Qwen Image LoRA

Qwen Image LoRA is the text-to-image node that lets you attach up to three LoRA model URLs to a Qwen Image generation, so the render comes out in a style or subject you've pre-trained - and all of it happens server-side on WaveSpeed. If you've trained a Qwen-Image LoRA locally (a workflow the community has had running since mid-2025, thanks to AI-Toolkit making it fit on consumer GPUs), this is how you carry that LoRA into cloud generation: upload the weights to a public URL, paste it in, generate.

This is one of the more interesting nodes in the pack precisely because it bridges the local and API worlds. The usual split is "train locally, generate locally" or "everything in the cloud." Here you can train a LoRA on your own GPU, then use it for paid cloud renders at $0.02 per image - which is great when you need the LoRA's consistency but don't want to babysit a local diffusion pass.

How it works

The node sends your prompt plus the LoRA paths and scales to WaveSpeed's Qwen Image LoRA endpoint, waits for the render, and returns a ComfyUI IMAGE tensor. Model choice matters: Qwen Image (the 20B MMDiT) or Qwen Image 2512 (the 7B variant with better text rendering). lora_1_path and lora_1_scale are required; LoRAs 2 and 3 are optional with their own scales (0.0–4.0, default 1.0).

The inputs that matter

• model - Qwen Image or Qwen Image 2512.
• prompt - the description, Chinese or English.
• lora_1_path / lora_1_scale - the required first LoRA and its influence.
• lora_2_path / lora_3_path - optional additional LoRAs, each with a scale.
• width / height - 256 to 1536, default 1024×1024.

Standard pack extras: output_format (jpeg/png/webp), enable_sync_mode, enable_base64_output, a cache-control seed, and an optional client input.

Installing it

Part of the ERPK Collection (eRepublik-Labs/comfyui-nodes-erpk). ComfyUI Manager: search erpk → ERPK Custom Nodes → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk
pip install -r requirements.txt

WaveSpeed API key required (Settings → ERPK → API Keys). No local models - the LoRAs live at URLs.

Common issues

• LoRA fails to load - WaveSpeed needs a public, direct-download URL. Hugging Face resolve/main/... links work; expiring signed links don't.
• LoRA too strong / too weak - tune the _scale. 1.0 is full influence; drop to 0.3–0.5 if it overpowers the prompt, raise toward 2+ if it's barely visible.
• Text rendering - if the image has text in it, try the 2512 model; that's its stated strength.
• Billed per image - $0.02 per render plus LoRA processing. Keep the seed pinned when re-testing plumbing so ComfyUI serves the cached result.