FAL Qwen Edit Plus

FAL Qwen Edit Plus runs Qwen-Image-Edit over fal.ai - the API edition of the model the KB calls the default open instruction editor since late 2025. You hand it up to three images and a sentence, and it edits. The whole pitch of Qwen-Image-Edit is that it killed a stack of older machinery: object removal, garment swaps, relighting, re-posing - jobs that used to take a mask plus a ControlNet plus an IP-Adapter are now "say what you want." This node gives you that without downloading the 20B model, which is the part that matters if you don't have the VRAM.

The endpoint is fal-ai/qwen-image-edit-2511/lora - the 2511 line, i.e. the newer Qwen-Image-Edit checkpoint, with the lora suffix exposing LoRA support. The source uploads your reference images, submits the prompt, polls, and returns the edited image.

The inputs that matter:

• api_key - FAL key (node input or FAL_KEY env var).
• prompt + image1 - the instruction and the source. Up to 3 input images total (image2, image3 optional); the Qwen edit models genuinely use multiple references well - this is where "change the background to match this other image's style" becomes a two-image job.
• width / height - 64 to 4096, which is a much wider range than the Flux 2 edit nodes in this pack. Qwen-Image-Edit re-emits the whole frame at whatever geometry you set.
• num_inference_steps (default 28), guidance_scale (default 4), seed, output_format (jpeg/png), enable_safety_checker, acceleration (none/regular).
• lora_url / lora_scale (default 0.84) - hosted LoRA for style. The KB is emphatic that Qwen-Image-Edit's LoRA library is what separates it from its own parent model - pose, angle, and expression LoRAs trained on the editor do not work on plain Qwen-Image. The non-integer default scale (0.84) hints that the author tuned this against the model's real behavior.

Outputs: image, image_path, generation_info, generation_time, saved to output/API/FAL/qwen-image-edit/.

Context that matters for actually using it. The KB's coverage of Qwen-Image-Edit is unusually rich and unusually honest: it won the editor category on licensing (Apache 2.0 - no anti-circumvention nonsense like Kontext's), but its structural flaw is that it re-emits the entire frame, so pixels nobody asked about move and faces drift across a chain of edits. The 2026 fix is to bolt a mask back on around it. This node doesn't have a mask input - it's pure instruction editing - so if you need bit-identical unmasked regions, this is the wrong tool and the pack's Kontext conditioning node is the right one. If you're restyling, swapping, or removing, this is the one to reach for.

Other practical notes: it's paid per call, and the Qwen editor is a 20B model - even via API, "plus" jobs with multiple references and high resolution are the slow/expensive end of the pack. Keep guidance around the 4 default; the edit models get weird at high guidance. And the KB's workflow wisdom transfers directly: short prompts undersell it, so describe the edit precisely - these models actually parse the text.

Install is the pack standard: Manager search "Node Fun", or clone https://github.com/nofunstudio/Node_Fun_ComfyUI into custom_nodes, pip install -r requirements.txt, restart, set your FAL key. If you already run Qwen-Image-Edit locally this is a redundant convenience; if you've been hearing everyone say "just use Qwen Edit" and can't fit the weights, this is the version that fits.