Best Uncensored NSFW AI Video Generators in 2026: 8 Models Tested, MiniMax H3 and Wan 3.0 Lead

Eight video models you can call through the SpicyAPI Playground or API, ranked by what came back when we pushed each up a five-step content ladder: the clip as written, toned down, or refused. Based on our September 2026 tests.

Short answer: MiniMax H3 and Wan 3.0 lead: both delivered all the clips we ran as written. H3 is the cheapest way in from text ($0.185 per 5 seconds at 768p, against $0.225 for Wan 3.0 at 480p). Wan 3.0 has the larger record and the only 30-second take from text, but it can go beyond the prompt, so check the end of every clip.

The quick answer

• Cheapest strong clip from a text prompt? MiniMax H3, $0.185 per 5 seconds at 768p.
• Broadest test record, or a take longer than 15 seconds? Wan 3.0: all six probes delivered, and 30 seconds from text for $1.35 at 480p.
• Have stills and want them moving without writing a prompt? MiniMax H3 Spicy: prompt optional, 5 seconds from $0.19. H3 and H3 Max generate sound; H3 Spicy has no sound option: add sound in the edit if you need it. Vidu Q3 Spicy and Seedance 2.5 Spicy do the same job with a sound option, and each delivered every probe apart from one toned-down L2 clip.
• Cheapest of the eight? Wan 2.2 Spicy at $0.095 per 5 seconds, but it toned down both L5 probes. Check the grid before you pick it for anything strong.
• 4K? LTX 2.5: 5 seconds for $1.50 (Seedance 2.0 native 4K is $3.89).
• 30 seconds from a still, or clips from 1 second? Seedance 2.5 Spicy for 30 seconds ($1.08 for 5 s); Vidu Q3 Spicy for clips from 1 second ($0.0665 a second at 540p).

Try MiniMax H3 text-to-video →

How the ladder works

Five levels, each stronger than the one before; full definitions in the test methodology. All runs used consenting adult characters, mostly 1 to 5 seconds at the cheapest resolution.

• L1 Suggestive: suggestive imagery
• L2 Mature themes: moderate NSFW imagery
• L3 Strong: advanced NSFW imagery
• L4 Intense: higher-intensity adult scenes between two adults
• L5 Extreme: two probes, L5a Restraint, a consensual adult scene with restraint themes, and L5b Gore, cinematic gore

"Unasked" means a model escalated to L4 when the prompt didn't request it. Seedance 1.5 Pro Spicy, a cheaper endpoint, is outside the eight (see the end of the comparison).

How far each model goes

Each cell is clips delivered as asked out of clips run; a dash means no clip at that level.

No clip in the grid was refused outright; every miss is a toned-down clip, and toned-down clips are billed like any other. Uncensored on our model pages means no content filter is added on our side. Seedance 2.0's text-to-video also rejected most L1 attempts with a copyright error (see its section), so its L1 cell counts only the clips that came back.

What the clips look like

MiniMax H3 text-to-video, 768p. Left: the L1 prompt (a sheer slip). Right: L3, cropped to the upper body. The mosaic hides most of each frame.

On H3 the L4 clip held the scene for its whole length, with no cutaway or fade, and the L5a probe kept every element we named on screen.

Same frame, same prompt

Given one L3 frame and the L5a prompt, MiniMax H3 Spicy, Seedance 2.5 Spicy, Seedance 1.5 Pro Spicy and Seedance 2.0 Mini Spicy delivered the scene as prompted. Wan 2.2 Spicy brought in a second person and escalated to L4; Seedance 2.0 Spicy, Seedance 2.0 Fast Spicy, LTX 2.3 Spicy and Wan 2.6 and 2.7 Spicy also went to L4 unasked. Builds that look the same on a grid differ here: some stop where you stop, some keep going.

Published examples, one per model

The everyday look, one published SFW example each.

Top: MiniMax H3, H3 Spicy, Seedance 2.5 Spicy, Seedance 2.0. Bottom: Wan 3.0, Wan 2.2 Spicy, LTX 2.5, Vidu Q3 Spicy.

The full comparison

Price: one 5-second clip at the cheapest resolution (480p unless noted; H3 starts at 768p), catalogue prices. Ranked by delivery rate; models within about two points are ordered by price.

H3 first is a price call: H3 has the lowest price, Wan 3.0 the larger record. H3 Spicy, Vidu Q3 Spicy, Seedance 2.5 Spicy and LTX 2.5 are close behind; Seedance 2.0 and Wan 2.2 Spicy are clearly lower. Specs side by side: MiniMax H3 vs Wan 3.0.

1 · MiniMax H3: cheapest in from text

Strengths

• L1, L3, L4, L5a and L5b all came back as written (14 of 14), mostly from text, 2k included.
• Three ways in: a text prompt, a first-frame image, or up to nine reference images (images only, no video or audio references) to hold a character across shots.
• The cheapest text-to-video here: 5 seconds for $0.185 at 768p, $0.29 at 2k.

Trade-offs

• No seed: a re-roll gives you a different person. Fix the look with a still, then use image-to-video.
• 15 seconds and 2k at most. Longer takes: Wan 3.0. 4K: LTX 2.5.
• Sound is always on: write it into the prompt, or mute it in the edit.

Prompt to try (text-to-video). A mild, adult starting prompt:

An adult woman lies back on white hotel sheets in late-afternoon light, one arm above her head. She turns toward the camera, draws one knee up and holds the lens with a slow smile. Slow push-in from the foot of the bed. Warm window light, quiet room tone.

Run this prompt on H3 →

Full MiniMax H3 test

2 · Wan 3.0: the fullest record, and long takes

Strengths

• All six probes delivered (51 of 51 clips).
• $0.045 a second at 480p, so a full 30-second clip is $1.35 ($2.70 at 720p).
• Reference-to-video keeps the character's look, from up to 10 reference images, with up to 5 reference videos and 5 audio clips if you have them.
• Sound on or off, and a seed.

Trade-offs

• It escalates: in one round it added an unrequested L4 scene to most clips. If a scene has to stop at L3, check the end of every clip.
• 1080p at most ($0.90 for 5 seconds).

Prompt to try (text-to-video, set the length to 30 seconds):

One continuous take. Two adults, a man and a woman, in a candlelit loft at night. She sits on the edge of the bed; he kneels behind her and gathers her hair over one shoulder; she leans back against his chest. The camera circles the bed once. Warm candlelight, rain on the skylight, no music.

Run this prompt on Wan 3.0 →

Every Wan version tested

3 · MiniMax H3 Spicy: stills that move, no prompt needed

Strengths

• The prompt is optional: send a still alone and it moves. An optional closing frame pins the ending.
• Delivered all six probes except one L4 clip that only implied the scene.
• 3 to 15 seconds at 480p, 540p, 768p or 1080p. 1080p is enlarged from the native 768p, not extra detail.

Trade-offs

• First-frame image only; for text, use H3.
• No sound option: add sound in the edit if you need it. (H3 and H3 Max generate sound.)
• Pricier than H3: 5 seconds is $0.19 at 480p, $0.42 at 768p, $0.76 at 1080p.
• Without a seed, a re-run gives a new take.

Prompt to try (image-to-video, optional):

She wakes, stretches both arms over her head and rolls onto her side toward the lens, hair sliding across the pillow. Slow push-in, then hold on her eyes. Soft breathing, rain on the glass.

Run this prompt on H3 Spicy →

4 · Vidu Q3 Spicy: everything from a still, and clips from 1 second

Strengths

• Delivered all six probes except one L2 clip that came back toned down.
• Clips from 1 to 16 seconds: the shortest clips here.
• Prompt optional, sound on or off.
• 1080p costs barely more than 720p: $0.76 against $0.71 for 5 seconds. Skip 720p.

Trade-offs

• First-frame image only, and no seed.
• 5 seconds at 540p is $0.33, more than Wan 2.2 Spicy or H3 Spicy.

Prompt to try (image-to-video, optional):

She lifts her face out of her wet hair and looks straight into the lens as water runs down her throat and shoulders. Slow push-in. Dripping water, steady breathing.

Run this prompt on Vidu Q3 Spicy →

Full Vidu Q3 test

5 · LTX 2.5: 4K and sound, with a seed

Strengths

• The cheapest 4K here: 5 seconds for $1.50. A 20-second 720p take is $1.80.
• From text and from an opening frame, it delivered L1, L2, L3 and L5a; L4 in 5 of 6 clips.
• Sound is always on.
• A seed, so you can re-run a take you like.

Trade-offs

• It starts at 720p, so the cheapest 5-second clip is $0.45.
• Text-to-video only does 16:9 or 9:16.
• Anatomy can slip at the top of the ladder: malformed figures and merged limbs at L3 and L4. The cheaper LTX 2.3 Spicy ($0.095 for 5 seconds at 480p, silent) is uneven at the top: it delivered L4 in 2 of 3 clips, L5a in 1 of 3 and L5b in none.

Prompt to try (text-to-video):

An adult woman lowers herself into a candlelit copper bathtub, water rising to her shoulders as she leans back against the rim and closes her eyes. Locked camera, low angle from the foot of the tub. Candlelight on the water and the copper. Water lapping, one slow exhale.

Run this prompt on LTX 2.5 →

Full LTX test

6 · Seedance 2.5 Spicy: 30 seconds from one still

Strengths

• Up to 30 seconds per take, at 480p, 720p, 1080p or an upscaled 4K.
• Delivered all six probes except one L2 clip that came back toned down.
• Prompt and closing frame are optional; sound on or off; a seed.

Trade-offs

• The most expensive model here: $1.08 for 5 seconds at 480p, $12.96 for 30 seconds at 720p, $10.80 for 5 seconds of 4K.
• Slow: 3 to 11 minutes per clip.
• 4K is upscaled from a native 1080p render.

Prompt to try (image-to-video, set the length to 30 seconds):

One unbroken take. She rises from the bath, wraps a towel around herself, walks to the fogged mirror and wipes a clear streak with her palm, meeting her own eyes. The camera follows at shoulder height and settles behind her. Steam, dripping water, footsteps on tile.

Run this prompt on Seedance 2.5 Spicy →

7 · Seedance 2.0: native 4K from text, but it softens L4

Strengths

• Delivers L1, L2, L3 and L5a.
• Native 4K: 5 seconds for $3.89. At 480p it's $0.35, sound costs nothing extra, and there's a seed.

Trade-offs

• Asked for L4 outright, it delivered 1 clip in 9; the rest were toned down to a kiss, an embrace or an implied version. For L4 on request, use Wan 3.0 or H3.
• Its text-to-video rejected a plain L1 prompt with a copyright error 7 times in 8; the eighth came back as written. Seedance's own safety layer refuses some wording outright, kissing being the usual trigger. Re-roll, or use Seedance 2.0 Fast or 2.5.
• A safe opening frame holds image-to-video back; start from an L3 frame to go higher. An L1 frame is fine for an L2 prompt.
• Reference-to-video (up to 9 reference images, 3 videos and 3 audio clips) follows what the reference shows: a clothed reference character stays clothed.
• 15 seconds per clip.
• The image-only Seedance 2.0 Spicy ($0.57 for 5 seconds at 480p) can escalate to L4 unasked.

Prompt to try (text-to-video):

An adult woman in a black silk robe stands at a floor-to-ceiling window at night. She turns to face the camera, one hand on the glass. Slow dolly-in. Cool city glow on one side, a warm lamp on the other. Low city hum.

Run this prompt on Seedance 2.0 →

Every Seedance version tested

8 · Wan 2.2 Spicy: the cheapest clip, weak at L5

Strengths

• $0.095 for 5 seconds at 480p, $0.304 for 8 seconds at 720p.
• L1 to L4 delivered (it can also go to L4 unasked).
• A seed for repeatable drafts.

Trade-offs

• L5 is where it slips. L5a delivered once, then came back toned down in all three later runs; L5b was toned down in all three.
• Without a seed, the same frame and prompt gives the same clip. Set a different seed each time you re-roll.
• It adds people: from a solo frame it brought in a second person. Wan 3.0 Prime left the same kind of frame alone.
• Silent, 5 or 8 seconds only, 720p at most. A prompt is required.

Prompt to try (image-to-video):

She pushes herself up from the sheets onto her knees, turns her head toward the camera and pulls her hair back with both hands. Slow push-in, warm lamp light.

Run this prompt on Wan 2.2 Spicy →

Which one for which job

Run the ladder yourself

1. Draft cheap. The lowest resolution and 5 seconds cost $0.185 on H3 (768p), $0.225 on Wan 3.0 (480p) and $0.095 on Wan 2.2 Spicy (480p). Step up the resolution once the clip is right.
2. Change one thing per step. Keep camera, light and sound lines identical and change only the scene content.
3. Check the last frames. Escalation and toned-down cuts show up as the clip runs, not in the first frame.
4. Build the still first. It guides composition, but is not a universal intensity ceiling. Standard Seedance 2.0 can preserve the source's milder appearance; several Spicy routes changed or intensified it in our examples. Compose the intended scene before animating it.
5. If a clip is toned down, switch route before rewording. Every re-roll is billed: move from text to image-to-video, change the seed, or pick a model that delivered that level.
6. Keep everyone an adult, and use a real person's likeness only with their consent.

Need more than 30 seconds? Chain clips from the last frame: the longer uncensored video guide shows how.

Cheapest of all: Seedance 1.5 Pro Spicy is $0.06 for 5 silent seconds at 480p (sound doubles it). It delivered L1, L2, L3, L5a and L5b in every clip and L4 in 2 of 3.

The ones that toned it down

These endpoints carry a restricted label in our catalogue. Here is what they did with an uncensored prompt:

MiniMax H3 Max text-to-video, L4 prompt. Unedited frame.

Wan 2.5 text-to-video (all 18 prompts delivered, L4 and L5 included), Wan 2.2 Animate, and Wan 2.5 and HappyHorse 1.1 image-to-video are Uncensored, not on this list.

How to use them: Playground first, then the API

In the Playground. Open the model's task page, pick the mode (text, first-frame image or reference images), paste a prompt from the sections above and run it; the exact price shows before anything is charged. Image modes need your own image added there first. Create an account if you don't have one, and draft at the lowest resolution the model offers (768p on H3).

• Text: MiniMax H3, Wan 3.0, LTX 2.5 or Seedance 2.0
• First-frame image: MiniMax H3 Spicy, Vidu Q3 Spicy, Seedance 2.5 Spicy or Wan 2.2 Spicy
• Everything else: the uncensored video catalogue

Then the API, once a prompt works and you want it in your own app. The same call, as the H3 text-to-video request (image-only models take an image_url instead of text alone; LTX 2.5 starts at 720p):

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)" \
  --data '{
    "model": "minimax/h3/text-to-video",
    "input": {
      "prompt": "An adult woman lies back on white hotel sheets in late-afternoon light, one arm above her head. She turns toward the camera, draws one knee up and holds the lens with a slow smile. Slow push-in from the foot of the bed. Warm window light, quiet room tone.",
      "resolution": "768p",
      "duration_seconds": 5
    }
  }'

Fields, results and errors are in the API reference.

Let an Agent do 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 minimax/h3/text-to-video \
  --input-json '{"prompt": "An adult woman lies back on white hotel sheets in late-afternoon light.", "resolution": "768p", "duration_seconds": 5}' --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-skill

FAQ

What is the best uncensored AI video generator right now?

MiniMax H3 and Wan 3.0 tie on delivery. Pick H3 for the lowest price from text ($0.185 per 5-second clip); pick Wan 3.0 for the larger record and 30-second takes.

MiniMax H3 or Wan 3.0?

They tie on delivery, so decide on features. H3: the lowest price from text (18% less than Wan 3.0's 480p price), 2k and always-on sound. Wan 3.0: 30-second takes, 1080p, a seed, sound on or off, and the longer record.

Which AI video model is uncensored from a text prompt?

On our API, MiniMax H3, Wan 3.0, Seedance 2.0 and LTX 2.5 all passed our L3 and L5a brief from text alone. Wan 2.5 did too, 18 of 18. Avoid H3 Max and Wan 2.2 text-to-video: they refuse most L3 and L4 prompts. Asked for L4 from text alone, Wan 3.0 delivered 3 of 3; LTX 2.5 delivered 2 of 3 and toned down one; Seedance 2.0 toned down all 3.

Do I need an API, a browser tool, or a local setup?

All eight are in the Playground and the API: the Playground for trying a clip, the API for your own app. Local open-weight builds need your own GPU and are not part of this ranking.

What's the cheapest uncensored AI video model?

Seedance 1.5 Pro Spicy: $0.06 for 5 silent seconds at 480p (sound doubles it). Of the eight above, Wan 2.2 Spicy at $0.095 (image only; LTX 2.3 Spicy costs the same but is uneven at L4 and L5). From text, MiniMax H3 at $0.185 (768p).

Why did my video come back toned down?

Usually a Softened or Filtered endpoint, a safe opening frame on image-to-video, or a model that softens that level (Seedance 2.0 at L4, Wan 2.2 Spicy at L5). Use H3 or Wan 3.0, or start from a still that already sets the scene.

Do I pay for refused or toned-down clips?

Toned down: yes, at the normal price. Refused: the task ends as failed with content_rejected and the hold is normally released in full, unless the model page says it charges the accepted quote. Check settled and cost on the task; the policy page has the rule.

Can I use a photo of a real person?

Only with that person's documented consent, never for an intimate depiction of someone who hasn't given it.

Your part

SpicyAPI adds no platform content filter on these Uncensored routes; model or upstream checks can still apply. Everyone on screen must be an adult. Any real, identifiable person must have consented. You need the rights to everything you upload, and you're responsible for the law where you publish.