EigenAI Kontext Generator

The EigenAI Kontext Generator is the image-to-image node in this pack's modular line, and it's the one with the most moving parts. Kontext is the FLUX.1-Kontext-dev family - Black Forest Labs' edit/instruction model, the one the KB's BFL panel notes is more censored than the base Flux models - and here it's served to you as another remote API call. You feed it an image and a prompt, it base64-encodes your image into the JSON payload, POSTs to /generate, and comes back with a tensor.

Unlike the text-to-image siblings, this one does real client-side work before the request goes out: it converts your ComfyUI image tensor to a PIL image, and if auto_downscale_large_images is on (it is by default) it halves any image larger than 512px on the long edge before encoding. That's a server-API design decision leaking into your workflow - the endpoint expects smaller inputs, so the node shrinks yours for you.

Inputs that matter

• image (IMAGE) - your source. Can come from any loader or from a previous generator.
• prompt (PROMPT) - the edit instruction, from the Text node.
• lora_config (LORA_CONFIG) - from the LoRA node. The Kontext default LoRA is an HF repo ID (21j3h123/realEarthKontext), a reminder that these names are server-side.
• inference_steps - default 20, range 1–50. This is the one generator that lets you tune steps, because img2img gives you something to spend them on.
• guidance_scale - default 7.5 here, not the 3.5 the text-to-image nodes default to. That's the SDXL-ish convention this API inherited; treat it as a starting point.
• background_removal / background_removal_strength - flip the boolean and the server runs a background-removal pass; strength 0–1 (default 0.5) controls how aggressive it is. Per the KB's background-removal essay, a dedicated removal pass is usually cleaner than letting the model improvise a background, so this is a genuinely useful knob if your endpoint implements it.
• seed and api_url - seed -1 for random; api_url defaults to http://74.81.65.108:9000 (note the different port from the FLUX/Qwen nodes, since Kontext is a separate service).

Output: image (IMAGE).

Gotchas

• Same silent-failure behavior as every generator here: on error you get a light-red placeholder, not an exception. Pink output = check the server.
• The modular node posts to /generate with the image inline. The unified EigenAIKontextNode hits a different endpoint (/generate-with-image-and-return) - if you switch between them, the server you're talking to has to support both routes.
• auto_downscale_large_images silently changes your composition. If you're editing a 1024px image and the output looks downsampled, that's why.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/jialuw0830/flux_api_comfyui_plugin

Restart, or use ComfyUI Manager and search "Eigen AI FLUX API". Same light dependency set as the rest of the pack - nothing local to download.