flux1-kontext

Flux Kontext is the 12B model that made instruction-based editing a local commodity - the one the community found "shockingly effective" at removing watermarks and genuinely good at keeping a character consistent across generations. It works by feeding reference images in-context rather than through adapter embeddings, so the model sees your reference and the target simultaneously, which is why it became the backbone of turnaround-sheet character work. This node is that model behind Civitai's orchestration API: images in, sentences in, edited generations out, billed in Buzz, no local weights.

There's a second reason people might specifically want the hosted version, and it's a spicy one. Kontext is equally remembered for its license update - the one that prohibited circumvention of content restrictions and turned every NSFW fine-tune into a violation, with BFL threatening platforms' Flux access until CivitAI and HuggingFace purged the offending models. Run it through this node and the licensing story is Civitai's problem, not yours; you're calling an API, not redistributing weights. For the watermark-removal use case especially, that's a meaningful weight off your shoulders. The honest note: the open editing crown has since moved to Qwen-Image-Edit (Apache 2.0, roughly four times the community traffic), so this node is for people who specifically want Kontext's behavior - not a search for "the best editor."

The inputs that matter

• model - required enum: pro / max / dev, default pro. These are the hosted Kontext variants. pro is the default for a reason; max costs more and is a bit better; dev is the budget door.
• prompt - required. The instruction.
• images - the reference input. This is where Kontext's in-context magic lives: wire one or more images and the model consumes them as references for character/style consistency. Optional, but the node's whole personality is here.
• aspect_ratio - nine choices, default 1:1, when you're generating rather than editing.
• guidance_scale - default 3.5 (range up to 20); the Kontext family is low-guidance by design.
• quantity, seed, output_format, image_metadata - standard.

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

Install and auth

The pack-wide one-time setup. 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.

Where people get burned

Most people trip on the reference wiring: images is optional, so you can happily run pure text-to-image and wonder why it "doesn't feel like Kontext" - the reference input is the point. Feed it the character sheet or style source and the behavior snaps into place. Second, cost: hosted Kontext edits add up fast if you're generating reference-heavy turnaround sheets; check the per-job cost on the node before a long batch. And set expectations on the license question in the other direction too - this node makes your use clean, but your outputs still come from a model whose moderation and content-restriction heritage you can't edit away. That's the trade for "hosted."