KIE Seedream 4.5 Text-To-Image

Some days you want a model you could never run locally. Seedream 4.5 is one of those - ByteDance's closed image generator, only reachable through an API and a credit meter. This node is the front door: type a prompt, get a ComfyUI IMAGE. It's the plain text-to-image half of the KIE Seedream pair (its sibling does editing), and it's the one to reach for when you want Seedream's particular brand of prompt-following without firing up its edit pipeline.

Honest framing: everything about this node is a trade of control for convenience. No local sampler, no seed slider, no ControlNet - you get prompt, aspect ratio, and quality, and the model decides the rest. If you live for diffusion knobs, this isn't aimed at you. If you want a strong closed model in the graph with zero setup, it is.

How it works

The KIE job flow, images edition. The node wraps your prompt into a createTask payload with Seedream 4.5's T2I model id, polls recordInfo every 10 seconds until the job succeeds (300-second timeout, auto-retry on transient errors), then downloads the result and converts it to an image tensor. All the heavy inference happens on KIE's servers.

The inputs that matter

• prompt - the only required input, and the whole game. Seedream responds well to concrete visual description; write the scene, not just the subject.
• aspect_ratio - 1:1 default. Full list: 1:1, 4:3, 3:4, 16:9, 9:16, 2:3, 3:2, 21:9. Pick the canvas for the job; you can't crop your way out of a wrong ratio later without losing pixels.
• quality - basic (default) or high. Basic for drafts, high for the keeper. This is where most of the per-run credit cost difference hides.

Output: image, which feeds any IMAGE consumer - a preview, a saver, an upscaler, or one of the pack's I2V nodes to animate the result.

Installing

The whole KIE pack installs at once, and this node rides along. ComfyUI Manager: search "ComfyUI Kie API". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/gateway/ComfyUI-Kie-API
# restart ComfyUI

No models, no heavy dependencies. The one prerequisite is your Kie.ai API key at config/kie_key.txt inside the pack folder (copy from kie_key.example.txt) - no key, and every node in the pack fails before the first request.

Where people get burned

"KIE API key not found" is the first-run wall; make sure the file exists and isn't empty. After that, the failure mode is mostly financial: high quality is meaningfully more expensive per image, and it's easy to burn a balance on drafts. Run Get Remaining Credits before a batch and draft at basic. If a job submits but stalls, watch the console state transitions and check kie.ai/logs for the server-side record - the pack's retry logic handles the transient stuff, so a hard failure usually means something specific to your prompt or balance.