flux2 / editImage

You know the Flux 2 story: the 32B transformer is the most capable open image model you can download, and also the one that laughs at your GPU - 18-24GB of VRAM even quantized. If you've got a mid-range card, running Flux 2 Dev locally is a wall. This node is the workaround: it's the same engine, hosted on Civitai's fleet, and your job is just to describe the edit.

What it is. CivitaiImageGenFlux2EditImage is one of the ~160 nodes in the official civitai/civitai-comfy-nodes pack, which wraps Civitai's Orchestration API as ComfyUI nodes. The name tells you the shape: engine flux2, operation editImage. You feed it an image, tell it what to change, and it returns the result as a normal IMAGE tensor - no local checkpoint, no VAE, no model download at all. Your part of the bill is Buzz, Civitai's credit, and the generation happens in their cloud.

How it works. Wire a reference image into the images input (any IMAGE - from a Load Image node, another generator, anything in your graph), describe the change in prompt, and hit run. The node packages your prompt plus the image, POSTs a workflow to orchestration.civitai.com, then long-polls until it's done - so the graph blocks for however long the cloud job takes, which at launch people reported being slow. When it succeeds you get four outputs: images (the edited result, ready to wire into a Save Image or Preview), errors (a JSON string of any step errors), workflow_id (useful for tracking and cost lookup), and raw_json (the full workflow response).

Inputs that matter. The model dropdown is the one thing this node has that its siblings don't: pro or max, the two hosted Flux 2 sizes. Default pro is fine for most edits; max costs more Buzz and buys the top tier. width/height (512-2048) and quantity (1-4) behave like you'd expect. enable_prompt_expansion (default off) lets Civitai elaborate your prompt before generation - handy when your edit instructions are terse. Everything else - output_format, image_metadata (external metadata stored with the image), seed - you can leave alone on a first run. If you want to force a negative-prompt style no, that's not here; the edit variant keeps the input list tight.

Install. ComfyUI Manager → Custom Nodes Manager → search Civitai Comfy Nodes (publisher civitai) → Install, then restart. Or from the terminal:

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

The only dependency is requests. No models to download - that's the whole point.

Auth and gotchas. The node resolves credentials in order: a wired Civitai Auth node, then CIVITAI_API_TOKEN, then a stored key or OAuth login from the Civitai sidebar. If you're headless, set the env var. The realistic failure modes, from the pack's own error handling: 402 means not enough Buzz, 401 means your token's bad or expired, 429 means you're rate-limited. Generations time out after 30 minutes. And two things the launch thread flagged that are worth internalizing: it's early preview, so node behavior can change without notice, and it enforces the same content restrictions as the web generator - Flux 2 is censored at the source, so don't expect the edit node to do things the website won't.

Your data - the image and prompt - leaves the machine, which is the tradeoff of every cloud node. Use it for the edits your GPU can't run, and keep your local Flux 1 / Z-Image for everything it can.