Wan Video Model Guide: Settings, Variants and Licence
Wan is Alibaba’s family of open-weights video models (Wan 2.x) for text-to-video and image-to-video. You can run them locally in ComfyUI or through hosted services, and the settings are the same everywhere: prompt, negative prompt, resolution, frame count or duration, sampling steps, guidance scale and seed. Pick the variant your hardware or host supports, start with the defaults, and change one setting at a time.
Wan sits in a different category from Kling, Runway or Hailuo. It is not a website with a credit meter; it is a family of open-weights models from Alibaba that anyone can download, run and build on. That means you can generate on your own machine, pay a hosted service per second of output, or wire it into a pipeline, and the settings behave the same way in every case. This guide explains the Wan 2.x variants, what each setting actually does, and sensible starting values.
What Wan is
Wan 2.x is a set of video diffusion models released with open weights. The family includes text-to-video and image-to-video variants at more than one size: smaller models that fit on consumer graphics cards and larger ones that need more memory but produce finer detail and better prompt adherence. Newer releases have added capabilities such as first-and-last-frame conditioning and higher output resolutions; the exact list changes, so the model card for the version you download is the authority.
Because the weights are open, the ecosystem is broad. Hosted services such as Replicate, fal and Krea offer Wan behind a simple form, and ComfyUI has native support for running it locally. The local and hosted image-to-video guide compares those routes. Licence terms are set out in the model card for each release, and they differ between versions, so read the one that applies to the weights you are using before putting output into commercial work.
Choosing a variant
- Text-to-video or image-to-video. Separate checkpoints. Image-to-video is the one most creatives use because a still fixes composition and look; text-to-video is for exploration and atmosphere.
- Model size. The smaller variant runs on cards with modest memory and is quick enough for iteration. The larger variant is noticeably better at detail, hands and prompt following but needs a high-memory card or a hosted service.
- Resolution tier. Some checkpoints are tuned for a particular output size. Generating far from that size produces soft or unstable results, so match the resolution to the checkpoint.
- Quantised and distilled builds. Community versions trade a little quality for lower memory use or fewer steps. Useful for drafts; verify the keeper on the full model.
Every setting explained
- Prompt. Subject, action, camera, setting and light, in that order. Wan responds well to full sentences and film vocabulary. Keep style tags modest.
- Negative prompt. What to suppress: "blurry, low quality, distorted hands, text, watermark, static, flicker". Wan’s default negative prompts are long and worth keeping as a base.
- Resolution. Width and height of each frame. Stay close to the checkpoint’s native size; larger frames cost memory and time quadratically.
- Frames or duration. Number of frames generated, which with the frame rate gives the length. Wan models typically expect a frame count of the form 4n+1 (such as 81); the model card states the valid values. Fewer frames mean faster, more stable clips.
- Steps. How many denoising iterations the sampler runs. More steps add detail up to a point, then only add time. Somewhere in the twenties to thirties is a common range for the full model; distilled builds need far fewer.
- Guidance scale (CFG). How strongly the output follows the prompt. Too low gives generic motion; too high gives oversaturated, jittery frames. Mid single digits is the usual starting zone.
- Seed. The random starting point. Fix it to compare two prompts fairly; randomise it to explore variations of one prompt.
- Sampler and scheduler. The algorithm that walks the denoising steps. Use the one the model card or your workflow template recommends; changing it is an advanced move.
- Frame rate. Playback speed of the saved file. Wan is commonly rendered around 16 frames per second and interpolated afterwards if you need smoother motion.
Settings at a glance
| Setting | What it does | Start with |
|---|---|---|
| Variant | Text-to-video or image-to-video, small or large | Image-to-video; the largest your hardware or host allows |
| Prompt | Subject, action, camera, setting, light | Two or three full sentences |
| Negative prompt | Things to suppress | The model’s default list plus "text, watermark" |
| Resolution | Frame size | The checkpoint’s native size |
| Frames | Clip length in frames | A valid low count from the model card, then more |
| Steps | Denoising iterations | The template default; lower for drafts |
| Guidance (CFG) | Prompt adherence vs freedom | Template default, mid single digits |
| Seed | Repeatability | Fixed while comparing, random while exploring |
| Frame rate | Playback speed of the file | The model’s native rate, interpolate later |
Hosted services expose a subset of these and may rename them; the defaults they ship are usually close to the model card’s recommendations.
A method for tuning
- Start from a known-good template. ComfyUI example workflows and hosted defaults are tuned. Change them only when you can name the reason.
- Fix the seed, then vary one setting. Steps first (does more detail appear?), then guidance (does the prompt land better without jitter?), then frames.
- Draft small, finish big. Low frame count and fewer steps to find the shot; full settings for the keeper.
- Keep the whole recipe. Model file name, variant, every setting and the seed. Locally, ComfyUI can embed the workflow in the output; on hosted services, copy the request parameters into a note. Without the recipe a good clip cannot be reproduced.
Try it
Model: Wan 2.x image-to-video, largest variant available
Image: a vintage bicycle leaning against a sunlit brick wall, ivy above
Prompt: Leaves on the ivy flutter in a light breeze. Shadows shift slowly as a cloud passes. Slow push in towards the bicycle. Static background, calm natural motion.
Negative prompt: blurry, low quality, distorted, text, watermark, flicker
Resolution: checkpoint native. Frames: lowest valid count. Steps: template default. CFG: template default. Seed: 1234
Render it, then raise only the frame count and compare stability. Then, with the seed fixed, nudge guidance up one point and see whether the breeze reads more clearly or the frame starts to shimmer. Two comparisons tell you more than ten random generations.