flux2 / dev / editImage

Flux 2's whole party trick is that it's one checkpoint for two jobs - generation and instruction-based editing - and this node is the editing half of the dev variant. Take the flux2 / dev / createImage node and add exactly one input: images. Wire a picture in, describe the change, and the hosted dev checkpoint returns the edited version as an IMAGE tensor, billed in Buzz. Everything you read in the createImage article still applies - the sampler knobs, the LoRA socket, the "you could run this locally if you had 18–24GB of VRAM" framing - it's just that here the input is an image, not a blank canvas.

The editing story is where the dev checkpoint earns its reputation. Flux 2's in-context editing is strong: it holds character and style across the change, and because it's the same model as generation, edits stay consistent with what the model itself would draw. Combined with the loras socket, this is the node for "edit this existing image with a Flux 2 LoRA I picked from Civitai's browser" - a workflow that's genuinely awkward to build locally and trivially easy here.

The inputs that matter

Required: prompt and images (IMAGE socket - wire a LoadImage or any image output).

• prompt - the instruction. Sentence-level changes work well ("make it autumn", "change the car to red").
• guidance_scale (2.5) / num_inference_steps (28) - same dev defaults as createImage; edits want the same low-guidance, ~28-step recipe.
• loras - CIVITAI_LORAS socket for the Civitai LoRA Selector; style/character LoRAs apply on top of the edit.
• width / height - 1024×1024 default. Match these to your input's aspect ratio or the source gets resized and squashed - the classic silent quality killer on every image-input node in this pack.
• enable_prompt_expansion - off by default; for edits, consider explicitly keeping it off so the prompt doesn't get rewritten mid-task.
• quantity, seed, output_format, image_metadata - standard.

Outputs: images (IMAGE), errors, workflow_id, raw_json.

Install and auth

The one-time pack install. ComfyUI Manager → Civitai Comfy Nodes, or:

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

Authenticate (Civitai account with Buzz): export CIVITAI_API_TOKEN=..., a Civitai Auth node, or the sidebar sign-in.

The honest call

Where this node sits among the pack's editors: Qwen-Image-Edit is the open default, HiDream-O1 is the recursion specialist, and Flux 2 dev edit is the "strong all-rounder with sampler control and a LoRA ecosystem" option - arguably the most flexible editing node in the pack precisely because of that LoRA socket. Its limits are the dev checkpoint's limits: a step below pro/max on the very hardest edits, and the usual instruction-editing drift where unprompted pixels wander (the model re-emits the whole frame). And the standing question: if you have the VRAM and the privacy requirements, the local dev install is free per-edit - this node is for the hardware-constrained and the "I want Civitai's LoRA catalog wired in without managing files" crowd. Both are legitimate, and this is the cleanest door to Flux 2 dev editing the pack offers.