EigenAI LoRA Manager

The EigenAI LoRA Manager is where new users of this pack trip over the central fact about it: generation happens on a remote server, and the LoRA "manager" does not look at your disk at all. It doesn't list your models/loras folder, it doesn't apply LoRAs to any checkpoint, it won't even notice if a file doesn't exist. It just builds a little JSON blob and hands it to the generator node, which forwards it over HTTP to an Eigen AI FLUX API server.

So what you're actually managing here is a list of up to three LoRA names that the server knows about, plus a weight for each. The default name field is a dead giveaway:

/data/weights/lora_checkpoints/Studio_Ghibli_Flux.safetensors

That's a path on the server's filesystem - or, in the Kontext variant, a Hugging Face repo ID like 21j3h123/realEarthKontext. Either way it's not something you browse to locally. You type strings, and if the server doesn't have that LoRA, the server decides what happens.

Inputs that matter

• lora1_name / lora1_weight - required. Name is a string (server path or HF ID), weight is a slider from 0.0 to 2.0, default 1.0.
• lora2_name / lora2_weight and lora3_name / lora3_weight - optional. Leave the name blank and that slot is skipped.

One output: lora_config, typed LORA_CONFIG. Wire it into any modular generator's lora_config input. Internally the node validates the names, packs them into a dict with a total_count, caps the list at 3, and - here's a fun trap - if you leave all three name fields empty it silently substitutes the default Studio Ghibli LoRA instead of sending none. You can't opt out of LoRA entirely through this node.

How to use it without burning yourself

Start with weight 1.0, but don't treat it as gospel. The standard advice on LoRA weights applies here even more because you're at the mercy of the server's stacking: too hot and you get that greasy over-baked look. Drop to 0.6–0.8 if a single LoRA is overpowering, and be conservative when stacking two or three - each one multiplies the influence. If a LoRA needs a trigger word, the server-side model presumably expects it in your prompt, because nothing in this pack is going to add it for you.

Installing

Standard for the pack - no models, no heavy deps, just an API client:

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". Needs requests, pillow, numpy, torch, all already in ComfyUI.

One more thing: if you don't need modularity, the unified generator nodes (EigenAIFluxNode, EigenAIQwenNode, EigenAIKontextNode) have three lora1/2/3_name fields built in with the same semantics - same server-side paths, same 0–2 weights, same "none" sentinel for the optional slots. The LoRA Manager only earns its keep in a modular graph where you want one LoRA setup feeding several generators.