Catalog
Which uncensored video model to use, and which to skip
Seedance 2.5 is the one to reach for when the clip has to hold together, and the two Wan 3.0 models plus HappyHorse 1.1 cover most work below it at a fraction of the cost. The cheap end — Seedance 2.0 Fast and the base MiniMax H3 — is worth using to test an idea and worth avoiding for anything you intend to deliver.
Measured on September 8, 2026 against the catalog as it stood that day; models and their costs change.
How this was judged, and what it is not
This is a working note about our own catalog, written by the person who runs it, from generating with these models rather than from reading about them. It covers no platform but this one, and it makes no claim about any tool you might use elsewhere. Nothing here was benchmarked against a scoring rubric; it is judgement built up from output, and it is stated as judgement.
One thing deliberately absent: speed. Generation time is not something the catalog reports and not something measured properly here, so this page says nothing about which model is fast. Cost and length are structural — they come from the catalog and they are exact — and those are the numbers on each model's page.
The catalog also moves. Models are added, retired and repriced, so this reflects the models as they stood on the date shown above. If that date is old, treat the page as history.
Seedance 2.5 when consistency is the whole job
Seedance 2.5 is the best model in the catalog by a clear margin, and the gap is about consistency rather than any single frame. Other models produce good output; they do not hold it together across a clip as reliably. When a scene has to stay coherent from the first frame to the last, this is the one that does it.
It is also the most expensive per second by a wide margin, and that is the honest trade. It earns the price for professional work — the clip a client sees, the sequence that goes into a finished piece — and it is overpriced for a test you are going to throw away. Use it when a failed clip costs more than the credits do.
It carries the longest single-generation length in the catalog, matched only by the two Wan 3.0 models. That matters when a shot has to run without a cut: stitching two clips together introduces a seam the model never intended, and the longer ceiling avoids the problem entirely.
Wan 3.0 Pro and HappyHorse 1.1 for most work
Wan 3.0 Pro sits just below Seedance 2.5 and costs considerably less per second. It is not as consistent, but it is genuinely good, and for a great deal of work the difference does not survive contact with the finished piece. If the budget matters at all, start here rather than at the top.
It reaches the same maximum clip length as Seedance 2.5. Worth stating plainly, because the base Wan 3.0 reaches it too — length is not what separates Pro from the base model in this family. Quality is. Both go long; Pro holds together better while it does.
HappyHorse 1.1 costs the same per second as Wan 3.0 Pro and is genuinely strong on consistency and quality. Its ceiling on clip length is lower, so it is the wrong choice for a long unbroken shot and a good one for everything shorter. Between these two the decision is length first, then taste.
The middle: MiniMax H3 Max and Wan 3.0
MiniMax H3 Max and the base Wan 3.0 share a price point and sit in the middle of the range. Both are usable for real work. Both are a step down from the three above in how well the result holds together, and you notice it on longer or busier clips rather than on simple ones.
The two differ in feel. Motion from the MiniMax family reads more synthetic — movement that is technically correct and slightly wrong in a way that is easier to see than to describe. Wan 3.0's motion sits closer to natural, which for the same price is worth weighing if the clip has people or animals in it.
Wan 3.0 also reaches the long ceiling, which H3 Max does not. If the shot needs to run, that decides it.
The cheap end, and when not to use it
Seedance 2.0 Fast and the base MiniMax H3 are the least expensive video models in the catalog, and they are useful. Use them to find out whether a prompt is pointing anywhere before you spend real credits on it, or to check that a composition works at all.
For anything you intend to deliver, do not. Both produce motion that reads as artificial, and the quality gap against the models above is not subtle enough to argue about. A cheap clip you cannot use is more expensive than an expensive clip you can — that is the entire calculation, and it is why these two live in the drafting stage rather than the final one.
This is the only recommendation on this page phrased as a warning, and it is deliberate. A page that finds every model excellent is not telling you anything.
Starting from an image instead of a description
Every video model here has an image-to-video counterpart that animates a still you supply rather than composing from text alone. The ranking is not identical to the text-to-video one: Seedance 2.5, HappyHorse 1.1, Wan 3.0 Pro and MiniMax H3 Max are the strongest at staying faithful to the frame you give them.
The rest work. They are simply less reliable about keeping the reference intact, which is the whole point of starting from one. If the still is something you already committed to — a character you established, a composition you framed — that reliability is what you are paying for.
The practical route is to settle the frame first with an image model, then hand it over. The guide on going from a still to a clip covers the workflow, and the image model guide covers what to use for the frame itself.
What uncensored means across the catalog
Every video model here runs without a prompt filter: the description you write reaches the model as written, motion wording included, and nothing paraphrases it into something safer first. That is a property of the platform rather than of any one model, so it is not a reason to pick between them.
Quality on that material does differ. HappyHorse 1.1, Seedance 2.5 and Wan 3.0 Pro handle it noticeably better than the rest — the same three that lead on general consistency, which is not a coincidence: material a model handles badly is usually material it was trained thinly on.
The refusals that remain are narrow, fixed and absolute: no depiction of anyone under age, and no real identifiable person without consent. Neither is a setting on any model. The rules page has the policy in full.
Frequently asked questions
Is Seedance 2.5 worth the price difference?
For work you are delivering, yes. Its advantage is consistency across a whole clip rather than a prettier frame, and that is exactly what fails expensively when it fails. For testing an idea it is poor value — the cheap models exist for that.
What is the difference between Wan 3.0 and Wan 3.0 Pro?
Quality, not length. Both reach the same maximum clip length, so a long unbroken shot is possible on either. Pro holds together better across that length and costs more per second; the base model is a reasonable choice when the clip is simpler.
Which model should I never use for finished work?
Seedance 2.0 Fast and the base MiniMax H3. Both are cheap and both produce motion that reads as artificial. They earn their place as drafting tools — check a prompt, check a composition — and they will disappoint you anywhere else.
Do any of these models handle adult content better than others?
Yes. All of them generate it, since nothing filters the prompt, but HappyHorse 1.1, Seedance 2.5 and Wan 3.0 Pro produce noticeably better results on it. The same three lead on general consistency.
Which is best for animating an image I already have?
Seedance 2.5, HappyHorse 1.1, Wan 3.0 Pro or MiniMax H3 Max. All four stay closer to the frame you supply than the rest do, which is the property that matters when the still is one you already committed to.
How long can a single clip be?
It depends on the model, and the ceilings differ enough to drive the decision. Seedance 2.5 and both Wan 3.0 models reach the longest single generation in the catalog; the others stop shorter. Each model page states its own range, read from the catalog as the page loads.