Qwen Image 1.0 LoRA · Text to Image

Qwen-Image's launch pitch was text rendering - clean, commercial-grade glyphs in English and Chinese, decades ahead of the smear that earlier models produced on a shop sign. That reputation is why a 20B model stayed relevant even after smaller models outran it on pure image quality: nothing else lettered as reliably.

This node is that model, generation task, with a loras input bolted on. 4.5 credits ($0.045) an image. If your work involves words in pictures - packaging, posters, signage, UI mockups - this is the sensible starting point, and the LoRA socket is how you stop it looking like every other Qwen render.

Why the LoRA socket is the interesting part

Qwen-Image is Apache 2.0, every weight of the family, and that's the structural reason the community built on it rather than on Flux Kontext: the ecosystem is enormous and it isn't going to get taken down. The LoRA library for this family runs from photographic styles to character sheets to type treatments.

Here you connect a PopNow LoRA node to the loras socket, set its source and scale, and chain up to three of them. Source can be a local .safetensors in ComfyUI/models/loras - uploaded once and reused for twelve hours - or a Hugging Face repo, a Civitai page, a direct link. Gated files need HF_TOKEN or CIVITAI_API_KEY in the environment, deliberately not in a widget.

Two rules the pack's own tooltips give you: put any trigger word first in the prompt, and keep LoRA stacking modest - a style at 0.7 plus a character at 0.6 usually behaves better than three at full strength.

Fields, and the one that traps people

prompt - "Describe the subject, composition and lighting." Write sentences, not tag soup. The encoder behind this family is a language model, and it rewards a paragraph the way an LLM rewards a clear brief.

aspect_ratio - 1:1 by default. Change it early; it's baked into the render, and re-rendering at a new ratio is a new charge.

output_format - JPEG default, PNG when you care about banding or plan more processing.

seed - fixed by default. Queue the same graph twice and you get the cached first result instead of a second bill. That's a feature that reads like a bug: same graph, same picture, no new charge. Flip control after generate to randomize when you want another roll.

loras - optional; up to three.

Outputs are images and task id. No alpha channel on this one - if you want cutouts, the 2.1 models in this pack expose an alpha mask output.

Install

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

Restart, or install through Manager by searching PopNow. The pack has no dependencies to install beyond what ComfyUI already bundles (it does its HTTP with the aiohttp your ComfyUI already has), runs on Python 3.10+, and requires a ComfyUI recent enough for the V3 node API. It works on machines with no GPU at all, because nothing runs locally.

Then the key: popnow.ai/api/console/keys → Settings → PopNow → paste → restart. The second restart is the one people skip; the node list is built from your account's catalog at startup, so new models don't appear until you do it. On a server, use POPNOW_API_KEY instead so the key never lands in a shared file.

Where it goes wrong

Complaints about missing models are almost always that skipped restart. No PopNow API key is set means the key didn't save; PopNow 401 means it saved but is wrong or revoked. 40201 is an empty wallet - 4.5 credits an image goes fast when you're exploring, so set Settings → PopNow → Max credits per run to something like 50 before you start.

And one expectation to set: this is the original Qwen-Image. There are newer revisions in this pack (2512, 2.1, 3.0) that render better type and take more references. What 1.0 gives you is the most mature LoRA ecosystem, which is a real reason to stay on it if a specific LoRA is what you're building around.