across the batch. Reference images have size/aspect-ratio constraints the node checks before submitting, so errors here usually mean "resize that input," not "something's broken."

The inputs that matter

• model_version - doubao-seedream-4.5 or doubao-seedream-4.0. The 4.5 is the newer, better one; you're probably picking it.
• prompt - what you want, in the context of your references.
• images - the autogrowing reference input. Single or batch.
• enable_group_generation - off by default. Flip it on to get a set of content-related images in one run.
• max_images - how many images per group, up to 15.
• size / width / height - preset sizes or manual overrides (defaults at 2048×2048, up to 8192).
• seed, generation_count (up to 2048), watermark - the usual suspects.

Outputs

• images - the generated image(s) as a ComfyUI IMAGE.
• response - the raw API response JSON for debugging.

Installing and where people get tripped up

Install is the pack-standard route: ComfyUI Manager or git clone https://github.com/fkxianzhou/ComfyUI-Jimeng-API into custom_nodes, restart. The SDK self-installs; no models to download. The pack quirks - legacy UI only, content moderation - apply.

The realistic confusion points: first, group generation and generation_count interact, so count on reading the console output to see what you actually got. Second, reference-image constraints are real - the node refuses oversized or extreme-aspect images before spending a request. And third, moderation applies to references as much as prompts, so an innocent-looking source image can still bounce. Once those are handled, Seedream 4 is a solid pick for anything where consistency across images matters - and it's a clean step up toward Seedream 5 when you're ready for web search.