mageflow / 4b-edit / editImage

Everything interesting about MageFlow is in the edit variant. Microsoft built the 4B model to both generate and edit from the same weights, and the community's take was unanimous: for editing it outclasses Krea 2, which was never trained for it. This node is the 4B-edit model running on Civitai's cloud - feed it an image and a sentence, get back the edited IMAGE, no local weights and no VRAM.

It's the right tool when you want clean instruction editing ("move the subject right", "turn this into a night scene") without standing up a Qwen-Edit or Kontext chain locally. The 4B size keeps it cheap to host, which is why this specific editor is worth trying even if you've got a big GPU sitting idle - the hosted job often beats the effort of running a 20B editor locally.

How it works

The civitai-comfy-nodes pattern: upload your source image and prompt to the Civitai Orchestration API, run MageFlow 4B-edit on Civitai's fleet, long-poll, and return the result as a native IMAGE tensor. This is the non-turbo edit variant - default 30 steps at CFG 5.0. That's the quality-oriented setting; the turbo sibling does 4 steps if you want speed.

The inputs that matter

• prompt and images are both required - an editor without a source is just a generator.
• steps / cfg_scale - 30 / 5.0 defaults. If edits come back rough, this is where you'd nudge quality up; if they're slow, the turbo variant is the speed lever.
• diffusion_model (CIVITAI_AIR) and loras (CIVITAI_LORAS) - optional typed sockets. Wire Civitai Model Selector / LoRA Selector nodes to swap in a MageFlow finetune or stack LoRAs on the cloud job.
• sampler / scheduler / seed / width / height / quantity (max 4) - standard.

Outputs

images wires into Preview/Save Image. errors, workflow_id, and raw_json are diagnostics; raw_json has the full workflow result when you need to audit cost or behavior.

Installing and authenticating

Pack-wide install - no separate MageFlow package:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

or ComfyUI Manager → "Civitai Comfy Nodes" → Install → restart. Dependencies are just requests and python-socketio[client].

Auth: Civitai Auth node → CIVITAI_API_TOKEN → stored key/OAuth in ~/.civitai/ → browser login. Headless:

export CIVITAI_API_TOKEN=your_token

Where people get burned

• 30 steps is the quality default, and it's billed per job. The non-turbo edit is the expensive end of the MageFlow family. If you're iterating on prompts, use the turbo edit for drafts and this one for finals.
• Edits re-emit the whole frame. Like all instruction editors, unedited regions come back close but not identical, and drift compounds across a chain of edits. Keep your source clean and consider re-cropping.
• Your image leaves the machine. Standard API-node deal - the reference and prompt go to Civitai's servers, subject to hosted moderation.
• Don't confuse the create and edit nodes - if you've wired images into the createImage variant expecting it to matter, you're on the wrong node.