Qwen Image 2.1 LoRA · Edit

This is the node for the specific job of "edit a photo, but in my style". Qwen Image 2.1 Edit brings the reference count up to ten and the resolution tier up to 2K; the LoRA socket brings whatever you trained or downloaded. Roughly 6.3 credits ($0.063) an image, which is the edit premium over the plain 2.1 t2i.

The reason people care about the LoRA half: Qwen-Image-Edit's own LoRA library is enormous precisely because it's the part of the family that gets fine-tuned. Pose, angle, expression, next-scene, garment and style LoRAs exist for the editor, and mostly don't transfer to the base generation model - so if the LoRA you want is edit-trained, this node is the one that can use it.

Wiring the LoRA

loras is a custom socket that only accepts a PopNow LoRA node. Drop one in, set two things:

• source - a .safetensors filename in ComfyUI/models/loras (uploaded once, up to 2 GiB, reused for twelve hours), a Hugging Face repo id or page, a Civitai page, or any direct https:// link. Hugging Face and Civitai pages are imported server-side, so a gated file needs HF_TOKEN or CIVITAI_API_KEY in the environment. Not a widget - the pack is explicit that widget values are saved into every workflow and image you share.
• scale - the strength. This is where you'll spend your fiddling time.

Chain up to three LoRA nodes to stack them. In practice 0.7 + 0.6 beats 1.0 + 1.0; two LoRAs at full strength fight over the same conditioning.

The rest of the fields

prompt - "Describe the subject, composition, lighting, and any wording to render." Trigger word first if the LoRA has one. Because the whole point is applying a trained look, the prompt should describe the finished image, not just the delta.

image_urls - Autogrow, one to ten, ordered. Where the 2512 and 1.0 editors cap at three, this one lets you pile on the references: person, wardrobe, set, product, colourway. Each extra reference raises the quote, so the badge price is a floor, not a total.

aspect_ratio - 1:1 default, and unlike the plain 2.1 edit's model default behaviour this one will reframe your source. Check it on the first run of any workflow you inherit.

resolution - 1k tier by default, 2K available.

output_format - JPEG default. PNG if the render goes back into another edit.

seed - fixed by default, so re-queueing the same graph reuses the cached image instead of buying another. Control after generate → randomize when you want a variation.

No transparency switch on this variant, and no mask input - if you need a hard edit boundary or an alpha cutout, use Qwen Image 2.1 · Edit and skip the LoRA. That split is deliberate in the API schema, not an omission in the node.

Outputs: images and task id.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Restart, or search PopNow in ComfyUI Manager and install from there. The pack declares no dependencies and installs nothing beyond itself - it leans on libraries ComfyUI already ships. Python 3.10+ and a ComfyUI on the V3 node API are the real requirements.

Key at popnow.ai/api/console/keys → Settings → PopNow → paste → restart again so the catalog loads. On a server, export POPNOW_API_KEY instead.

Notes from the sharp end

A LoRA that renders beautifully at 1k can fall apart at 2K if it was trained small; if the 2K tier looks plasticky, drop the scale before blaming the model.

Because references and images are both priced in, the fastest way to burn credits here is a ten-reference stack on a prompt you haven't tested. Test with one reference, then add.

Codes: PopNow 40201 = no credits, 40004 = unservable combination (it names the field), 400 = a rejected value, 40310 = verify your email. generation_failed and content_rejected are refunded. Keep task id - it's the only handle support has.