⭐ Star Qwen / WAN Ratio

The most common local image-to-video pipeline in 2026 looks like this: generate a still with Qwen-Image (or edit one with Qwen-Image-Edit), then feed that latent into Wan to animate it. Both models are Alibaba, both have their own native resolutions, and neither shares them with the other. Qwen wants 1328×1328; Wan HD wants 960×960; Wan Full HD wants 1440×1440 at 1:1 and 1920×1080 at 16:9. Memorizing two resolution tables so your still doesn't get mangled on the way to video is exactly the kind of bookkeeping a node should do for you.

StarQwenWanRatio is that node: an empty-latent generator that knows the native resolutions for all three targets and picks the right set based on a model dropdown.

What you set

• model - Qwen, Wan HD, or Wan Full HD. This switches which ratio presets apply. It's the most important field, and the one to check first when the latent comes out the wrong size.
• qwen_ratio, wan_hd_ratio, wan_fhd_ratio - ten options each, from 1:1 through 16:9, 9:16, 4:3, 3:2, 5:7, 7:5, down to Free Ratio (custom) with custom_width/custom_height (16–8192, stepped by 16). Each dropdown's labels carry the exact pixel dimensions, so you can't misremember them.
• use_nearest_image_ratio - feed an optional image in and it snaps to the closest preset ratio to that image's aspect. This is the "match my source" convenience: drop in your Qwen still and it sizes the Wan latent to the nearest matching preset instead of you doing the math.
• batch_size - 1–64.

Outputs: latent (for your sampler), width, height (integers for downstream nodes). Wire latent into Wan's sampler and you're done.

Why you'd bother

Two reasons, both practical. First, the obvious one: a single node that covers both halves of a Qwen→Wan graph removes a whole class of "why is my video squished" errors - the resolution tables are in the dropdown labels, not in your head. Second, the model switcher is the honest core: Wan is where open video consolidated (the family froze at Wan 2.2 - the numbered series went API-only, so 2.1/2.2 checkpoints are what everyone actually runs), and its resolutions genuinely differ from Qwen's. One node, both families, zero table lookups.

Install & gotchas

Pack standard: ComfyUI_StarBetaNodes is the retired beta repo; the nodes now ship in ComfyUI_StarNodes - install via ComfyUI Manager (search Starnodes) or git clone https://github.com/Starnodes2024/ComfyUI_StarNodes into custom_nodes, then restart. No deps, no models.

The gotcha is the model dropdown itself: it only changes the preset table, not your loaded checkpoint. If you set model to Wan Full HD but your workflow is actually running the HD checkpoint, you've made a 1440×1440 latent the 960-model has opinions about. Match the dropdown to what you loaded - that's the entire trick to this node.