Wan2.6 Image

Wan is Alibaba's model family, and the line most people know is video - Wan 2.1 and 2.2 shipped open weights and became the base of half of 2026's open video models. But the family extends further, and 2.5/2.6/2.7 are API-only. Runware_alibaba_wan2_6_image (alibaba:[email protected]) is the image-generation rung of that closed tier, served through Runware so it behaves like any other node in your graph. Same Alibaba-ecosystem DNA, zero weights to manage.

What makes this node distinctive at a glance is the resolution range: width and height go from 1 to 4096 in steps of 1. No other image node in this pack gives you that. In practice you're not generating 1px images - the range reflects that Wan2.6 Image is a flexible-aspect model that will produce whatever frame you ask for, up to very large sizes. If your workflow needs a 2K+ image in an unusual ratio without a separate upscale step, this is the node.

How it works

imageInference with the alibaba:[email protected] AIR. Prompt and referenceImages go up, image comes back as an IMAGE tensor. One toggle is worth reading before you queue anything:

• providerSettings.alibaba.promptExtend - on by default. An LLM rewrites and expands your prompt before generation, which typically improves results but breaks reproducibility: the same seed won't reproduce because the prompt itself changed. If you need deterministic runs, switch it off and write the full prompt yourself.

The inputs that matter

• positivePrompt (required) - given promptExtend is on, you can be telegraphic and let the rewriter expand; turn it off and you'd better be complete.
• width / height - 1024×1024 defaults, 1–4096, step 1. Big numbers cost more compute.
• referenceImages - the editing/consistency input.
• negativePrompt - supported; how much it helps depends on the model's guidance behavior.
• numberResults - up to 20 variations per run.

Installing it

Install the pack. ComfyUI Manager → search Runware → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Key from runware.ai/api-keys → ComfyUI Settings → Runware API key (or RUNWARE_API_KEY). Wan2.6 Image is API-only and hosted; nothing downloads.

Where people get burned

• promptExtend vs reproducibility is the real trap. People flip seeds, get different images, and blame the node - it's the prompt rewriter doing its job. Decide which you want: quality (on) or determinism (off).
• Huge resolutions are huge bills. The 4096 ceiling is capability, not a suggestion to live there. 1024 is a fine default; 4K is a different price class.
• This is image, not video. Easy to confuse given Wan's video fame - the node's category and task type are imageInference. The pack's separate Runware_alibaba_wan2_6 node is the video one.
• Wan 2.6 is closed. Unlike Wan 2.1/2.2's Apache weights, this tier is API-only - which is precisely why this node exists, and why you can't move it local later.

Wan2.6 Image is the flexible-canvas node in the pack: huge aspect range, LLM prompt boosting on by default, and Alibaba-ecosystem quality without the Alibaba onboarding. Just mind the resolution knob.