WaveSpeedAI Qwen Image Text-to-Image LoRA

Qwen-Image's superpower is text rendering, but its personality comes from the LoRA ecosystem around it - realism, style, and character adapters that give the 20B base model a look. The catch has always been the same: 20B of VRAM plus a LoRA folder to manage. This node runs the base model on WaveSpeed AI and lets you attach up to two LoRAs by name - user/model format, resolved server-side - so you can audition the community's style library without downloading a single adapter file.

What it is

A hosted Qwen-Image text-to-image node with LoRA slots, sitting alongside the pack's plain Qwen Image Text-to-Image. The differences: lora_1_path / lora_2_path (with scales 0.0–2.0, default 1.0) attach adapters by reference, and the size dropdown uses friendlier aspect labels - "1:1 (Square)," "16:9 (Widescreen Landscape)" - instead of raw pixels. There's still a custom_size escape hatch if you need exact dimensions.

How it works

client from the WaveSpeedAI Client node, then a POST to the Qwen image generation endpoint with prompt, size, LoRA paths, and scales. The generated image downloads back as an IMAGE tensor. LoRA paths follow the same rule as the pack's other LoRA nodes: a full URL or username/model-name. If the path is wrong, the API rejects the whole request - no partial generation.

The inputs that matter

• prompt - the image to generate; the tooltip example is a character portrait, and realism LoRAs pair well with detailed subject descriptions.
• size - labeled aspect ratios, or custom_size for exact pixels.
• lora_1_path / lora_2_path - up to two adapters in user/model format. flymy-ai/qwen-image-realism-lora is the tooltip's example.
• lora_1_scale / lora_2_scale - start at 1.0; style overwhelms fast at 2.0.
• seed, output_format, enable_sync_mode (default on).

Output name is output_image, type IMAGE.

Install & setup

cd ComfyUI/custom_nodes
git clone https://github.com/razvanmatei-sf/razv-wavespeed.git
cd razv-wavespeed
pip install -r requirements.txt

or ComfyUI Manager → "razv-wavespeed" → Install, restart. API key from wavespeed.ai in config.ini (from config.ini.tmp), the Client node, or WAVESPEED_API_KEY. Deps: requests, pillow, numpy<2.0.0, torch.

Common issues

• Invalid LoRA path - must be a URL or username/model-name. Double-check the exact casing of the model name.
• Style fights the subject - lower the scale, don't rewrite the prompt around the LoRA.
• "No output received" - job timeout or transient API error; retry.
• 401 - key issue.

The use case that justifies this node: LoRA window-shopping. Five renders, five style LoRAs, no downloads, and you learn which direction you actually like before committing to a local Qwen-Image install. Once you know the look you want and you're generating constantly, the local route is cheaper - but as a tasting menu for the LoRA ecosystem, hosted is hard to beat.