Eigen AI Qwen Generator

EigenAIQwenNode is this pack's unified Qwen-compatible text-to-image node - the "one node does everything" counterpart to the modular EigenAIQwenGeneratorNode. Prompt on the panel, three LoRA name slots, an upscale toggle, and a single image output. Same remote-API architecture as everything else here: build a JSON payload, POST to /generate on an Eigen AI FLUX API server, download the image, return a tensor. Your hardware contributes nothing but the network connection.

Two small things make this node distinct from the FLUX sibling on the same panel. First, its default api_url is http://74.81.65.108:8010 - port 8010, matching what the README documents for Qwen, while the modular Qwen generator ships with the shared 8000 default. In practice it doesn't matter, because you're replacing whichever hardcoded IP you get with your own endpoint anyway; just don't assume the two Qwen nodes point at the same service out of the box. Second, the upscale handling: flip upscale on and the payload includes upscale plus upscale_factor (2 or 4), and the server returns a larger image. That's a server-side operation - you don't get to see the 512 result before it's blown up.

Inputs that matter

• prompt (STRING) - type it right on the node.
• width / height - 256–1024 in steps of 64, default 512.
• guidance_scale - default 3.5, ±10 range.
• seed - -1 for random.
• upscale (boolean) + upscale_factor (2 or 4) - server-side, as above.
• lora1_name / lora1_weight - required. Default lora1_name is the same server-side Studio Ghibli path as the FLUX node. Optional lora2_name / lora3_name default to "none" (the sentinel that gets skipped).
• api_url - default http://74.81.65.108:8010. Set it to a live endpoint.

Output: image (IMAGE).

Where it bites

Everything you'd expect from this pack's family of nodes: LoRA names are server-side strings with no validation on your end, and any failure produces a light-red placeholder instead of an error. The sneaky one is the default lora1_name - it's a required field with a server path pre-filled, so if you build a workflow and forget it, you're silently requesting a LoRA that may not exist on your server. Change it to something your API actually knows, or the server decides what that means. Guidance-wise, 3.5 is a reasonable first guess, but if your Qwen endpoint behaves like a diffusion model with different CFG conventions, don't be precious about the default - the slider is there for a reason.

Installing

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

Restart, or search "Eigen AI FLUX API" in ComfyUI Manager. Dependencies are already in ComfyUI (requests, pillow, numpy, torch). No models, no VRAM, no downloads - the whole point is that the heavy lifting is someone else's problem.