Qwen Image Edit 2509 LoRA · Edit

What it is

Qwen-Image-Edit-2509 is the build the open community standardised on through late 2025: the instruction editor that ate a lot of work masks and ControlNet checkpoints used to do. It takes plain language - "change her dress to blue", "remove the watermark" - and has a genuinely huge LoRA library behind it precisely because Alibaba released it Apache 2.0 with no content clause attached, so nobody takes your adapter down.

This node is that model, hosted, with a loras socket. Two reasons you'd want it: the 20B editor is not something most cards can hold, and the adapters people actually want to use - pose changers, angle sliders, clothing transfer, relight - are scattered across Hugging Face and Civitai as single .safetensors files that the pack can fetch for you.

How it works

The node mirrors the model's schema, so what you see is exactly what the endpoint accepts. The interesting half is the LoRA path. You don't upload anything through this node - you add a PopNow LoRA node, point its source at a file in ComfyUI/models/loras, a Hugging Face repo id, a Civitai page, or a direct https:// link, set a scale, and connect it to loras. A local file gets uploaded once and reused for twelve hours; a Civitai or HF page gets imported server-side, once, and reused. Chain up to three.

When the job runs, images are uploaded, the edit is performed on PopNow's hardware, and the result comes back as a normal IMAGE batch plus a task id.

The inputs that matter

prompt is an instruction, not a caption. Say what changes and what stays. And note the tooltip: refer to your sources as Picture 1, Picture 2, Picture 3. That's how 2509 indexes multiple references, and using it explicitly beats hoping the model works out which person you meant.

image_urls is an auto-growing socket group - one to three source images, and the result follows the first one's framing unless you say otherwise with aspect_ratio (default model default, which keeps the first image's shape). Multi-image editing is 2509's headline: person + product, person + scene, person + garment.

loras (optional) is where the personality is. output_format is jpeg, png or webp - pick png if the image is going into another stage, because a jpeg round-trip on a texture you're about to upscale is wasted detail.

Outputs: images and task id.

Installing it

# Manager: search "PopNow" > Install > restart
cd ComfyUI/custom_nodes && git clone https://github.com/popnow-ai/popnow-comfyui

Then a key from popnow.ai/api/console/keys in Settings > PopNow (or POPNOW_API_KEY), and restart so the catalog loads. Nothing to download; the LoRA node doesn't need the weights locally either if you point it at a URL.

Gated files on HF or Civitai need HF_TOKEN or CIVITAI_API_KEY in ComfyUI's environment. The pack reads those from the environment and never from a widget, deliberately - widget values get baked into workflows and image metadata.

Where it goes wrong

The family's known limitation is that it re-emits the whole frame rather than patching a region: unedited areas come back close but not identical, and the error compounds if you chain edits. 2509 specifically drew complaints about geometric drift that skews proportions unless your input resolution is sane - which means every edit round-trip nudges things. If you need pixel-exactness on the untouched parts, composite the edit back through a mask instead of trusting the whole-frame return.

Second trap: LoRAs for the editor are trained on the editor. A Qwen-Image base T2I LoRA will not behave here, and an SDXL or Flux LoRA is a different architecture entirely - the file will load and do nothing useful. Third: a 40004 usually means your three references plus that aspect ratio plus that LoRA count isn't a servable combination; drop one and retry. And 40201 is just an empty wallet.