Seedream5.0 Pro - Image to Image

The pitch that made people pay attention

Seedream 5.0 Pro landed in mid-2026 and the honest framing of it, from people testing it for real production work rather than benchmarks, was this: the pain isn't a completely unusable output, it's an output that's 90% there with one broken detail. Wrong text on a sign. A product colour that drifted. A face that shifted slightly. The old response was rerolling the entire prompt and hoping.

This node is the direct answer to that: reference image in, plain-language change out, everything else left alone. It's the same instruction-editing category that Qwen-Image-Edit dominates locally, but with ByteDance's model - and since Seedream has never shipped weights, there is no local version of this. ByteDance's split is consistent: generation products stay API-only, components go open. You want Seedream, you call it.

Inputs

• prompt - the instruction, documented at 3 to 5000 characters. The stock default is a genuinely instructive example ("keep the model's pose and the flowing shape of the liquid dress unchanged; change the clothing material…"): name the thing that must stay put.
• image_urls - the reference, on an IMAGE batch socket.
• aspect_ratio - required, defaulting to 1:1.
• quality - required. basic is 1K, high is 2K. Yes, that's lower than the Lite node's 2K/3K/4K ladder, which looks backwards until you remember Pro charges more per pixel for a better model. Set expectations accordingly: if you want four megapixel output from Seedream, Lite is currently the tier that advertises it.
• output_format - png or jpeg.
• nsfw_checker - default True.
• timeout_seconds (1200 default, 30–7200) and callback_url (leave empty).

Outputs: image, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

Restart ComfyUI, then set your KIE key once at Settings → KIE.ai Nodes Next → Connection. It's validated with KIE before it replaces a saved key, and it lives in ComfyUI's user directory instead of your workflow - so sharing a graph doesn't leak billing credentials. KIE_API_KEY in the environment overrides the saved key for headless installs.

Honest read

The "keep the parts that work" framing is the right one for editors, and it's why this node is worth the credits when a client has approved an image except for one thing. But two caveats from the same testing: text rendering is still the weak spot on busy layouts - usable for a simple element, not something to rely on for a dense one - and everyone comparing Seedream 5.0 to Nano Banana Pro or GPT Image has an obvious horse in the race. Plenty of that thread traffic reads as placement.

Practically: bracket your work at basic, make one change per call, and watch the on-node credit receipt. Instruction editors re-emit the whole frame, so identity can slide on a chain - for faces, crop to a region, edit the crop, stitch it back. That's the standard 2026 hybrid and it applies here as much as it does locally.