Qwen Image and Z-Image in ComfyUI: Brand-Consistent Product Images on a Private Server

Qwen Image and Z-Image are two open-licence image models that ComfyUI supports natively. Qwen Image is the one for text, posters and precise edits, with native 2K output. Z-Image Turbo is the fast one, producing a finished image in eight steps on a 16GB GPU. Both carry an Apache 2.0 licence, so a business can use the output commercially and run the models on its own server.

Why the model matters more than the prompt

When a team’s AI images disappoint, the prompt gets the blame. In our experience the model was the wrong choice for the job more often than the prompt was wrong. Text that comes out garbled, a product that changes shape when you edit the background, a generation that takes a minute per image: each of these is a model trait, and no prompt fixes it.

The good news for a business in 2026 is that two open models now cover most of the work, and both run on a single workstation-class GPU. Here is how we choose between them.

Qwen Image: text, layouts and edits

Qwen Image is the model we reach for when words have to be right. Qwen-Image 2.0, announced in February 2026, unified generation and editing in a single model and outputs at native 2K resolution. It follows long instructions, up to about a thousand tokens, which is what lets it lay out an infographic, a poster or a slide with the copy already in place.

The editing side is where product teams get the most value. Qwen-Image-Edit-2511 accepts several input images at once and holds a subject’s identity across pose and style changes, so a packaging shot can be moved into new scenes without the label drifting. ComfyUI has carried native Qwen Image nodes since August 2025 and has since added the newer 3.0 releases as they arrived.

Use it for: promotional posters with real text, catalogue and packaging edits, layouts that need to match a brief exactly.

Z-Image: speed on a modest GPU

Z-Image is a 6-billion-parameter family built for efficiency. The Turbo version finishes an image in eight sampling steps and fits within 16GB of consumer GPU memory, which is why it is the default model on our standard server plan. The base model followed in January 2026, and the Omni variant handles both generation and editing. ComfyUI supports Turbo, Omni and a lightweight ControlNet for pose and layout control.

Use it for: volume work, dozens of variations of a social visual, quick concepting where the team wants to see ten options before lunch.

Flux 2 Klein as the third option

Flux 2 Klein 4B is a small model from Black Forest Labs that does text-to-image and multi-reference editing in one network, runs in roughly 13GB of memory, and is released under Apache 2.0. It is a strong second opinion when Qwen Image’s edit is too heavy for the machine. One caution: the larger 9B version carries a non-commercial licence, so check which weights you are loading before client work goes through it.

A product-image pipeline with these models

This is the pipeline we build most often, and each stage maps to one model.

• Shoot once. Each product photographed on a plain background, one good frame each.
• Place with Qwen Image Edit. The product is masked and moved into the scenes the campaign needs. The label stays legible because the model holds the subject.
• Multiply with Z-Image Turbo. Once a scene is approved, Turbo produces the colourways, crops and seasonal variations at speed.
• Add copy with Qwen Image. Posters and banners get their headlines rendered in the image where the brief calls for it.
• Upscale, review, publish. A batch upscale, then a person approves what goes out. Nothing ships unreviewed.

All of it runs from a shared style and prompt library in SecondBrain, so the whole team draws on the same approved looks. The graphs behind it are described in ComfyUI workflows for product and brand images.

Licences and commercial use

Qwen Image, Z-Image and Flux 2 Klein 4B are all released under Apache 2.0, which permits commercial use of the models and their output. That is the reason we standardise on them for client work. Keep a note of which model and version produced each campaign asset. It takes a minute and it answers the question a brand manager or a lawyer will eventually ask.

Hardware

A 16GB-class GPU handles Z-Image Turbo and Flux 2 Klein 4B comfortably. Qwen Image is heavier and is happier with more memory, which is what the larger of our private AI server plans carries. Either way the machine lives in your office, the photos never leave it, and the number of images the team makes does not change the bill. The trade-offs against renting GPU time are in ComfyUI cloud or a private server.

Where these models fall short

Photoreal people in advertising still need care, both for likeness rights and for the uncanny moments a reviewer will catch. Exact brand colours should be checked after generation, because every model drifts a little. And where an image carries a legal claim, a regulated product, a medical device, a nutrition label, the real photograph stays. Generated imagery is for campaigns, concepts and variations, with a person signing off.

For what ComfyUI itself is and is not, see ComfyUI for business use.

MT Labs helps companies across Singapore deploy AI tools they actually own. Private infrastructure, no recurring cloud subscriptions, and a setup built around how your team already works. Whether you want these models running on a server in your office next week or just want to know which one fits your brand work, we’ll walk you through it. Get in touch and bring three of your product photos.