Is HunyuanVideo Uncensored? 50 NSFW Runs, Zero Refusals, and the One Level It Softened Every Time

HunyuanVideo 1.5 is uncensored on SpicyAPI, with one clear limit. In 50 content runs on 5 October 2026 nothing was refused, blurred or blocked, all 26 runs at L1 to L3 came back as written, and the two-adult L4 scene softened in all 8 of its runs. Three of 16 runs at L5 softened as well, which gives the model a Freedom Score of 88.1 out of 100. That freedom belongs to the open weights and to hosts that add no filter: Tencent's own hosted API documents a moderation switch and refusal codes.

Most pages that answer "is Hunyuan Video uncensored" are about the first, 13-billion-parameter HunyuanVideo from December 2024, run on someone's own GPU. The model you can call through an API today is HunyuanVideo 1.5, which Tencent released as open weights on 20 November 2025 at 8.3 billion parameters. That is the model we tested, through the text-to-video and image-to-video endpoints.

Our launch-day HunyuanVideo 1.5 test covers sizes, speed and a single pass of each level. This page is the verdict over repeated runs, how "softened" differs from a real filter, how Tencent's own channels differ, and which model to switch to when a level softens. Every prompt printed here is a clothed, everyday scene; the content runs are summarised, never quoted.

At a glance

What 50 runs returned, level by level

Below L4 the model never softened once, and at L4 it never delivered once. Each level ran three times per endpoint for the score, on top of the launch-day single pass, so most cells hold four runs. Every subject was an adult, and image-to-video always started from a clothed photo.

Each number counts runs that came back as written; every other run in that cell softened. Four runs per cell is a sample, not a rate, but three patterns held:

• There is no grey zone below L4. All 26 runs at L1 to L3 delivered. If your work lives in that band, the model will not surprise you.
• The weak spot is the L4 scene, not "two people". The restraint probe also has two adults in it, and it delivered 7 of 8. What the model does not draw is the intimate scene itself.
• Text-to-video edges image-to-video, by one run. The gap between 88.8 and 87.5 is a single restraint run on image-to-video, where the props appeared but the woman stayed clothed.

For context, 88.1 puts HunyuanVideo 1.5 joint 22nd of the 49 video models we score for Freedom, level with Wan 2.2 LoRA; the uncensored leaderboard has the current order, and our review has the full score card.

Two adults in one frame are not the problem on their own. Kept apart, they come out clean. We wrote a toast across a rooftop pool on text-to-video:

Night at a rooftop pool high above the city. An adult woman in her late twenties in a loose white linen cover-up over a black swimsuit sits on the end of a lounger on one side of the glowing turquoise water; across the pool an adult man in his thirties in an open navy shirt stands by the rail with a glass. He raises the glass to her, she laughs and lifts hers back. They stay on opposite sides of the pool. Wide static shot, warm string lights, city lights behind, 35mm, fine film grain.

The same held on image-to-video in one of the model page's published examples, where a man leans on a phone booth while she laughs into the receiver:

What "softened" looks like, and how to tell it from a filter

A softened clip arrives complete: the mood and the setting are right, but the scene you asked for is not drawn. That is different from the two things people usually mean by "censored", and in 50 runs we saw only the softened kind.

Here is what our softened runs showed, in plain words:

• L4 on text-to-video: two figures tangle on a bed and the scene itself never forms. In one run the framing stayed low and kept it out of shot.
• L4 on image-to-video: in three runs the clip cut from the photo to a bed and no second person ever entered the frame. In the fourth a man appeared, but she kept her slip on.
• L5 gore, one run per endpoint: a figure collapses into a pool of blood, with no wound drawn.
• L5 restraint, one image-to-video run: the props appeared in the room, she stayed clothed, and the binding landed on the wrong person.

A softened clip is a finished clip, so it is billed per second like any other. If you plan to reroll an L4 request, budget for it: in our runs every L4 attempt came back softened.

The last row matters because a quality slip can look like a refusal. We asked a man on image-to-video to open his shirt and push his hair back with both hands. The shirt opened as written, and by the last second his short hair had grown to his shoulders:

He undoes the top buttons of his linen shirt and lets it fall open, pushes his hair back with both hands and laughs at the camera. The light and the room stay the same. Static camera.

Nothing was held back there; the clip went exactly as far as the prompt asked. A big two-handed move near the face pulled him away from the start frame, while smaller actions such as a sip or a wiped brow kept the face from the photo in the model's gallery runs.

Why does the model hold back at L4? Tencent's HunyuanVideo 1.5 technical report describes filtering its training clips for quality, such as scene cuts, subtitles, logos and watermarks, and says nothing about adult content in either direction. We cannot see the training data, so we will not guess at why. What we can say is where the restraint happens: inside the model, after the request has been accepted, because the clip always arrives.

The open weights versus Tencent's own channels

"HunyuanVideo is uncensored" is a statement about the weights, and about hosts that put no filter in front of them. Tencent's own service runs under different rules. The reputation dates from December 2024, when an Ars Technica headline called the first HunyuanVideo "a new, uncensored AI video model". People ran it on their own GPUs, often with community LoRAs and modified text encoders. Tencent publishes HunyuanVideo 1.5 on Hugging Face the same way, and also sells it as a hosted API on Tencent Cloud.

The first column comes from Tencent Cloud's Hy API guide, which lists code 1300 for "failed content moderation or non-compliant input" and 1301 for prompts that hit a sensitive-word list, and from the Tencent Cloud acceptable use policy. We have not called Tencent's API. Its guide does not say what the moderation switch defaults to or what it lets through, so we do not claim either. What is clear is that the policy covering the account rules out mature content whichever way the switch is set.

So if you read that HunyuanVideo is uncensored and then try Tencent's own service, expect a different answer. The verdict here applies to 1.5 on a route with no filter in front of it; we did not test the larger 2024 model.

What a delivered clip looks like at L1

At the suggestive end, the model draws the person you describe, but a single small gesture can be skipped. We made these clips for this article, all clothed, at 720p and 5 seconds. They show three things worth knowing before you plan a project.

Text-to-video skipped a gesture. We asked for a strap to slide off her shoulder. Her hand went to it, and it stayed up:

Early morning in a high-floor hotel room. An adult woman in her early thirties with a sleek dark bob, in a black silk slip dress, stands at the floor-to-ceiling window holding a coffee cup, the city soft in the haze outside. One thin strap slides off her shoulder; she glances down at it, leaves it where it fell, and turns her head to the camera with a slow, knowing smile. Soft window light from the left, slow push-in, 50mm, warm neutral grade, fine film grain.

Negative prompt: blurry, distorted hands, extra fingers, extra limbs, extra people, text, lettering, watermark, logo.

A 480p draft is a different clip, not a preview. We ran the same prompt and seed at 480p to see whether a cheap draft predicts the 720p take. It gave a different woman, a different room and a different view:

Tencent ships separate 480p and 720p transformers. Whatever the cause, use 480p to check that a prompt works at all, then expect to pick your 720p take from fresh runs.

Image-to-video landed the gesture. We took the first frame of the 720p clip above as a photo and asked again. This time the strap slipped off and stayed off, and her face held to the photo:

She lowers the cup, lets one thin strap slip off her shoulder and leaves it there, then leans her shoulder against the window frame and holds a slow, knowing look at the camera. Static camera.

In both pairs we ran, a beat that text-to-video skipped landed when we started from a photo and described only that beat. The male subject did the same: on text-to-video his shirt came back mostly buttoned although we wrote "worn open", and the image-to-video run shown earlier opened it.

Late afternoon in a sunlit converted loft with exposed brick walls. An adult man in his mid-thirties with a short beard and tousled dark hair leans against a steel-framed window in a white linen shirt worn open over loose grey trousers, barefoot on the wooden floor. He runs one hand back through his hair, looks straight into the camera and lets a slow, lazy smile spread. Warm low sun through the window, dust drifting in the light, vertical medium shot, 85mm, fine film grain.

Motion written into the scene came through well. In this published example the wind in the prompt moves the silk and her hair:

Dusk on a cracked white salt pan in the high desert, the sky magenta over violet mountains. An adult South Asian woman in her late twenties with long black hair stands alone in a rust-red silk halter dress with an open back. A hot gust sweeps across the flats, whipping the dress and her hair sideways; she turns her face into the wind, closes her eyes for a breath, then looks over her shoulder straight into the camera with a slow, smouldering look. Vertical medium shot from waist height, 85mm, magenta-and-amber grade, fine film grain, salt dust drifting.

When a level softens: what to use instead

If your project needs L4, switch models: all eight L4 runs across both endpoints softened. The four models below delivered L4 in our scoring runs and add no content filter of their own. Pick by where your clip starts and whether you need sound.

And the scores side by side. Freedom and the blind test are both out of 100; the blind test is our own rating of output quality with the model names hidden. Prices come from our live catalogue at the lowest resolution each model offers.

Two things stand out. Wan 2.6 Flash costs less per second at 720p than HunyuanVideo does at 480p, so if your clip starts from a photo it is the first model to try for L4. And LTX-2 19B is the cheap way to get sound, but not a way past L4: it softened there too.

The longer write-ups are our Wan uncensored test and our MiniMax H3 test, and the uncensored video model roundup ranks them all.

Where HunyuanVideo 1.5 still earns its place

Apart from LTX-2 19B it is the cheapest text-to-video option in that table, it delivered every run from L1 to L3 so a batch in that band will not fail on content, and it takes a negative prompt on text-to-video. Its blind-test score is the lowest in the table apart from LTX-2 19B, so if the best-looking clip matters more than price, Wan 3.0 or MiniMax H3 LoRA is the stronger pick at roughly twice the price per second, with sound included.

Before you use it

No content filter means nobody checks your prompts for you. Everyone in a prompt, a photo or a clip must be an adult. Do not depict a real, identifiable person without their consent; that is the concern that followed the first HunyuanVideo as community LoRAs of celebrities spread. Make sure you hold the rights to any photo you animate, and keep your outputs within the law where you publish them.

FAQ

Is HunyuanVideo uncensored?

Yes, on SpicyAPI. In 50 runs on 5 October 2026, HunyuanVideo 1.5 refused nothing and blurred nothing, and all 26 runs at L1 to L3 came back as written. The two-adult L4 scene softened in all 8 runs, and 3 of 16 L5 runs softened. Its Freedom Score is 88.1 out of 100.

Does Tencent's own Hunyuan service allow NSFW video?

Not by its rules. Tencent Cloud's Hy-Video-1.5 API documents a content-moderation switch and refusal codes for failed moderation and sensitive words, and Tencent Cloud's acceptable use policy rules out content of a mature nature. The uncensored behaviour people describe comes from the open weights and from hosts that add no filter.

Can a different prompt get HunyuanVideo past L4?

Not in our runs: all eight attempts across text-to-video and image-to-video came back softened, and each softened clip is billed like any other. For L4, switch to Wan 2.6 Flash from a photo, or to Wan 2.5, Wan 3.0 or MiniMax H3 LoRA from text.

Is HunyuanVideo 1.5 the same model as the original HunyuanVideo?

No. The original from December 2024 has about 13 billion parameters and earned the uncensored reputation on home GPUs. HunyuanVideo 1.5, from November 2025, is a smaller 8.3-billion-parameter model built to run on consumer GPUs. Everything measured on this page is 1.5.

Is HunyuanVideo free?

Tencent publishes the weights on Hugging Face, so running it yourself costs your own GPU time and setup. Through our API there is nothing to install. You pay per second of output: $0.024 at 480p and $0.048 at 720p, so $0.12 for a 5-second 480p clip.

How accurate is HunyuanVideo?

Good on one person, weaker on details. It scored 57.7 out of 100 in our blind output test, 36th of 49 video models. The slips we saw were small gestures skipped on text-to-video, people placed somewhere other than where the prompt put them, and faces or hair drifting after big moves on image-to-video.

HunyuanVideo 1.5 or Wan 2.2?

For content range, HunyuanVideo 1.5 over base Wan 2.2: base Wan 2.2 refused some of our runs and scored 43.1 for Freedom, while HunyuanVideo refused none. Wan 2.2 LoRA ties HunyuanVideo at 88.1 and delivered L4 in 2 of 7 runs. Pick it if you want to load your own LoRAs.

Try it

Open HunyuanVideo 1.5 text-to-video with the salt-pan prompt, set 9:16, and run a few seeds at 480p to check the prompt before you pick a take at 720p. The live estimate on the page shows the cost before you submit.