Jib Mix Qwen LoRA · Text to Image

If you know the name, you know why this exists

J1B is the maker behind the Jib Mix line - Jib Mix Realistic XL, then Jib Mix Flux, the finetunes people reached for specifically because base Flux gave everyone plastic skin. That's a real, well-documented gripe: "Flux needs to solve that 'plastic' look" was the thread title, and Jib Mix Flux was the answer people recommended in it. A Qwen-Image mix from the same hand is the obvious next move, because Qwen-Image is the open base with the best skin and texture in the current generation.

This node gives you that mix with your own LoRAs stacked on top. Four and a half credits an image, no download, no 20B model on your disk. If you've been eyeing Qwen-Image finetunes but don't have the VRAM or the patience for a 20 GB base plus a stack of adapter files, this is a cheap way to see what the fuss is about.

The mechanism

An API call with a ComfyUI face. The node resolves whatever's on the loras socket - local .safetensors from ComfyUI/models/loras get uploaded once and cached for twelve hours, Hugging Face repo ids and Civitai pages get imported server-side - then quotes the run, submits it under an idempotency key, polls, downloads the image and hands you a tensor. Zero dependencies, no GPU, no local model.

Worth knowing what you're giving up. A local Qwen-Image workflow with a LoRA gives you full control over sampler, CFG and the rest; here you get the widget set the model publishes, and that's it. The listing is deliberately short: prompt, aspect ratio, output format, seed, and LoRAs.

Inputs

• prompt - describe the person, place and light. The tooltip's other half matters more: if a LoRA has a trigger phrase, put it first. Mix models are already opinionated about faces; hand them a trigger word early or it'll get buried.
• aspect_ratio - 1:1 by default. Photoreal portraiture usually wants portrait; go 2:3 or 3:4 unless you're deliberately doing a square crop.
• output_format - jpeg by default. Fine for output you're delivering as a photo; switch if you're planning on grading afterwards.
• seed - default 0, and it starts on fixed control, so re-queueing the same graph returns the same billed image rather than quietly charging for a second attempt. Flip to randomize when you want variations.
• loras (optional) - up to three weight files, each with its own strength from a PopNow LoRA node. The tooltip is specific: trained for Qwen-Image or Qwen-Image-2512. An SDXL LoRA here is a wasted run.

Outputs: images (IMAGE) and task id (STRING).

Install

comfy node install popnow-comfyui

Manager users: search PopNow, Install, restart. Then create a key at popnow.ai/api/console/keys, paste it under Settings > PopNow (or set POPNOW_API_KEY), and restart again so the catalog is fetched for your account. If you still have the old custom_nodes/popnow_api.py, delete it. You'll want a recent ComfyUI - the pack is built on the V3 node API.

Reality checks

LoRA strength is the usual culprit. A style LoRA at 1.0 on top of an already-tuned mix tends to flatten skin back into the plastic look you came here to avoid; start at 0.6 and work up. And stacking three LoRAs is a good way to get a face that resembles all of them and none of them.

PopNow 400 names the field the model rejected - most often a LoRA that isn't Qwen-Image-compatible. 40201 is credits; at 4.5 an image you'll get through a small pack quickly if you're exploring with randomized seeds. 40004 is an unservable combination, 401 is the key, 40310 is your email. Declines are content_rejected and refunded.

Nobody's written the edge cases for this pack yet - searching for PopNow in the usual places returns nothing at all - so the practical advice is: small credit pack, key in the environment, and keep an eye on the printed charge after each run until you trust the numbers.