SF WaveSpeed Qwen Edit LoRA

The plain Qwen Edit node is great, but "edit in a specific style" is where hosted editing usually falls flat - the model has no idea what your aesthetic even is. SF WaveSpeed Qwen Edit LoRA fixes that by letting you attach up to two LoRA models to the hosted Qwen-Image-Edit call. You get the model's editing chops and a style you've chosen, without running anything locally.

Think of it as the pack's answer to style-consistent editing: a character LoRA, an art-style LoRA, or a realism LoRA rides along on the API call, and the edited output stays in that lane. If you've been bouncing between local Qwen-Image edit runs and hosted ones, this is the hybrid you've been looking for.

It's part of the comfyui-stillfront pack, under Stillfront/WaveSpeed.

How it works

The node takes your source image URL and edit prompt, attaches the LoRA paths and scales you specify, and sends the whole bundle to WaveSpeed's Qwen edit endpoint. The LoRA is applied server-side - you just supply the reference.

The LoRA inputs are where the model's conventions matter:

• lora_1_path / lora_2_path - the tooltips give the exact formats. A Hugging Face-style username/model-name path (the examples are flymy-ai/qwen-image-style-lora), or a full URL. The pack validates this format and raises a clear error if you pass a bare filename.
• lora_1_scale / lora_2_scale - 0.0 to 2.0, default 1.0. This is your style-strength knob: 1.0 is the usual starting point, higher for aggressive style, lower to keep the edit more faithful to the source.
• LoRA 2 is optional - the node handles one or two.

Everything else matches the non-LoRA sibling: prompt, image_url (from an Upload Image node), seed, output_format, enable_sync_mode. Output is a single output_image (IMAGE).

Installing it

Pack install, then your WaveSpeed key:

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

Restart, and wire in an SF WaveSpeed Client node with your key (or use config.ini / WAVESPEED_API_KEY).

Troubleshooting

• "Invalid LoRA path format." The node's validation is strict: it wants either a full http(s):// URL or username/model-name. A raw filename or a Windows path will get rejected - and that's the pack doing you a favor, since those never work on the server side anyway.
• Style is too strong or invisible. That's the scale. Crank lora_1_scale up toward 2.0 for a strong look, down below 1.0 for subtle. Two LoRAs at full scale can fight each other; if output gets weird, drop LoRA 2 to 0.5 or remove it.
• LoRA doesn't exist on the server. The path must be one WaveSpeed's model host can resolve - a community LoRA that lives only on your disk isn't available to the API. Use a hosted/public LoRA or upload it somewhere WaveSpeed can fetch.
• Edit quality drops vs. the plain node. LoRA guidance and edit fidelity pull against each other; it's a trade, not a bug.

One honest caveat: you can't train here. If the LoRA you want isn't already hosted, this node won't help - but for the growing catalog of public Qwen-image LoRAs, it's the easiest way to get style-consistent hosted edits in one call.