qwen / 20b / createVariant

The img2img variant of the Qwen-Image cloud node. You feed it an image plus a prompt, and it re-renders that image with Qwen's 20B MMDiT on Civitai's sdcpp fleet - same text-rendering chops, no 20B download. This is the node for "I like this render, now redo it with different words on the sign / a different mood / a different composition," where the words actually need to stay legible.

It's the natural companion to the Qwen create node, and it inherits the same defining strength: legible text survives into the variant. An img2img pass on most models will smear a logo into noise; Qwen re-renders it properly. One structural quirk to know: this node makes width and height required - unlike the create node, you must declare the output size, and there's no auto-inferring from the input.

How it works

Standard pack client: submit engine: sdcpp, ecosystem: qwen, model: 20b, operation: createVariant with the source image, prompt, size, and strength; long-poll; get a ComfyUI IMAGE back.

The inputs that matter

• prompt (required) - the change you want. Include any text that should appear, spelled exactly.
• image (required, IMAGE) - the source to re-render.
• width / height (required) - the output size, no default inference. Match the source's aspect or Qwen will reframe it.
• strength - how much of the original survives. Low (~0.4) restyles, high (~0.7) rewrites.
• version - latest / 2509 / 2512, pinning the Qwen build.
• steps / cfg_scale - 20 / 2.5 defaults.

Install

Pack install: ComfyUI Manager → "Civitai Comfy Nodes" → Install → restart, or clone https://github.com/civitai/civitai-comfy-nodes.git into custom_nodes, pip install -r civitai-comfy-nodes/requirements.txt (just requests), restart.

Common issues

Credentials: Civitai Auth node into api_config or CIVITAI_API_TOKEN env var (from civitai.com/user/account); headless needs the env var.

Two variant-specific traps. First, the required width/height: people wire the image and queue without setting size, and the node blocks with a missing-required-input error - this is one of the few nodes in the pack that validates it locally. Second, text preservation is good but not telepathic: if you change the composition substantially (high strength), expect the text to be re-set in the new frame rather than transferred pixel-for-pixel. Spell it out in the prompt if the words matter.

Billing: 20B variant jobs are mid-weight, and there's no free retry. Nail the strength at quantity: 1 before scaling up.