Wan Video LoRA NSFW: Which Wan Builds Take LoRAs, and Which Are Checkpoints

You found a Wan video LoRA NSFW download on Civitai or Hugging Face and want to know where it can run without installing ComfyUI or owning a high-end graphics card. Short answer: a hosted API can load a LoRA, but not the full merged Wan checkpoints the community passes around, and if all you want is a model with no content filter, you may not need a LoRA at all.

• Where a LoRA loads: Wan 2.2 LoRA, Wan 2.2 Spicy LoRA, Wan 2.2 5B LoRA and Wan 2.1 LoRA. No other Wan video build has a LoRA field.
• What does not load: the three most-searched community builds are full model checkpoints of 9 to 27 GB per file, not LoRAs.
• If you only want no filter: Wan 2.2 Spicy ($0.095 for 5 seconds at 480p) or Wan 3.0 ($0.225) need no LoRA.

We ran six same-seed clips for this guide: one still, one prompt, one seed, one change per run. The test prompts are not published, and the character is a fictional adult. Prices were checked on 21 October 2026.

The short answer

LoRA or checkpoint? What Wan 2.2 Enhanced NSFW, Wan 2.2 Remix and Rapid AIO really are

A LoRA is a small add-on file, usually 150 MB to 1 GB, that is attached to the base model when a clip is generated. A checkpoint is the whole model in one file. The three names that dominate the searches are checkpoints. We read their public Civitai and Hugging Face pages on 21 October 2026 and downloaded nothing.

A hosted API cannot swap its base model for yours. Our LoRA fields take weight files only, up to 1 GB each on image-to-video, and a 9 to 27 GB file is a different kind of object, not a bigger LoRA. So these builds cannot run here, and we would rather say that than send you to a workaround that does not exist.

Two routes get you close to what they give you:

1. Use a build that has no filter to begin with. Wan 2.2 Spicy and Wan 3.0 delivered every request in our Mild to Very strong tests without a LoRA.
2. Load a general-purpose unfiltered LoRA pair on Wan 2.2 LoRA or Wan 2.2 Spicy LoRA. Our LoRA guide for Wan 2.2 lists the community files we have checked and whether they run.

You do not need to add a speed LoRA either. We bill per clip, so a speed LoRA can only shorten the wait.

Which Wan endpoints take a LoRA

Four families take a LoRA on video, and two more take one on a single still. Each LoRA is a direct link to a .safetensors file, with a strength (scale) from 0 to 4 and a default of 1. Loading one costs nothing extra.

Plain Wan 2.2, Wan 2.2 Spicy, 2.5, 2.6, 2.7, Animate and 3.0 have no LoRA field. The $0.102 price includes a 15% discount on Wan 2.2 LoRA image-to-video and Spicy LoRA image-to-video that is running now; the regular price is $0.12.

High-noise, low-noise or both

Wan 2.2 generates a clip in two passes. The high-noise pass lays out motion and composition; the low-noise pass refines detail. A LoRA published as two files, one marked high noise and one marked low noise, goes into two slots: the high file in high_noise_loras, the low file in low_noise_loras. A single file goes in loras, which applies it to both passes.

Our high-noise and low-noise slot guide shows what each slot changes in a four-clip test.

Will a LoRA made for another Wan version work?

The Wan 2.1 to Wan 2.2 row is the one we had not tested until this guide. We loaded one Wan 2.1 camera-orbit LoRA (an SFW motion file) on Wan 2.2 LoRA image-to-video at 480p for 5 seconds, with the same still, prompt and seed in every run, and measured how far each clip's frames differed from the no-LoRA run. Running the no-LoRA clip twice gave a difference of 0.4%, so that is the noise floor.

• In loras or high_noise_loras: the clip changed measurably, from about 1% at the start to about 7% by the end. The orbit was not visibly wider than the prompt alone already produced.
• In low_noise_loras: about 1.75% at most, four times the noise floor and not enough to see. Motion LoRAs do little in the detail pass.

So a Wan 2.1 file will load on Wan 2.2 and not fail, but treat it as unpredictable, and look for a Wan 2.2 version of the same LoRA first. This is one file and one test, so read it as a direction, not a rule. The same file did what its author promised on Wan 2.1:

To find out which base a file was trained for, use the LoRA compatibility checker.

What a LoRA adds when the model already has no filter

Less than you might think, and something different from "unlocking".

• It redraws the clip. On Wan 2.2 LoRA image-to-video, the base model had already delivered our Strong-level request as written. With a general-purpose unfiltered LoRA pair loaded in the two noise slots, the same prompt, still and seed came back redrawn from the first second: a different camera path, framing and rendering of the person. Frame differences from the no-LoRA run were 6% at half a second and 11 to 13% from 2.5 seconds on.
• It can lift a toned-down result. In an earlier same-seed test on 28 September, the base Wan 2.2 LoRA endpoint toned a Very strong request down twice; with the same LoRA pair loaded, both runs came back as written. That is two runs against two, a small sample. On Wan 2.1 video-to-video a general LoRA turned one toned-down Strong request into one delivered as written, which our Wan 2.1 guide covers.
• It adds a camera move or a look. A camera-follow pair, a vintage-film style LoRA and a Technicolor pair all have examples on their model pages.
• It does not give you a character. We did not test a general character LoRA, and we found no public one that is worth recommending. In image-to-video the person comes from your starting frame. To keep a character across clips, train a LoRA on a fictional adult character of your own. We do not recommend LoRAs of real people.

If your goal is just no filter, skip the LoRA. Wan 2.2 Spicy and Wan 3.0 delivered our Mild to Very strong tests without one. Wan 2.2 Spicy delivered only one of four Maximum runs as written; Wan 3.0 delivered that level too. The full per-level results are in our Wan 2.2 test and the all-versions table.

How to load one on SpicyAPI

1. Open Wan 2.2 LoRA image-to-video in the Playground and paste the link to the file into the LoRA row. Use the file's direct download link, or a Hugging Face file page.
2. For a pair, put the high-noise file in one slot and the low-noise file in the other. For a single file, use the general slot.
3. Pick 480p and 5 seconds for your first run; it is the cheapest.

Through the API, the request looks like this. The paths are placeholders:

{
  "model": "alibaba/wan-2.2-lora/image-to-video",
  "input": {
    "image_url": "https://example.com/opening-frame.jpg",
    "prompt": "She turns from the balcony rail to face the camera as the lanterns sway behind her; slow push-in, warm light.",
    "resolution": "480p",
    "duration_seconds": 5,
    "seed": 20261021,
    "high_noise_loras": [{ "path": "https://huggingface.co/<owner>/<repo>/resolve/main/<file>_high_noise.safetensors", "scale": 1 }],
    "low_noise_loras":  [{ "path": "https://huggingface.co/<owner>/<repo>/resolve/main/<file>_low_noise.safetensors",  "scale": 1 }]
  }
}

If a file sits behind a Civitai or Hugging Face sign-in, import it once with your own Hugging Face token or Civitai API key. The token is used for that one download and never stored. The imported file is kept as one of your uploads for the usual upload retention period, and you put the spicy:// address it returns in path. We did not run the import endpoint for this guide, so check the API reference for its limits. The full request flow, with polling and preflight checks, is in Run your own LoRA through the API.

A link that cannot be downloaded fails the run, and a failed download is not charged.

What it costs

Prices checked on 21 October 2026, for 5 seconds at 480p unless noted:

The six clips behind this guide cost $0.102 each, $0.612 in all. You can compare that with the graphics card and the 9 to 27 GB downloads a local checkpoint asks for.

Frequently asked questions

Which Wan 2.2 NSFW LoRAs work on a hosted API?

Any LoRA trained for Wan 2.2 A14B, as a direct link or an import, in the right slot: a pair into the high and low slots, a single file into loras. Files must be 1 GB or smaller on image-to-video. Our Wan 2.2 LoRA list shows the community files we have checked and whether each one runs.

Do Wan 2.1 LoRAs work on Wan 2.2?

They load without an error. In our same-seed test, a Wan 2.1 camera-orbit file changed the clip only from the high-noise slot or loras, almost not at all from the low-noise slot, and it did not widen the orbit. Look for a Wan 2.2 version first. Style and look LoRAs were not part of this test.

Do Wan 2.2 14B LoRAs work on the 5B model?

No. They load without an error and the clip stays the same. Use LoRAs trained for the 5B model on the 5B endpoint.

Can I load a merged Wan checkpoint on a hosted API?

No. The LoRA fields accept LoRA weight files only, and a merged checkpoint is the whole model at 9 to 27 GB. Pick a build with no filter, or load a LoRA on a Wan 2.2 LoRA endpoint.

Do I need both the high-noise and the low-noise file?

If the author published a pair, load both, each in its own slot. If there is one file, put it in loras.

Can I use a LoRA that needs a Civitai or Hugging Face sign-in?

Yes, through the import endpoint with your own Hugging Face token or Civitai API key. The token is not stored; the file is kept as one of your uploads.

Is there a character LoRA for Wan?

We did not test a general character LoRA. In image-to-video the person comes from your starting frame. For a recurring character, train your own LoRA on a fictional adult character.

Do I need a Lightning or speed LoRA?

No. Billing is per clip, so a speed LoRA can only shorten the wait.

Does a LoRA cost extra?

No. A failed download is not charged either.

Bottom line

• You want no filter and little fuss: start with Wan 2.2 Spicy at $0.095, or Wan 3.0 for every level with sound.
• You have a Wan 2.2 LoRA file: load it on Wan 2.2 LoRA image-to-video at 480p for $0.102, high file in the high slot, low file in the low slot.
• You only have a full merged checkpoint: it cannot run on a hosted API. Use one of the two routes in the checkpoint section, or run it locally.