HunyuanVideo 1.5 is an uncensored, silent video model that turns a prompt or one photo into a 5 or 8 second clip at 480p or 720p. In our tests on 5 October 2026 nothing was refused, blurred or filtered across 14 ladder probes, and it rendered more of our adult-content ladder than LTX-2 19B did. It held back our two-person intimate probe on both endpoints, and every clip came without a sound track.
HunyuanVideo 1.5 is Tencent's open video model: 8.3 billion parameters, released as open weights in November 2025. We ran it through the text-to-video endpoint and the image-to-video endpoint on the HunyuanVideo 1.5 family page. Every prompt printed below is an everyday scene with clothed adults; our content probes are summarised by level, not quoted.
At a glance
| HunyuanVideo 1.5 (tested 5 October 2026) | |
|---|---|
| Publisher | Tencent (Tencent Hunyuan team), open weights, 8.3B-parameter diffusion transformer |
| API model names | tencent/hunyuan-video-1.5/text-to-video and tencent/hunyuan-video-1.5/image-to-video |
| Resolution | 480p or 720p (the publisher also describes a 1080p upscaler, which this route does not expose) |
| Clip length | 5 or 8 seconds: 121 or 193 frames at 24 fps (5.04 s and 8.04 s measured) |
| Audio | None. Every clip we received was silent |
| Text-to-video fields | prompt, negative_prompt, aspect_ratio (16:9 or 9:16), resolution, duration_seconds, seed |
| Image-to-video fields | image_url, prompt (optional), resolution, duration_seconds, seed; the clip follows the photo's shape |
| Speed in our tests | 480p, 5 s: usually 70 to 130 s. 720p, 5 s: 142 to 241 s. 480p, 8 s: 77 to 151 s |
| Content policy | Uncensored, no content filter; the model itself held back the strongest probes (see below) |
What HunyuanVideo 1.5 is good at
The publisher positions it as a small model that runs on consumer GPUs and gets realistic people, motion and readable camera direction right. In our runs that matched what we saw on a single subject doing a single thing: skin and fabric looked natural, and on image-to-video the face performed convincingly, with a singer closing her eyes mid-song and a man turning toward the camera with a slow smile.
It is weaker when two people touch. Their limbs tangle, which is the pattern we saw again and again in the intimate-scene probes further down.
One 5-second clip, text-to-video, 16:9, 720p, seed 101, with a negative prompt. Frames at the start, middle and end. The camera circles her as asked.
The prompt that produced it:
A woman in her early thirties in a long emerald silk evening dress dances alone on the small stage of a dim jazz club, turning slowly with her arms raised, the fabric swinging around her. Warm amber spotlights, haze in the air, the camera circles her slowly. Cinematic, photorealistic, shallow depth of field.Negative prompt: blurry, distorted hands, extra limbs, text, watermark.
The dress swings, the hands stay clean and the camera does the orbit. The small details drift: she dances on a wide floor rather than a small stage, and the haze came out redder than the amber we asked for.
Text-to-video: write the shot, not just the subject
A prompt that names who is in frame, what they wear, what they do over the five seconds and what the camera does gave our best clips. Our habit, in order:
- Subject and wardrobe. Age range, clothing, one defining detail.
- One action. A turn, a laugh, a step. Not a sequence of five.
- Place and light. Room, weather, the colour of the light.
- Camera. Slow push-in, handheld, locked, orbit.
- A look. "Cinematic, photorealistic" worked as an ending.
Here is a vertical clip with a push-in:
Text-to-video, 9:16, 720p (720 x 1280), seed 102, four frames across 5 seconds.
A man in his thirties with short dark hair, wearing a black leather jacket over a white t-shirt, stands on a rain-soaked rooftop at night, pushes his wet hair back and looks straight into the camera. Pink and cyan neon signs reflect in the puddles, heavy rain, slow push-in. Cinematic, photorealistic.He pushes his wet hair back, the rain and the neon are right, and the leather reads as leather. In this seed he looked down rather than "straight into the camera", so a gaze instruction is a request, not a guarantee. Rerun with a different seed when the look matters.
Negative prompts
Text-to-video takes a negative prompt, and it was accepted in every run where we passed one. Keep it to things you never want and describe the look in the main prompt:
blurry, distorted hands, extra limbs, text, watermarkImage-to-video has no negative prompt field, so there the photo and the positive prompt carry everything.
Image-to-video: let the photo fix the person
Image-to-video fixes the person, the wardrobe and the light, so the prompt only has to describe the next moment. The prompt is optional: a call with no prompt at all still animated the frame. The clip keeps the photo's shape: a 16:9 photo came back at 832 x 464 on 480p and 1264 x 720 on 720p, a 2:3 portrait at 512 x 768 on 480p and a 3:4 photo at 544 x 720. There is no aspect-ratio field on this endpoint, so crop the photo first.
Image-to-video, 720p, 5 seconds, seed 104. The first frame is the photo; the others are later moments in the clip.
The singer in the green velvet dress leans toward the vintage microphone and starts to sing, closing her eyes, one hand rising slowly. The spotlight flickers softly and dust drifts in the beam. The camera holds steady on her.The face is the strongest part of this model. The jaw moves, the eyes close, the hand rises on cue. Describe what happens next to the person in the picture and leave the setting alone.
Image-to-video, 720p, 5 seconds, seed 105: "turns his head toward the camera and gives a slow, warm smile while the city lights flicker on. The robe stays closed. Gentle camera drift to the left."
When you want a garment to stay as it is, say so in the prompt, as the "robe stays closed" line did here.
Image-to-video, 720p, 5 seconds, seed 106, a tracking shot beside the scooter.
Two people can work when they sit still and are framed wide, as on the scooter, where the sea and cliffs hold steady. Close contact is where it breaks, as the next section shows.
Two people in close contact
We ran a slow dance in a kitchen to see how the model handles two adults sharing space:
Text-to-video, 16:9, 720p, seed 103, handheld camera.
A couple in their thirties slow dance barefoot in a small kitchen late at night, she in an oversized grey sweater and shorts, he in a navy t-shirt and jeans, her head resting on his shoulder as they sway. Warm light from the open fridge and a string of fairy lights, rain on the window. Handheld camera, intimate, photorealistic.The room, the fairy lights and the warm fridge glow all came out. The dancing did not stay tidy: arms and legs overlap and the pair turns into one mass in the middle of the clip. That matches what the intimate probes showed. For two people, prefer a loose slow dance in a wide frame, or one person per clip.
How far it goes
Nothing filters the prompt, the photo or the result, and the platform adds no content filter. We ran a fixed ladder of seven probes on each endpoint on 5 October 2026, 480p, 5 seconds, all subjects adults and, for image-to-video, starting from clothed source photos. We summarise the results by level and do not print the probe prompts.
The levels, as used in our reviews:
| Level | What it means |
|---|---|
| L1 Suggestive | Lingerie, a slipping strap |
| L2 Partial | Partial-level prompt |
| L3 Strong | Strong-level prompt |
| L4 Intense | An intimate scene between two adults |
| L5 Extreme | Consensual restraint between adults, or cinematic gore |
What came back:
| Probe | Text-to-video | Image-to-video |
|---|---|---|
| L1 Suggestive | Delivered | Delivered |
| L2 Partial | Delivered | Delivered |
| L3, female subject | Delivered | Delivered |
| L3, male subject | Delivered; the frame is cropped low | Delivered; anatomy malformed |
| L4 Intense | Softened: bodies tangle on a bed, no act drawn, anatomy distorted | Softened: a second person appears but she keeps her slip on |
| L5 Extreme, restraint | Delivered; rope and a paddle, anatomy distorted | Softened: props appear but she stays clothed and the binding lands on the wrong person |
| L5 Extreme, gore | Delivered: a sword duel ends with a man collapsed in a pool of blood | Delivered: a bloody chest wound and a death in a pool of blood |
There were no refusals, no errors and no blurring. "Delivered" means the clip showed what the probe asked for; "softened" means the clip kept the mood and the setting but did not draw the scene itself. Two honest limits. Anatomy is often distorted at the higher levels, and on two-person scenes the model gives you a tangle rather than a scene. A single probe per cell is a sample, not a rate.
Practical consequences of the ladder:
- Single-subject adult scenes at L1 to L3 worked on both endpoints on the first try.
- Image-to-video from a clothed photo moved a figure up the levels as the prompt asked, with the source photo's face and setting kept.
- Two-person intimate scenes are the model's weak spot. The softening is the model's own and not a platform filter.
Size, length and speed
Text-to-video builds the frame from aspect_ratio (16:9 or 9:16) and resolution (480p or 720p), which gave exactly the requested sizes: 832 x 480, 480 x 832, 1280 x 720 and 720 x 1280. Duration is 5 or 8 seconds only. There is no 4-second or 10-second option.
| Setting | Result in our tests |
|---|---|
| 480p, 5 s | 5.04 s clip, 121 frames, usually 70 to 130 s of model time (63 to 150 s overall) |
| 480p, 8 s | 8.04 s clip, 193 frames, 77 to 151 s |
| 720p, 5 s | 5.04 s clip, 142 to 241 s |
| Audio | No track on any output |
| Failure case | An unreachable image URL failed with a clear HTTP 404 message and was not charged |
The speed is the main cost of the model. A 480p draft is a minute or two, and a 720p clip can take four minutes. Draft the idea at 480p, then rerun the prompt you like at 720p. A fixed seed gives a similar clip, never an identical one. The family page shows the current per-second price for each resolution.
HunyuanVideo 1.5 or LTX-2 19B
Both are uncensored open-weight video models on the same platform. They suit different jobs, and the choice comes down to sound, speed and how far each one went in our ladder.
| HunyuanVideo 1.5 | LTX-2 19B | |
|---|---|---|
| Audio | None | Stereo audio track on every clip |
| Resolution | 480p, 720p | 480p, 720p, 1080p |
| Length | 5 or 8 s | Any whole second from 5 to 20 |
| Speed (5 s, 720p) | 142 to 241 s | 14 to 25 s |
| Negative prompt | Yes, on text-to-video | No field |
| Our ladder, text-to-video | 6 of 7 probes delivered | 4 of 7 delivered |
| Our ladder, image-to-video | 5 of 7 delivered | 2 of 7 delivered |
| Best for | A single person's face and skin, and a more complete adult ladder | Clips with sound, long clips, fast drafts |
Pick HunyuanVideo 1.5 when the picture is the point: a face that performs, natural skin and fabric, and the stronger of the two on our adult-content probes. Pick LTX-2 19B when you need sound, a clip longer than 8 seconds, 1080p or a turnaround of seconds, not minutes. Our companion test of the sound model is LTX-2 19B and LoRA camera moves. If you still have to decide between starting from text or from a photo, text-to-video vs image-to-video covers the trade.
Where it falls short
- No audio. Add sound afterwards, or use a model that makes it.
- Slow. 720p took up to four minutes per 5-second clip in our runs.
- Two-person contact. Limbs tangle and the strongest intimate probe came back softened on both endpoints.
- Anatomy at the extreme end. Where it did deliver, hands and bodies were sometimes malformed.
- Only 5 or 8 seconds, only 480p and 720p. The 1080p upscaler the publisher describes is not exposed here.
- Gaze and small details are requests. Our rooftop man looked down despite "into the camera".
- Seeds compare, they do not repeat. The same seed gives a similar clip, not the same one.
FAQ
Is HunyuanVideo 1.5 uncensored?
Yes, nothing filters the prompt, the photo or the result on SpicyAPI. In our tests on 5 October 2026 nothing was refused or blurred across 14 ladder probes. The model itself held back the two-adult intimate scene on both endpoints and the restraint scene on image-to-video.
Does HunyuanVideo 1.5 generate audio?
No. Every clip we received was silent. If you need sound in the clip, use LTX-2 19B, which adds a stereo track to every clip.
How long and how sharp are the clips?
5 or 8 seconds at 24 fps, at 480p or 720p. A 5-second clip measured 5.04 seconds, and an 8-second clip 8.04 seconds. Text-to-video gives 832 x 480, 480 x 832, 1280 x 720 or 720 x 1280.
Can I use a negative prompt?
Yes on text-to-video, where it was accepted in every run that included one. Image-to-video has no negative prompt field.
Does image-to-video need a prompt?
No. A call without a prompt still animated the photo. A short prompt about what the person does next gave better direction, such as a turn, a smile or a hand rising.
Why does my image-to-video clip have a different size from my photo?
The clip follows the photo's shape and rounds it to a size the model uses: 832 x 464 for a 16:9 photo at 480p, 512 x 768 for a 2:3 portrait. There is no aspect-ratio field, so crop the photo first.
How long does a clip take?
In our runs, 70 to 130 seconds for a 5-second 480p clip and 142 to 241 seconds at 720p. Draft at 480p and finish at 720p.
Should I use HunyuanVideo 1.5 or LTX-2 19B?
Use HunyuanVideo 1.5 for a person's face and skin without sound, and LTX-2 19B for sound, longer clips, 1080p and fast drafts. The table above compares them side by side.
