Generate 4o Image

4o image generation is the thing that went viral inside ChatGPT in March 2025 - the Ghibli-style portraits, the whole wave - and this node calls it through KIE. Two reasons you'd want it here rather than in a browser tab: reference-image editing with up to five inputs, and the ability to put a closed model in the middle of a local pipeline instead of at the end of one.

What you're actually buying

OpenAI's 4o image is a strong generalist with a well-earned reputation for unpredictable refusal. The KB's closed-source-models page records the community's verdict in one quote - "Chat gpt 4O sucks and everything trips its baby mode content filters" - alongside the praise for output quality. Both halves are real. It renders text, follows compound instructions with people and objects in frame, and edits a supplied image coherently; it will also occasionally decide your prompt is a problem and hand you a lecture instead of a picture.

That's the trade for using a closed model at all. The same doc's framing applies: an API node is the right tool for a model you cannot run, and a bad default for one you can.

Inputs

Only three fields are on the card.

prompt - your description. The tooltip is specific: required if neither filesUrl nor fileUrl is supplied, and at least one of prompt or filesUrl must be present. So this is text-to-image, or the instruction half of an edit.

size - required, and limited: 1:1, 3:2, 2:3. KIE documents no other values on this page, so if you were hoping for 16:9 out of the node, you're cropping afterwards.

filesUrl - up to 5 publicly reachable image URLs to use as reference or source material. Note that this one stayed a JSON array string rather than becoming a native image socket, matching KIE's documentation for the endpoint. If you don't have hosting, KIE's docs point at their file-upload quickstart to get URLs first.

fileUrl - a single URL, explicitly marked deprecated on the endpoint's own page. Prefer filesUrl in anything new.

Plus timeout_seconds (1200 default, 30–7200) and callback_url (leave empty for normal ComfyUI use).

Outputs

image is a real IMAGE - batch-ready, goes straight into a Save Image, a preview, or an img2img stage. url is the hosted result; all_urls_json keeps every URL the response reported, which matters because 4o image jobs can return more than one artifact. task_id is what you feed the companion Get 4o Image Details node if you want to poll the job separately. raw_json is the whole response - where you look when the model refused, because a refusal bodies its way back rather than always erroring. credits_consumed and credits_left are floats, plus the on-node receipt.

Install

ComfyUI Manager (search the pack title), or:

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

Restart. The dependency list is requests, Pillow, numpy, PyYAML - no weights, no local inference, no VRAM. The cost is entirely per-call on KIE's side.

Key in Settings → KIE.ai Nodes Next → Connection - Paste / replace KIE.ai API key. The pack validates it against KIE before replacing a working key, then shows a masked fingerprint and CONNECTED. Servers and headless setups can set KIE_API_KEY, which takes priority over the saved key.

Gotchas

Refusals arrive as text, not as an exception. If a run finishes and your image looks wrong or missing, read raw_json before assuming a node bug.

Reference images must be reachable. filesUrl takes URLs. Local files need the KIE file-upload route first - this is the one node in the pack where a local image doesn't just get uploaded for you.

Sizes are three and only three. Plan your composition around 1:1, 3:2 or 2:3, and crop downstream.

Card drift. The pack builds its node library from KIE's API docs at startup and re-checks about daily, so fields follow the provider's documentation rather than the plugin's release notes.