SFWaveSpeedQwenTextToImageLora

The plain Qwen text-to-image node is fine, but plain generation is where hosted image APIs underdeliver - anyone can do "a person, generic style." SF WaveSpeed Qwen Text to Image LoRA is the version that lets you attach up to two LoRA models to the hosted Qwen-Image call, so your output can actually match a style, a character, or a realism look instead of whatever the base model defaults to.

This is the node that makes "hosted but stylistically mine" a real option. The tooltip's example LoRA - flymy-ai/qwen-image-realism-lora - is a hint at the ecosystem: there's a growing catalog of public Qwen-image LoRAs you can reference by name, no download required. If you've built up Qwen LoRA workflows locally, this gets you a similar lever without the VRAM.

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

How it works

You give it a prompt, pick the size, and optionally supply up to two LoRAs as username/model-name paths (or full URLs) with strength scales. The node bundles all of that into a call to WaveSpeed's Qwen text-to-image endpoint with LoRA support, and the server applies the style during generation.

The inputs that matter:

• lora_1_path / lora_2_path - LoRA references. Format is strict: username/model-name (like flymy-ai/qwen-image-realism-lora) or a full URL. The pack validates this and rejects bare filenames - the server can't fetch your local disk, so don't try.
• lora_1_scale / lora_2_scale - 0.0–2.0, default 1.0. Your strength control.
• size - the Qwen native presets (1:1 (Square) default through the 2:3 portrait), with custom_size as an override.
• prompt, seed, output_format, enable_sync_mode - same as the base node.

Output is a single output_image (IMAGE).

Installing it

Pack install plus WaveSpeed key:

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

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

Troubleshooting

• "Invalid LoRA path format." The strict validator again - it wants username/model-name or a full URL. That error is the pack protecting you from a call that was never going to work.
• LoRA has no visible effect. The referenced model may not exist on WaveSpeed's host, or its scale is too low. Bump lora_1_scale toward 2.0 to see if it's actually loaded; if nothing changes, the path is wrong.
• Output looks over-styled or broken. Two LoRAs at full scale fighting each other. Drop one to 0.5 or remove it entirely.
• Custom size hurting quality. Same as the base node - non-native sizes make Qwen-Image fall apart. Presets are your friend.

The honest caveat: LoRA choice is the whole game here, and it's constrained to what's hosted. But for the public Qwen-image LoRA catalog, this is the easiest path to stylistically-controlled generation inside ComfyUI - a real answer to "I want Qwen quality but my own taste."