Seedream 4.5 (Replicate)

Seedream 4.5 is ByteDance's top closed image model, and the first thing to understand is that "download it" is not an option. It's API-only - the weights never leave ByteDance's servers. This node is a thin wrapper that lets you call it from inside a normal ComfyUI graph, so your workflow can jump from a local KSampler to a cloud 4K generator and back without leaving the canvas. That's the whole point of the Replicate Select pack: closed models you'll never run locally, sitting next to the local ones you can.

Honest framing first. You're paying per call, your prompt and reference images leave the machine, and the model's safety checker runs before you ever see a result - that's the price of a model that has no open weights. The community's line on API nodes is basically "right tool for a model you can't run, wrong default for one you can." Seedream is firmly in the "can't run" column.

How it works

The pack is built as a catalog (nodes.json) plus a code generator, so every Replicate_* node shares one engine. This node serializes your inputs, uploads any reference images as PNG (small files go inline as data-URIs, big ones via Replicate's Files API), fires off replicate.run("bytedance/seedream-4.5", input=...), downloads the result, and converts it to a normal IMAGE tensor.

That last part matters for how you wire it: the output plugs straight into SaveImage, a face detailer, an upscaler - anything that expects an IMAGE. No video-saver gymnastics here, it behaves like a local generator node.

The inputs that matter

• prompt (required) - up to 4000 characters. BytePlus suggests keeping it under 600 English words; Seedream reads long descriptive prompts well but starts drifting past that.
• image_input - the killer feature. Feed 1–14 images and it does multi-reference generation: character consistency, style reference, scene composition in one call. This is why people reach for Seedream over a local model.
• size - 2K (default) or 4K. Note the trap: 1K is not supported in Seedream 4.5, so don't try to dial it down to save money. If you need specific dimensions, pick custom and set width/height (1024–4096).
• aspect_ratio - defaults to match_input_image, which is what you want for img2img. Otherwise pick from the usual 1:1 → 21:9 list.
• sequential_image_generation - auto makes the model decide whether to output multiple related images (story scenes, character variations), capped by max_images (1–15, and total input+generated can't exceed 15).
• disable_safety_checker - defaults off. Turning it on only relaxes moderation down to blocking illegal content; it doesn't remove the filter entirely.

The single output is named image - an IMAGE tensor.

Install

This node ships with the whole Replicate Select pack - no per-model downloads, because nothing runs locally.

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Or use ComfyUI Manager → Install Custom Nodes → search "Replicate Select". Restart ComfyUI. The only real dependency is the replicate Python package (≥1.0.7), which installs with the pack.

Then point the pack at your Replicate token via an environment variable - there's no key field in the node:

export REPLICATE_API_TOKEN=r8_...

Restart again after setting it; the node reads it from the environment at call time and errors with a clear message if it's missing.

Where people get burned

• The bill. 4K generation is the expensive tier, and unlike a local run there's no "free retry." Iterate at 2K, reserve 4K for finals.
• Refusals. ByteDance's filter is opaque. A prompt that "should be fine" can silently come back blocked; that's the model's policy, not a node bug, and no setting removes it.
• Forgetting the reference images go to a server. If your source images are sensitive, this isn't the node for that job.