flux2 / klein / editImage

Take the Klein create node, bolt an image input onto it, and you get this: the most control-heavy edit node in the Civitai image lineup. Flux 2 Klein earned its reputation as the editing front of the open-weights ecosystem - the distilled 4B/9B models that people actually fine-tune and use for img2img when the full 32B Dev is out of reach. CivitaiImageGenFlux2KleinEditImage serves that same model as a hosted edit operation, so you can do image-guided edits without a local checkpoint or a VRAM budget.

How it works. Wire a source image into images, describe the edit in prompt, and the node packages both into a workflow submitted to orchestration.civitai.com. While it runs, you get none of the local model loading - the first few seconds are just submission and queue - and then an edited IMAGE comes back on the images output, ready for Preview, Save, or downstream processing. Standard pack outputs: images, errors, workflow_id, and raw_json.

Where it's richer than the other edit nodes. Klein's edit variant carries over the create node's tuning surface, which is rare for a cloud edit endpoint. You get cfg_scale (default 5.0), steps (default 20), a full sample_method list (euler, dpm++2m, heun, lcm, tcd…), a schedule dropdown (simple, discrete, karras…), and a real negative_prompt - the field most of the cloud nodes simply don't have. Then the two headliners: model_version (4b vs 9b, plus the base variants) and loras, a CIVITAI_LORAS socket.

The LoRA gotcha, again. You can't type a LoRA name here. Wire a Civitai LoRA Selector's loras output into the loras socket, one node holding as many LoRAs as you want with per-LoRA strength and enable toggles. The selector's keywords field is auto-filled as a reminder only - the LoRA applies by its AIR reference and strength, so if you want its trained concept to show up in the edit, paste those trigger words into this node's prompt. A common beginner trap is wiring the selector, seeing "connected", and wondering why nothing changed.

Inputs that matter. prompt (required) should describe the edit precisely, and for a tight edit consider raising cfg_scale above the 5.0 default so the model stays anchored to your source. width/height (512-2048), seed, and quantity (1-4) are standard. enable_prompt_expansion (off by default) has Civitai rewrite your prompt before generating - usually fine, occasionally rewrites an edit instruction you wanted literal.

Install and auth. It ships with the pack - ComfyUI Manager → search Civitai Comfy Nodes → Install, or git clone https://github.com/civitai/civitai-comfy-nodes.git into custom_nodes and pip install -r requirements.txt. Credentials resolve from a Civitai Auth node, CIVITAI_API_TOKEN, or stored sidebar login. Keep the error codes in your head: 402 insufficient Buzz, 401 bad token, 429 rate limit, and jobs are capped at a 30-minute timeout.

The honest take. Cloud Klein edits are the slow way to do something a 13GB local card can handle - the launch reports of slow generations weren't wrong. But if your GPU can't run Klein at all, or you want a LoRA stack applied to an edit without managing model files, this is the most capable edit node Civitai publishes. Just budget for the round trip and don't forget the trigger words.