Run models

Best NSFW image-to-video AI models, tested

By the SpicyAPI team · Updated · Test data as of · How we test

For NSFW image to video, Wan 2.6 Flash and Wan 2.7 Spicy delivered every test run we gave them, and Wan 3.0 adds the strongest overall quality. All eight models here animate a still you supply; the table shows how far each one went and how the blind test rated its output.

Quick picks

#ModelBest forRuns deliveredBlind testSpicy IndexFromLink
1Wan 3.0Best overall51/5170.987.7$0.045/sec at 480pRun
2Seedance 2.5Best for reference stills45/4665.882.2$0.074/sec at 480pRun
3MiniMax H3 LoRABest for a consistent character58/6068.779.6$0.05/sec at 480pRun
4Wan 2.6 FlashMost permissive19/1969.278.1$0.022/sec at 720pRun
5Vidu Q3Best second opinion17/1967.374$0.07/sec at 540pRun
6Seedance 2.0 Mini SpicyBest value Seedance build17/1965.372.9$0.039/sec at 480pRun
7Wan 2.7 SpicyUnfiltered at every level19/196172.5$0.12/sec at 720pRun
8Wan 2.2 SpicyBest budget Wan13/1963.969.2$0.019/sec at 480pRun

Runs delivered counts test runs returned as requested across all five intensity levels; softened and refused runs are not counted. “From” is the lowest listed price on SpicyAPI, checked 2026-11-04; higher resolutions cost more. Sample sizes are small for some models — see how we test.

How we picked

Only models that take a start image were eligible, which left 62 candidates with range-test results. From those we chose models whose start-frame workflow differs in a way you would notice, such as reference stills, LoRA loading or clip extension, rather than listing five builds of the same model.

The ranking

Public sample from the Wan 3.0 page on SpicyAPI. It shows the model’s look; it is not one of the test runs counted here.
1

Best overall

Wan 3.0 by Alibaba

  • Highest Spicy Index here
  • Best blind-test score here
  • Text and image input

Highest Spicy Index of any image-to-video model and no refusals in 51 runs. It can extend the result once you are happy with the first seconds, so a short test take can grow into the final clip.

Spicy Index
87.7
Runs delivered
51/51
Blind test
70.9
From
$0.045/sec at 480p
  1. Mild 9/9 all delivered
  2. Moderate 9/9 all delivered
  3. Strong 6/6 all delivered
  4. Very strong 9/9 all delivered
  5. Maximum 18/18 all delivered

Delivered every run at all five levels. Inputs: Image to video, Reference to video, Text to video, Video extend.

Try Wan 3.0 on SpicyAPI Full test record

Public sample from the Seedance 2.5 page on SpicyAPI. It shows the model’s look; it is not one of the test runs counted here.
2

Best for reference stills

Seedance 2.5 by ByteDance

  • Text and image input
  • Small sample: 3 runs or fewer at some levels

Accepts several reference images and a start frame together. Softened one run at Very strong in our tests.

Spicy Index
82.2
Runs delivered
45/46
Blind test
65.8
From
$0.074/sec at 480p
  1. Mild 9/9 all delivered
  2. Moderate 9/9 all delivered
  3. Strong 16/16 all delivered
  4. Very strong 8/9 some held back
  5. Maximum 3/3 all delivered

Clean up to Strong; held back on some runs from Very strong. Inputs: Image to video, Reference to video, Text to video, Video extend.

Try Seedance 2.5 on SpicyAPI Full test record

Public sample from the MiniMax H3 LoRA page on SpicyAPI. It shows the model’s look; it is not one of the test runs counted here.
3

Best for a consistent character

MiniMax H3 LoRA by MiniMax

  • Loads your LoRA
  • Text and image input

Load a character or style LoRA so the subject looks the same from clip to clip. Delivered 58 of 60 test runs.

Spicy Index
79.6
Runs delivered
58/60
Blind test
68.7
From
$0.05/sec at 480p
  1. Mild 9/9 all delivered
  2. Moderate 9/9 all delivered
  3. Strong 11/11 all delivered
  4. Very strong 11/13 some held back
  5. Maximum 18/18 all delivered

Clean up to Strong; held back on some runs from Very strong. Inputs: Image to video, Reference to video, Text to video, Video extend.

Try MiniMax H3 LoRA on SpicyAPI Full test record

Public sample from the Wan 2.6 Flash page on SpicyAPI. It shows the model’s look; it is not one of the test runs counted here.
4

Most permissive

Wan 2.6 Flash by Alibaba

  • No refusals in testing
  • Small sample: 3 runs or fewer at some levels

No refusals and no softened runs at any level. It also accepts reference images alongside the start frame, which helps when a character has to match an earlier shot.

Spicy Index
78.1
Runs delivered
19/19
Blind test
69.2
From
$0.022/sec at 720p
  1. Mild 3/3 all delivered
  2. Moderate 3/3 all delivered
  3. Strong 3/3 all delivered
  4. Very strong 3/3 all delivered
  5. Maximum 7/7 all delivered

Delivered every run at all five levels. Inputs: Image to video, Reference to video.

Try Wan 2.6 Flash on SpicyAPI Full test record

Public sample from the Vidu Q3 page on SpicyAPI. It shows the model’s look; it is not one of the test runs counted here.
5

Best second opinion

Vidu Q3 by Vidu

  • Small sample: 3 runs or fewer at some levels

A different model family from the Wan picks, so it is worth a second take from the same still when a Wan clip misses. Delivered 17 of 19 runs; the other two were softened, not refused.

Spicy Index
74
Runs delivered
17/19
Blind test
67.3
From
$0.07/sec at 540p
  1. Mild 3/3 all delivered
  2. Moderate 3/3 all delivered
  3. Strong 3/3 all delivered
  4. Very strong 1/3 some held back
  5. Maximum 7/7 all delivered

Clean up to Strong; held back on some runs from Very strong. Inputs: Image to video.

Try Vidu Q3 on SpicyAPI Full test record

Public sample from the Seedance 2.0 Mini Spicy page on SpicyAPI. It shows the model’s look; it is not one of the test runs counted here.
6

Best value Seedance build

Seedance 2.0 Mini Spicy by ByteDance

  • Small sample: 3 runs or fewer at some levels

The Spicy build of Seedance 2.0 Mini. It delivered 17 of 19 test runs and can extend the clip it makes.

Spicy Index
72.9
Runs delivered
17/19
Blind test
65.3
From
$0.039/sec at 480p
  1. Mild 3/3 all delivered
  2. Moderate 3/3 all delivered
  3. Strong 3/3 all delivered
  4. Very strong 1/3 some held back
  5. Maximum 7/7 all delivered

Clean up to Strong; held back on some runs from Very strong. Inputs: Image to video, Video extend.

Try Seedance 2.0 Mini Spicy on SpicyAPI Full test record

Public sample from the Wan 2.7 Spicy page on SpicyAPI. It shows the model’s look; it is not one of the test runs counted here.
7

Unfiltered at every level

Wan 2.7 Spicy by Alibaba

  • No refusals in testing
  • Small sample: 3 runs or fewer at some levels

Delivered every run. Its blind-test score is well below Wan 3.0, so it is the choice when range matters more than polish.

Spicy Index
72.5
Runs delivered
19/19
Blind test
61
From
$0.12/sec at 720p
  1. Mild 3/3 all delivered
  2. Moderate 3/3 all delivered
  3. Strong 3/3 all delivered
  4. Very strong 3/3 all delivered
  5. Maximum 7/7 all delivered

Delivered every run at all five levels. Inputs: Image to video, Video extend.

Try Wan 2.7 Spicy on SpicyAPI Full test record

Public sample from the Wan 2.2 Spicy page on SpicyAPI. It shows the model’s look; it is not one of the test runs counted here.
8

Best budget Wan

Wan 2.2 Spicy by Alibaba

  • Small sample: 3 runs or fewer at some levels

An older Wan generation that still delivered most runs. Useful for testing a start image before committing to a newer model.

Spicy Index
69.2
Runs delivered
13/19
Blind test
63.9
From
$0.019/sec at 480p
  1. Mild 3/3 all delivered
  2. Moderate 3/3 all delivered
  3. Strong 3/3 all delivered
  4. Very strong 3/3 all delivered
  5. Maximum 1/7 some held back

Clean up to Very strong; held back on some runs from Maximum. Inputs: Image to video, Video extend.

Try Wan 2.2 Spicy on SpicyAPI Full test record

Before you upload a start image

Image-to-video models treat the first frame as a contract: clothing, lighting and framing in the still largely decide what the clip can become. A model that delivered a level in our text tests may still hold back if the start image itself trips its input check, so the results here are a guide to how permissive each model is, not a promise for every image.

Use images you created or have the rights to. Photos of real, identifiable people are not allowed on any of these models, whatever the setting.

Picking between motion and range

For video, the blind-test score weighs motion and consistency over time alongside prompt adherence and visual quality, judged without the model name. The delivery count measures how often the model went where the prompt asked. The two rarely peak in the same model, which is why the list mixes polished picks with fully permissive ones.

Well-known models that held back

These three accept a start image and are popular for general animation work. With this test set they held back on most runs, which is why none of them made the eight picks.

  • Kling 4.0 by Kling

    3 of 8 runs delivered. Clean up to Mild; blocked from Moderate. Image to video, Reference to video, Text to video.

    1. 2/2
    2. 0/1
    3. 0/1
    4. 0/1
    5. 1/3
  • PixVerse V6 by PixVerse

    1 of 3 runs delivered. Clean up to Mild; blocked from Moderate. Image to video, Reference to video, Text to video.

    1. 1/1
    2. 0/1
    3. 0/1
    4. —
    5. —
  • P-Video 2 Pro by Pruna AI

    17 of 23 runs delivered. Clean up to Mild; held back on some runs from Moderate. Image to video, Reference to video, Talking avatar.

    1. 4/4
    2. 2/3
    3. 3/3
    4. 2/5
    5. 6/8

Every scored model is in the filter tracker.

Questions

What is the best NSFW image to video AI?

Wan 3.0 has the best overall score. If you need every run delivered, Wan 2.6 Flash and Wan 2.7 Spicy had no refusals in our tests.

Why did my image get blocked even on a permissive model?

Models check the input image as well as the prompt. Images that appear to show real people or minors are blocked everywhere, and some models also screen start images more strictly than prompts.

Can I keep the same character across clips?

Use a model with LoRA loading, such as MiniMax H3 LoRA, or one that accepts reference images, such as Wan 2.6 Flash or Seedance 2.5.

Open the model picker →