Wan is not one model. The family runs from Wan 2.2 to Wan 3.0 and is sold as 33 separate endpoints on our API (text, image and reference input, plus "Spicy", "Prime" and "Pro" builds), and they do not all treat adult content the same way. Give Wan 3.0 a text prompt for a consensual scene between two adults with restraint themes and the clip comes back with every element you named, then goes further than you wrote. The same kind of prompt on Wan 2.2 text-to-video is refused, and you aren't charged.
Short answer: mostly yes, but the endpoint matters more than the version number. Based on our September 2026 tests, 28 of the 33 Wan endpoints carry the Uncensored label. Base Wan 3.0 delivered every level on text, image and reference input, from suggestive imagery up to our two most intense probes. Wan 2.5 delivered nearly all. Base Wan 2.2 is the weak spot: from a still it tones down a two-adult scene, and from text it refuses everything, even the mildest. The cheap "Spicy" builds do better, but the newest ones tend to change the clothing of people you asked to keep dressed.
How the levels work
Five levels, each stronger than the last; the grid below uses them.
- L1 Suggestive and L2 Mature themes: suggestive and moderate NSFW imagery.
- L3 Strong: advanced NSFW imagery of one adult.
- L4 Intense: a higher-intensity scene between two adults.
- L5a is a consensual adult scene with restraint themes; L5b is cinematic gore. L5b is a different kind of content, not a stronger L5a.
Every character is a fictional adult. A clip either came back as written or was toned down: it finished and was billed, but the scene was swapped for something tamer, a partner was dressed, or the key moment was left out of frame. A few endpoints refuse instead, and those jobs fail without producing a clip.
The 30-second version
- Refused: Wan 2.2 text-to-video (even L1) and Wan 2.6 reference-to-video. Base Wan 2.2 image-to-video and Wan 2.7 text-to-image (and its Pro version) are Softened. Those five endpoints are the only ones without the Uncensored label.
- "Uncensored" means no filter on our side, not "always as written". Wan 3.0 goes further than you ask, while Wan 2.2 Spicy and the LoRA builds miss some levels (see the grid).
- For most jobs: from a still, Wan 2.2 Spicy up to L4 (cheapest), Wan 2.6 Flash or Wan 2.5 when every level must render, Wan 3.0 for up to 30 seconds with sound. From text, Wan 3.0, or Wan 2.6 when you want it to stay closer to your script.
- 5 seconds costs $0.095 on Wan 2.2 Spicy (480p), $0.225 on Wan 3.0 (480p) and $0.45 on Wan 2.6 (720p).
Try Wan 2.2 Spicy → · Try Wan 3.0 text-to-video →
How far each Wan version goes
Each cell is clips that came back as written out of runs at that level; "–" means not run. A miss is a toned-down clip unless the row says refused.
| Wan version · task | L1 | L2 | L3 | L4 | L5a | L5b | Page label |
|---|---|---|---|---|---|---|---|
| 3.0 · text, image and reference, each | 3/3 | 3/3 | 2/2 | 3/3 | 3/3 | 3/3 | Uncensored |
| 3.0 Pro and 3.0 Pro Prime · text, image and reference, each | 3/3 | 3/3 | 2/2 | 3/3 | 1/1 | – | Uncensored |
| 3.0 Prime · text and reference, each | 3/3 | 3/3 | 2/2 | 3/3 | 1/1 | – | Uncensored |
| 3.0 Prime · image | 3/3 | 3/3 | 3/3 | 0/3 | – | – | Uncensored |
| 3.0 Spicy · image, one run per level | 1/1 | 1/1 | 1/1 | 1/1 | 1/1 | 1/1 | Uncensored |
| 3.0 Prime Spicy · image, one run per level | 1/1 | 1/1 | 1/1 | 1/1 | 1/1 | 1/1 | Uncensored |
| 2.7 Spicy · image | 3/3 | 3/3 | 3/3 | 3/3 | 4/4 | 3/3 | Uncensored |
| 2.7 · edit | 3/3 | 3/3 | – | 1/3 | 1/3 | 3/3 | Uncensored |
| 2.7 Pro · edit | 3/3 | 3/3 | – | 0/3 | 2/3 | 3/3 | Uncensored |
| 2.7 · text-to-image | 3/3 | 3/3 | 1/2 | 0/3 | 0/3 | 3/3 | Softened |
| 2.7 Pro · text-to-image | 3/3 | 3/3 | 0/2 | 0/3 | 0/3 | 3/3 | Softened |
| 2.6 Spicy · image | 3/3 | 3/3 | 3/3 | 2/3 | 4/4 | 3/3 | Uncensored |
| 2.6 Flash · image | 3/3 | 3/3 | 3/3 | 3/3 | 4/4 | 3/3 | Uncensored |
| 2.6 · text | – | – | – | 1/1 | 2/2 | – | Uncensored |
| 2.6 · image | – | – | – | – | 1/1 | – | Uncensored |
| 2.6 · reference | – | – | refused 0/1 | – | – | – | Filtered |
| 2.5 · image | 3/3 | 3/3 | 3/3 | 3/3 | 4/4 | 3/3 | Uncensored |
| 2.5 · text | 3/3 | 3/3 | 3/3 | 3/3 | 3/5 | 3/3 | Uncensored |
| 2.2 Animate · character animation | 3/3 | 2/3 | 3/3 | 3/3 | 3/4 | 3/3 | Uncensored |
| 2.2 Spicy · image | 3/3 | 3/3 | 3/3 | 3/3 | 1/4 | 0/3 | Uncensored |
| 2.2 Spicy LoRA · image, no LoRA loaded | 3/3 | 3/3 | 3/3 | 0/3 | 2/4 | 3/3 | Uncensored |
| 2.2 Spicy LoRA · video-extend | 1/1 | 1/1 | 1/1 | 1/1 | 2/2 | 1/1 | Uncensored |
| 2.2 LoRA · image, no LoRA loaded | 3/3 | 3/3 | 3/3 | 1/3 | 3/4 | 1/3 | Uncensored |
| 2.2 · image | 3/3 | 3/3 | 3/3 | 0/3 | 1/4 | 3/3 | Softened |
| 2.2 · text | refused 0/3 | refused 0/3 | refused 0/3 | refused 0/3 | refused 0/4 | refused 0/3 | Filtered |
Uncensored means we add no content filter on our side; the model can still tone a scene down, as the bold cells show. Softened succeeds and bills but tones the result down. Filtered refuses outright. Softened clips are billed; a refused job fails with the error code content_rejected and normally costs nothing. In the 3.0 family, Prime is listed as a faster tier at 36% more per second at 480p and 40% more at 720p or 1080p (worth it only if turnaround matters; on Spicy builds we did not see it) and Pro lifts resolution to 1080p, 2K or 4K.
What the grid shows:
- Text is where Wan splits. Wan 2.2 text-to-video refuses everything. Wan 2.5 and 3.0 text-to-video delivered at every level; base Wan 2.6 delivered the L4 and L5a prompts we ran. Wan 2.5 text-to-video occasionally tones L5a down, so rerun before you write an endpoint off. Wan 3.0 text-to-video generates up to 30 seconds with sound, so the same behaviour carries over to longer takes.
- From a still, the top of the ladder is where Wan 2.2 gives out. Base 2.2 image-to-video keeps L1 to L3 but tones L4 down. The Spicy build fixes L4 but not the L5 probes, and it is not a superset of the base model: base 2.2 delivered L5b where Spicy kept the aftermath but left the key moment out of frame. The two LoRA builds, run with nothing loaded, sit in between.
- Wan 3.0 Prime image-to-video is the odd one out of the 3.0 family: it swaps the requested L4 scene for something tamer. Wan 3.0 Prime Spicy delivered it.
- Wan 2.7 stills split by task. Text-to-image softened every L4 and L5a request on both base and Pro; base delivered one of two L3 images, while Pro softened both. Both passed L5b, the separate non-adult probe. The edit endpoint delivers L1 and L2 every time but only some L4 and L5a prompts.
Other sites that sell the same Wan version behave differently, and they change, so rerun your own prompt on any endpoint before you commit a batch to it.
Illustration, not a model output: the same model gives different results depending on what you feed it. On Wan, text, a still and reference footage are three different doors.
What the clips look like
Asking versus not asking. Wan 2.6 text-to-video held an L4 film-script scene for the full 5 seconds, with no cutaway and no change of subject, and stays much closer to what you wrote than Wan 3.0, which escalates and adds people. Wan 2.7 Spicy, 2.6 Spicy, 2.2 Spicy and Spicy LoRA video-extend also sometimes go beyond the prompt.
The prompt beats the still. If the still shows one person and the prompt names a partner, Wan 2.2 Spicy writes the partner in. All four Wan 3.0 tiers opened on a still of a woman on a white bed, then cut to the red room the prompt described. Describe the frame you gave it, not another scene.
What Softened looks like. Wan 2.5 text-to-video on an L5a prompt gave, in two runs, first a clothed couple kissing, then, with a new seed and new wording, a man in a shirt and no woman at all. Both finished normally and were billed.
Wan 2.5 text-to-video, L5a prompt, two runs, five frames each, start to end. Unedited.
Base Wan 2.2 image-to-video keeps an L3 still as it is, then stops short at L4 (toned down or only implied). Wan 2.2 Animate can dress the partner depending on the driving clip. Wan 2.7 text-to-image comes back as a tamer image, billed as usual.
Refused. Wan 2.2 text-to-video refuses every prompt, down to L1, and Wan 2.6 reference-to-video refuses L3 reference footage. Wan 3.0 reference-to-video, otherwise open, also turns away photos of real people with a face check ("The input content is suspected to include real human faces"); recurring characters need illustrated, 3D or generated faces. Refusals normally cost nothing.
The everyday look. Wan 3.0 image-to-video, 720p with sound, starts from a back view and turns the woman toward the camera, as the prompt asked.
Wan 2.6 Spicy at 720p with sound, from a still of a woman in a black swimsuit at a rooftop pool:
And Wan 2.2 Spicy, silent and portrait (784x1176, 5.8 seconds), with a fully clothed prompt, to show that the Spicy builds make ordinary clips too:
How to write Wan prompts
- Image-to-video: the still carries the content, the prompt carries the motion. Body, wardrobe, set and light come from the frame. Write the action, the pace and the camera. Re-describe the outfit and Wan redraws it; describe another room and Wan cuts to it.
- Text-to-video: name everything, in order. Who, what they're wearing, one action, one camera move, light, sound. Wan 2.5, 2.6, 2.7 Spicy and 3.0 make sound, so give them some; Wan 2.2 and Wan 2.6 Flash are silent.
- Negative prompts exist on some versions only. Wan 2.5, 2.6 and 2.7 Spicy have a negative-prompt field, and so does base Wan 2.2 image-to-video. Wan 3.0 and Wan 2.2 Spicy do not, so say what you want instead of what you don't ("one continuous take", not "no cuts").
1. Solo, from a still, at real speed (Wan 2.2 Spicy)
She pushes herself up from the pillows, sweeps her hair back with one hand, turns her head to the camera and exhales, then stretches both arms above her head. Real-time speed, every movement finished inside five seconds. Locked camera, warm bedside lamp.
Run this prompt on Wan 2.2 Spicy →
Targets slow motion. One small action stretched over five seconds plays slow, and "slowly" asks for it. Give it three or four beats and say "real-time speed".
Weak: "Woman on a bed moving slowly." Wan picks the body, the room and the camera, and you asked for slow. Strong: the prompt above: the same idea with every choice yours.
2. Keep the camera still (Wan 2.6 Spicy)
Static tripod shot: the framing stays exactly as in the first frame for the whole clip, head to knees in view. She rolls onto her side toward the camera, props her head on one hand and holds the look. Soft window light, quiet room tone.
Run this prompt on Wan 2.6 Spicy →
Targets the camera that creeps in. State the framing and say it doesn't change. On Wan 2.5, 2.6, 2.6 Flash, 2.6 Spicy and 2.7 Spicy, also put "zoom, push-in, close-up" in the negative prompt box.
3. Stay on your still (Wan 3.0)
Same woman, same bed, same room and light as the first frame, one continuous take with no cut. She turns her head toward the camera, draws one knee up and runs a hand through her hair.
Run this prompt on Wan 3.0 image-to-video →
Targets the clip that leaves your image. Describe the frame you gave it, then the motion; never a different room. Wan 3.0 image-to-video also runs with no prompt at all and simply animates the frame. For choosing between text, image and reference input on Wan 3.0, see our Wan 3.0 workflow guide.
4. Add a LoRA without losing the frame (Wan 2.2 Spicy LoRA)
She turns her shoulders toward the camera, lets her hair fall forward, then lifts her eyes to the lens. One continuous move, framing unchanged.
Run this prompt on Wan 2.2 Spicy LoRA →
Targets the LoRA that repaints your frame. Wan 2.2 renders in two passes: the early (high-noise) pass sets composition and motion, the late (low-noise) pass sets detail and texture. Follow the adapter author's high-/low-noise compatibility instructions rather than choosing a slot by its style alone; add one LoRA at a time, don't also name its style in the prompt, and test strength at 480p. The high-noise vs low-noise slot guide goes deeper on slot choice.
5. Two adults from text alone (Wan 2.6)
Two adults, a man and a woman, on a white hotel bed at night. She sits facing him, arms around his neck, and they share a slow kiss. One slow orbit around the bed. Warm bedside lamp, city lights through the window, quiet breathing and distant traffic.
Run this prompt on Wan 2.6 text-to-video →
Wan 2.6 stays close to what you write. Say "two adults", say who is who, one shared action, one camera move. On Wan 3.0 the same prompt starts at $0.225 for 5 seconds (480p), and Wan 3.0 fills every gap you leave.
6. Make the still first (Wan 2.7 edit)
The first image is the woman. Place her on a rumpled bed in a dim hotel room lit by one bedside lamp, sitting against the headboard and looking at the camera. Keep her face and body exactly as in the reference.
Run this prompt on Wan 2.7 edit →
The way around every text path that refuses or rewrites. Wan 2.7 edit delivers L1 and L2 reliably at $0.03 an image (Pro $0.075) but only some L4 and L5a prompts. Reroll for cents, check every image, then animate the keeper with prompt 1 or 3. For image generators that go further, see our NSFW image generator comparison, and for the editor itself the Wan 2.7 editing guide.
What gets toned down or refused, and what gets through
| This gets toned down or refused | This gets through |
|---|---|
| L4 from a still on base Wan 2.2 or Wan 3.0 Prime | Wan 2.2 Spicy or Wan 3.0 image-to-video from the same still |
| L5a or L5b on Wan 2.2 Spicy | Wan 2.5 image-to-video, Wan 2.6 Flash, Wan 2.7 Spicy, or Wan 3.0 |
| An L3 pose or L4 scene on Wan 2.7 text-to-image: comes back tamer, billed | Wan 2.7 edit from a reference for L1 and L2; for stronger levels expect rerolls |
| Any prompt on Wan 2.2 text-to-video, even L1: refused | Wan 2.2 Spicy from a still, or Wan 2.5, 2.6 or 3.0 text-to-video |
| L3 reference footage on Wan 2.6 reference-to-video: refused | Wan 3.0 reference-to-video with L3 reference images |
Where Wan trips up, and the fix
- Wan 3.0 goes past where you stopped. Name the one action you want and nothing vaguer. If it still escalates, switch to Wan 2.6, which stays closer to the script.
- The newest Spicy builds change the wardrobe. Wan 2.6 Spicy also sometimes shows more than the prompt asked for. Naming each garment and saying it stays on helps a little. Where the outfit has to survive, use base Wan 3.0, 2.5 or 2.6 and check every clip.
- Wan 2.7 edit tones the second person down. On L4 edits the partner usually comes back clothed. Reroll, or build the pose from a solo frame.
- Wan 2.6 multi-shot cuts away by design. Use a single take when continuity matters. Our Wan 2.6 reference guide covers the reference task.
- Wan 2.2 stops at 8 seconds. Wan 2.2 Spicy LoRA video-extend adds 5 or 8 more, or move to Wan 2.6 (15 s) or Wan 3.0 (30 s). The longer uncensored video guide covers chaining clips.
Which Wan should you use?
| You want | Use | Delivered | 5 s from | |
|---|---|---|---|---|
| The cheapest clip from a still, up to the two-adult scene | Wan 2.2 Spicy | L1 to L4; L5 mostly not | $0.095 (480p) | Try |
| The cheapest still-to-video where every level rendered, no sound | Wan 2.6 Flash | every level | $0.11 (720p) | Try |
| A still with sound, 5 or 10 s, every level | Wan 2.5 image-to-video | every level | $0.225 (480p) | Try |
| A still, any length from 2 to 15 s, with sound | Wan 2.7 Spicy | every level | $0.62 (720p) | Try |
| Text, close to what you wrote | Wan 2.6 text-to-video | L4, L5a | $0.45 (720p) | Try |
| Text, with sound, cheapest | Wan 2.5 text-to-video | every level; L5a occasionally toned down | $0.225 (480p) | Try |
| Text, up to 30 s with sound | Wan 3.0 text-to-video | every level | $0.225 (480p) | Try |
| A still, up to 30 s with sound, every level | Wan 3.0 image-to-video | every level | $0.225 (480p) | Try |
| 1080p to 4K | Wan 3.0 Pro (cheaper than base 3.0 at 1080p) | L1 to L5a | $0.72 (1080p) | Try |
| A still, and you want the model to lean adult | Wan 3.0 Spicy | every level; clothing often changed | $0.285 (480p) | Try |
| Your own LoRA files | Wan 2.2 LoRA or Spicy LoRA | no LoRA loaded: L4 mostly toned down | $0.10 (480p) | Try |
| Continue a clip | Wan 2.2 Spicy LoRA video-extend | every level | $0.19 per 5 s added (480p) | Try |
| Your character doing someone else's moves | Wan 2.2 Animate | nearly every level | $0.265 (480p) | Try |
| The still itself | Wan 2.7 edit | L1 and L2; L4 and L5a partly | $0.03 per image | Try |
One line: from a still, Wan 2.2 Spicy if L4 is as far as you go and every cent counts, Wan 2.5, Wan 2.6 Flash or base Wan 3.0 if every level has to render; from text, Wan 3.0, or Wan 2.6 when you want it to stay closer to your script. For uncensored work, skip base Wan 2.2 on both tasks, Wan 2.6 reference-to-video and Wan 2.7 text-to-image.
Is Spicy worth it? It depends on the version.
- Wan 2.2: yes, up to L4. Base image-to-video is $0.24 for a fixed, silent 5-second clip and tones L4 down. Wan 2.2 Spicy is $0.095, adds 8-second clips and a closing frame, and delivers L1 to L4. Neither build is the answer for L5.
- Wan 2.6: only if you want the lean. Wan 2.6 Flash delivers every level for $0.11 (silent); Wan 2.6 Spicy costs $0.475 and sometimes shows more than asked. Wan 2.7 Spicy is the only Wan 2.7 that makes video, and it delivers every level.
- Wan 3.0: not for the ladder. Base 3.0 delivers every level at $0.045 a second; Wan 3.0 Spicy ($0.057) adds the wardrobe drift above. The one gap Spicy closes is Prime image-to-video's L4: Prime Spicy delivers it, but at $0.09 a second against $0.045 for base 3.0 image-to-video, which also delivers L4, and it was no faster than Wan 3.0 Spicy in our first runs (126 to 148 seconds against 80 to 105 for 5-second clips). Start with base Wan 3.0 image-to-video.
Is LoRA worth it? Only with your own files. With nothing loaded, Wan 2.2 LoRA and Wan 2.2 Spicy LoRA gave frame-identical clips from the same still and prompt and neither beat plain Wan 2.2 Spicy on L4, which is also cheaper ($0.095 against $0.10 with the current offer). Spicy LoRA's extra is video-extend. Open weights exist for Wan 2.2 but not for Wan 2.5, 2.6, 2.7 or 3.0, which is why 2.2 is the one with community LoRAs; running your own LoRA on a hosted API covers hosting them.
Wan 3.0 is not the only family that delivered every level: see the MiniMax H3 uncensored test and every Seedance version, tested.
What a 5-second clip costs
| Wan | 480p | 720p | 1080p | 2K | 4K |
|---|---|---|---|---|---|
| 2.2, per fixed 5 s clip | $0.24 | $0.48 | — | — | — |
| 2.2 Spicy | $0.095 | $0.19 | — | — | — |
| 2.2 LoRA / Spicy LoRA | $0.10 | $0.20 | — | — | — |
| 2.2 Spicy LoRA extend, per 5 s added | $0.19 | $0.38 | — | — | — |
| 2.5 | $0.225 | $0.45 | $0.675 | — | — |
| 2.6 | — | $0.45 | $0.675 | — | — |
| 2.6 Flash | — | $0.11 | $0.17 | — | — |
| 2.6 Spicy | — | $0.475 | $0.71 | — | — |
| 2.7 Spicy | — | $0.62 | $0.95 | — | — |
| 3.0 | $0.225 | $0.45 | $0.90 | — | — |
| 3.0 Spicy | $0.285 | $0.57 | $1.14 | — | — |
| 3.0 Prime | $0.31 | $0.63 | $1.26 | — | — |
| 3.0 Prime Spicy | $0.45 | $0.90 | $1.80 | — | — |
| 3.0 Pro | — | — | $0.72 | $0.80 | $0.92 |
| 3.0 Pro Prime | — | — | $1.17 | $1.26 | $1.40 |
The LoRA prices include the current 15% "LoRA price drop" offer ($0.12 and $0.24 without it); video-extend has no offer. Wan 2.2 Animate bills by the length of the driving clip: $0.265 for 5 seconds at 480p, $0.55 at 720p. Wan 2.7 edit is $0.03 per image, Wan 2.7 Pro $0.075. Prices are per-second rates times five; each model page shows the live estimate before you submit.
How to use Wan: Playground first, then the API
In the Playground. Create an account, open the model's task page and stay on the Playground tab. For Wan 2.2 Spicy image-to-video:
- Drop your still into Input image (JPG, PNG, WEBP or GIF, up to 10 MB). It sets the body, wardrobe, set, light and aspect ratio.
- Paste one of the prompts above into Prompt. Describe motion and camera only.
- Set Duration (5 or 8 seconds) and Resolution (480p or 720p). Both change the price; draft at 480p.
- Optionally add a Closing frame to pin the ending, or a Seed to compare settings on the same still.
- Run it. The price shows before anything is charged.
Fields differ by Wan version and task, so the form on each page is the one to follow: Wan 3.0 text-to-video starts from the prompt alone, and the 3.0 and 2.6 reference tasks take reference images or reference video.
Then the API, once a prompt works and you want it in your own app. The same Wan 2.2 Spicy call, with your own still:
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "alibaba/wan-2.2-spicy/image-to-video",
"input": {
"image_url": "https://cdn.example.com/your-still.png",
"prompt": "She pushes herself up from the pillows, sweeps her hair back with one hand, turns her head to the camera and exhales, then stretches both arms above her head. Real-time speed, every movement finished inside five seconds. Locked camera, warm bedside lamp.",
"duration_seconds": 5,
"resolution": "480p"
}
}'Fields, polling, downloads and errors are in the API reference. A refused prompt fails with content_rejected: change the prompt or the model rather than retrying it unchanged.
Or let an Agent run it. An agent can make the same call through the CLI, the MCP server or the Agent Skill. All three read the live model schema, and CLI and MCP show the price and ask you to confirm before creating a task. With SPICY_API_KEY set:
# CLI
npx @spicyapi/cli tasks create --model alibaba/wan-2.2-spicy/image-to-video \
--input-json '{"image_url": "https://cdn.example.com/your-still.png", "prompt": "She turns her head to the camera and exhales. Locked camera, warm bedside lamp.", "duration_seconds": 5, "resolution": "480p"}' --wait
# MCP server: the command your assistant's MCP config runs
npx --yes --package=@spicyapi/mcp spicyapi-mcp
# Agent Skill
npx skills add spicyapi-ai/spicy-skillFAQ
Is Wan 2.2 uncensored?
Partly. Base Wan 2.2 image-to-video keeps L1 to L3 but is Softened: it tones L4 down. Wan 2.2 Spicy delivers L1 to L4; it matched one of four L5a clips and softened all three L5b probes; the LoRA builds and Wan 2.2 Animate are Uncensored too, with the misses in the grid. Wan 2.2 text-to-video is Filtered: it refuses everything, L1 included.
Is Wan 3.0 uncensored?
Yes: base Wan 3.0 delivered every level on text, image and reference input, and often adds more than you asked for. Wan 3.0 Prime image-to-video is the exception (L4 toned down), and the Spicy builds deliver every level but change the wardrobe.
Can Wan 2.6 do NSFW?
Yes. Base Wan 2.6 delivered L4 and L5a from text and L5a from a still. Wan 2.6 Flash delivered every level; Wan 2.6 Spicy missed one L4 clip. Wan 2.6 reference-to-video is Filtered and refuses L3 reference footage. Where you run it matters as much as the version; other sites differ.
Why does Wan refuse my text-to-video when image-to-video works?
On Wan 2.2 the text path is Filtered: it refuses everything, L1 included. Wan 2.5, 2.6 and 3.0 text-to-video deliver (Wan 2.5 occasionally tones L5a down, so rerun before you switch). Or make the still first (prompt 6) and animate it on Wan 2.2 Spicy.
Why are my Wan videos in slow motion?
Too little happening and too many pace words. Give it three or four actions, write "real-time speed" and delete "slowly". See prompt 1.
Is Wan open source, and can I run it uncensored on my own GPU?
The official Wan-AI organisation on Hugging Face lists open weights for Wan 2.1 and 2.2 (and Animate and Dancer variants) but none for Wan 2.5, 2.6, 2.7 or 3.0, which run only through hosted APIs. Running Wan 2.2 weights yourself means no hosted filter in front of them.
Is the Spicy version worse for everyday SFW videos?
It makes normal clips too: the gym clip above is Wan 2.2 Spicy with a fully clothed prompt. But the newer Spicy builds lean adult on their own and often change clothing, so for wardrobe-sensitive work use the base version.
Do I pay for refused or softened clips?
Softened: yes, at the normal price. Refused: normally no, because the failed task releases the hold in full. That is why you check the clip, not the status: a toned-down clip is a success as far as the API is concerned.
Your part
We add no extra content filter of our own. Everyone on screen must be an adult. Any real, identifiable person must have consented, including through a still or a LoRA trained on their likeness. You need the rights to everything you upload, and you're responsible for the law where you publish. SpicyAPI is not available in every country; check the availability list in our content and usage policy before you build on it.
Run your first clip
Open Wan 2.2 Spicy in the Playground, upload a still and paste prompt 1. Draft at 480p for $0.095, then step up to 720p, or move to Wan 3.0 image-to-video once the motion is right and you need every level or up to 30 seconds.
