Wan 2.7 Spicy Image to Video, Explained
Wan 2.7 Spicy is the uncensored branch of the Wan video model family. It takes one still photograph and generates motion from it, at 720p and above, in clips of five to fifteen seconds, with sound generated alongside the picture. This page covers what it does, where its limits actually are, and how to prompt it — then points you at a generator where you can make something in the same shape.
A reference page, not a spec reprint. Everything below is written for someone deciding whether this kind of model suits their footage.
What Wan 2.7 Spicy generates
The model’s published capabilities, and what each one means in practice once you are actually using it rather than reading about it.
- Input
- One still image, plus an optional motion promptImage-to-video only. Without a starting frame there is nothing for it to animate.
- Output resolution
- 720p and aboveHigh enough that faces survive. Which resolutions a given service exposes varies.
- Clip length
- Five to fifteen secondsFifteen is long for a single unbroken take. Plan what changes across it.
- Native audio
- Generated with the pictureRoom tone, dialogue and effects come out of the same pass, not a second one.
- Aspect ratio
- Taken from the input imageThere is no ratio parameter. Crop the still first if the shape matters.
- What "spicy" means
- A build tuned for adult creative workDifferent weights, not a general model with moderation switched off.
Specifications are the model’s, published by its developers. Individual services expose different subsets of them, so check whichever one you use rather than assuming every option is available everywhere.
Getting a usable clip out of it
Three steps. The third is where most disappointing first attempts are actually decided.
- 1
Choose a still that can survive being moved
This class of model frames the whole clip from your photograph, so the composition and crop you upload are the composition and crop you get back. Look for one clear subject, filling a decent share of the frame, in light even enough that detail is not crushed into black. Anything the photograph hides has to be invented as motion reveals it, and invention is where these models look worst.
- 2
Write the movement, not the picture
The scene is already settled by the image. What the prompt controls is what happens next. "She turns toward the camera and smiles, hair shifting in the breeze" gives the model something to do. "A beautiful woman in a red dress" only restates what it can already see, and clips prompted that way tend to drift or barely move at all.
- 3
Say what the sound is, if you want sound
Because audio is generated in the same pass rather than added afterwards, it takes direction from the same prompt. Naming the ambience — a room tone, rain against a window, a line of dialogue — gets you a soundtrack that belongs to the scene. Leaving it unsaid gets you whatever the model infers, which is usually generic.
What image-to-video looks like when it works
Both clips below were made on this site. They are here to show the shape of the output this kind of model produces, not to benchmark any particular one.
Wan 2.7 Spicy next to its neighbours
The models people weigh against it, on the axes that actually decide which one fits a job. Published specifications; no scores, because published benchmark claims in this category are not worth repeating.
| Model | Resolution | Lengths | Stands out for | Best suited to |
|---|---|---|---|---|
| Wan 2.7 Spicy | 720p and above | 5 to 15s | Native audio in the same pass; tuned for adult creative work | Material that general-purpose video models decline outright |
| Wan 3.0 | Up to 1080p | Up to 30s in one pass | The longest single take in the family, with ratio control | Scenes that need somewhere to go rather than a single beat |
| Wan 2.6 Spicy | 720p / 1080p | Around 15s | Multi-shot sequencing and audio-synced output | Short sequences that need more than one camera setup |
| Wan 2.2 Spicy | 480p / 720p | Short takes | The cheapest way to find out whether a photo animates at all | Iteration, before committing a still to a longer render |
| Seedance 2.5 Spicy | Up to 4K | Up to 30s | End-frame control and heavy multi-reference input | Work where the last frame matters as much as the first |
Every row is a different answer to a different question. The useful comparison is rarely "which is best" and almost always "which one has the constraint I care about".
What "spicy" actually denotes
The word gets used loosely enough to be worth pinning down, because the common assumption about it is wrong. A spicy build is not a general-purpose model with its moderation turned off. It is a separate set of weights, trained and tuned so that ordinary adult creative work is inside the distribution rather than a category the model was taught to refuse.
That distinction has a practical consequence. A general model asked for this material does not merely decline — when it does comply, it tends to comply badly, because the request sits at the edge of what it saw during training. A build tuned for the material handles it as a normal request, which is why the quality gap shows up in anatomy and motion rather than only in whether you get an output at all.
It is not a promise of no boundaries anywhere. Every service that exposes one of these models applies its own rules on top, and the serious ones all draw the same lines: consenting adults, lawful material, and no using a real person’s likeness without their agreement.
The prompt mistake that looks like a model limitation
One failure is common enough to name, because almost nobody guesses the cause. Words that work as quality boosters when generating an image from text — raw, unedited, unfiltered, real-life — behave completely differently once an input photograph is present. With an image in hand, the model reads them as an instruction to leave that image alone.
What comes back looks like a slow push into your original still rather than an animated scene. It is easy to read as the model being weak, or restricted, or as the wrong tier being selected. It is none of those. Strip those words out, name one concrete movement, and the same photo will usually behave.
- Cut from an image-to-video prompt: raw, unfiltered, unedited, real-life and close variants.
- Keep anything that names a movement, a camera direction, or how the shot ends.
- Give a fifteen-second take somewhere to arrive, not just something to do.
- Delete adjectives that only restate what the photograph already shows.
Where this class of model still falls down
Hands and hair remain the honest weak points across every build in this family, and they fail in a specific way: fine at rest, unreliable once they move quickly through frame. If a clip depends on a hand gesture reading clearly, plan on several attempts or on choosing a still where the hands are settled.
The other recurring problem is drift. A long take with nothing planned to happen inside it wanders — a face slowly stops being the same face, a background reorganises itself. This is not solved by a longer duration or a better model. It is solved by deciding, before you generate, what is different between the first second and the last.
✨ Make one here
Image to Video
Upload a photo, describe the motion, get a clip back
Compare the models
Which model suits which job, side by side
Image to Image
Fix or restyle the still before you animate it
Reference to Video
Several reference images held consistent in one clip
Wan 2.7 Spicy questions
The questions that come up most often around this model.
Make one and see
Reading about a model only gets you so far. New accounts get a video on the house — no card, no trial countdown. Upload a photo, describe the motion, and watch what comes back.