Jib Mix Qwen · Text to Image

A known mixer on a good open base

The Jib Mix name has been the answer to "how do I stop everything looking like plastic?" for a couple of generations now - Jib Mix Realistic XL in the SDXL era, Jib Mix Flux for the over-smooth Flux problem. This is J1B's pass at Qwen-Image, and Qwen-Image is a good base to pass at: a 20B MMDiT under Apache 2.0 that came out of the box with the best skin texture and the best literal text rendering in open weights.

The result is a photoreal generator you can call for about 3.6 credits an image, which is the cheapest quality-per-credit in this whole corner of the PopNow catalog. No 20 GB download, no quantization decisions, no offloading. What you don't get is a sampler - the node exposes the model's published fields and nothing else.

What happens under the hood

Your prompt goes to PopNow, the run gets quoted, the task is submitted with an idempotency key so a network retry doesn't double-bill, the node polls it until it settles, downloads the image and decodes it into an IMAGE. There's no model on your disk and no VRAM cost; the pack declares zero dependencies and uses the aiohttp, torch and Pillow ComfyUI already ships.

The trade you're making: a hosted endpoint instead of weights. That means your prompt leaves the machine, the vendor's filter has final say, and the prompt-only widget set is all the control you get. For a photoreal portrait workflow that's often fine - the thing that made Qwen-Image good was the model, not your CFG value.

Inputs

• prompt - the tooltip is a good checklist on its own: describe the person, clothing, place, light and camera feel. And quote any words that should appear in the image, because Qwen's text rendering will genuinely honour them, signage and all.
• aspect_ratio - 1:1 default. Portraits want 2:3 or 3:4; the model is comfortable at any of them.
• output_format - jpeg default. Keep it for delivery, change it if you're going to grade.
• seed - 0, with fixed control after generate, so queueing twice gets you the same image and one bill. Set randomize when you're hunting for a face.

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

No LoRA socket on this variant - if you want adapter stacking, use the Jib Mix Qwen LoRA node, about a credit more per image.

Install

Manager → search PopNow → Install, or:

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

Restart, then key: popnow.ai/api/console/keys into Settings > PopNow, or POPNOW_API_KEY in the environment for a server. Restart a second time so the model list loads from your account - a missing node after a fresh install is almost always that restart. Remove custom_nodes/popnow_api.py if it's there from the older single-file node.

You do need a current ComfyUI: the pack targets the V3 node API. Everything else - GPU, VRAM, model downloads - is irrelevant.

Prompt and troubleshooting notes

The photoreal failure mode with mix models is over-description. Asking for skin pores, individual eyelashes and subsurface scattering at once gets you a waxwork; ask for a person, a place, and light, and let the mix do its job. Camera language ("35mm, shallow depth of field, overcast window light") moves this model more than adjective stacking.

Error codes, briefly: PopNow 401 is a bad key, 40201 is an empty wallet, 400 names the field it refused, 40004 means the combination isn't servable, 40310 wants an email verification, and content_rejected or generation_failed means the model declined or broke - refunded either way.

Two behaviours people trip on. There's no seed randomization by default, which is a feature - no accidental double charges while you edit a prompt. And cancelling doesn't cancel: the task keeps running on PopNow and is charged if it succeeds.

Set Max credits per run in the PopNow settings as a sanity ceiling, and treat this as what it is - a very new pack from a vendor nobody's discussing yet. The upside is the price; the downside is that you're the one discovering the sharp edges.