Uncensored AI Video · NVIDIA NVlabs

Uncensored SANA-Video 2.0 720p: Cinematic Clips in Seconds, Not Minutes

Ask AI ▾

Waiting twenty minutes for a five-second AI clip kills every creative idea before it has a chance. NVIDIA's NVlabs team built SANA-Video 2.0 to fix exactly that. The SANA-Video 2.0 720p checkpoint is an open-weight model with about 4.47 billion parameters that writes eight seconds of 24 fps video from a short prompt, or brings a still image to life as the first frame, and on a single H100 the team reports a five-second 720p clip in roughly 13 seconds. Because the weights are public, you can run it on your own GPU with no cloud filter in between. Here is what it is good at, how fast it really is, how to run it and which prompts work best.

  • 720p, about 8 seconds
  • Open weights
  • Text and image to video
Animate a photo now →

Uncensored Alternative Video Generator

Image to video

Upload a photo and describe the motion, and Spicy AI animates it into a short clip.

Your photo will be animated.

From SANA-Video to version 2.0: what got better

The first SANA-Video, presented as an oral paper at ICLR 2026, proved that a small video model could run fast by swapping most of the heavy attention math for lighter linear attention. SANA-Video 2.0, released by NVlabs with partners including MIT and HKU in August 2026, keeps that speed idea but fixes the quality gap. The model more than doubles in size, from about 2 billion to about 4.5 billion parameters. Three out of every four layers still use the fast linear attention, while every fourth layer uses classic softmax attention as an anchor that keeps details and motion consistent. It also adopts the LTX 2.3 video autoencoder, adds joint training for text-to-video and first-frame image-to-video, and ships a 4-step distilled preview for instant drafts. The result is longer, sharper clips without losing the speed that made the series interesting.

Version 1 vs 2.0 at a glance

If you tried the original model and found the clips soft or short, SANA-Video 2.0 is worth a second look. These are the changes that matter for everyday creators.

SANA-Video (v1)SANA-Video 2.0
Model sizeAbout 2B parametersAbout 4.47B parameters (5B tier)
Attention designBlock-causal linear attentionLinear + softmax layers in a 3:1 mix
Video autoencoderEarlier VAELTX 2.3 (128 channels)
ModesText and image to videoText to video, first-frame image to video, 4-step preview
720p clip lengthShorter clips193 frames at 24 fps, about 8 seconds

What creators make with a fast video model

Speed changes how you work. When a clip renders in seconds instead of half an hour, you stop protecting every prompt and start exploring: ten camera angles for one scene, five lighting moods for one product, three endings for one story beat. That is why the first people talking about SANA-Video 2.0 720p are short-drama teams who storyboard dozens of shots per episode, e-commerce sellers who need fresh B-roll every week, illustrators who want a poster to breathe, and developers looking for a cheap, quick base for their own video tools. Adult (18+) creators like the same loop for glamour walks, fashion looks and romantic mood scenes, where trying many variations quickly matters more than any single perfect render.

Short drama storyboards

Preview a whole sequence of vertical-drama shots in minutes: draft with the 4-step version, then render the keepers in full quality.

Product and ad shots

Turntables, slow push-ins and moody light on watches, perfume or sneakers, ready as B-roll for stores and social ads.

Cartoon mini stories

Bright animated characters dancing, waving or running. NVIDIA's own demo is a cartoon rooster, and playful styles suit the model well.

Silver wristwatch turning on a black marble pedestal under soft rim light, an AI product video shot

How fast is it compared with Wan 2.2 and LTX-2.3?

Speed is the headline. NVIDIA's project page compares the time to render one five-second 720p clip at 40 steps on a single H100. With its Sol-Engine inference stack, SANA-Video 2.0 720p finishes in about 13 seconds, which the team describes as roughly 120 times faster than Wan 2.2 A14B. Without that stack, the paper reports 13.2 seconds for a 480p clip at 40 steps and a VBench score of 84.30. These are the lab's own charts, not independent benchmarks, and consumer-GPU timings have not been published yet. Still, if you are hunting for the fastest open source video model right now, this is the one everyone is measuring against.

Model (720p, 5 s, 40 steps, one H100)Reported render time
Wan 2.2 14B1556 s
HunyuanVideo788 s
LTX-2.3130 s
SANA-Video 2.0 14B (not yet released)69.3 s
SANA-Video 2.0 5B + Sol-Engine13.06 s

Figures come from NVIDIA's own project page and paper. Wan 2.2 still has the bigger community and LoRA ecosystem, and LTX-2.3 has a richer audio toolset; SANA's edge is raw speed.

What the 720p, 8-second output looks like

The SANA-Video 2.0 720p bucket renders at 1280×736, 193 frames at 24 frames per second, which works out to about eight seconds of footage. That is long enough for a full camera move or a small story beat, such as a character turning, smiling and walking out of frame. When you change the size, keep width and height divisible by 32, and pick a frame count where the number minus one divides by 8 (for example 121 or 193). Clean, cinematic shots with one clear subject look best: product turntables, landscapes, portraits with gentle motion and stylized animation. Crowded scenes with fast action are still the weak spot.

Full 50-step model vs the 4-step preview

Two SANA-Video 2.0 checkpoints share the 720p size, and they are meant to work as a team. The full model gives the best detail and supports first-frame image-to-video. The 4-step version, released in September 2026, is a distilled text-to-video model that drafts a clip almost instantly, so you can test ideas first and render the winners later.

Full model4-step preview
ModesText to video + first-frame image to videoText to video only
Steps and guidanceAbout 50 steps, CFG 8, flow shift 12, motion score 204 steps, CFG 1
Clip options720p, 193 frames, 24 fps5 s at 16 fps or 8 s at 24 fps
Best forFinal renders and animated stillsQuick drafts and storyboards

Image to video: make a still picture move

In the full SANA-Video 2.0 720p model you can hand over a picture as the first frame and describe what should happen next. Illustrations, posters, product photos and character art all work: hair moves in the wind, the camera pushes in, a smile appears. In the official prompt file, each line holds the text, then the tag <image>, then the path to your picture. Keep the motion modest, because the team notes that the video can drift away from the original details over eight seconds. The 4-step preview does not take images, so use the full model for this mode.

Cheerful cartoon fox dancing in a cozy vintage room with floral wallpaper, a short animated AI video idea

Download from Hugging Face and run it locally

Getting SANA-Video 2.0 running takes four steps, and you can test prompts in the free online demo on Hugging Face before downloading anything. Two common questions do not have official answers yet: NVIDIA has not published VRAM requirements, only H100 test results, and native ComfyUI or Diffusers support for version 2.0 has not been confirmed, so for now the official scripts are the reliable path.

  • Clone the NVlabs/Sana repository from GitHub and install its requirements.
  • Download the Efficient-Large-Model/SANA-Video_2.0_5B_720p checkpoint, or the _4step version for drafts, from Hugging Face.
  • Download the LTX 2.3 video autoencoder that the model card points to.
  • Run inference_video_scripts/inference_sana_video.sh with your prompt file, size and frame count.

Prompt templates that work

The team's advice is simple: short and specific beats long and vague, and prompt following drops when you stack many details. For SANA-Video 2.0 720p, a reliable formula is one subject, one action, one camera move and one light source. Copy these templates and swap in your own subject.

  • Product: "A sleek silver wristwatch rotating slowly on a black marble pedestal, soft rim light, shallow depth of field, 24fps cinematic."
  • Short drama: "Close-up of an adult woman on a rainy city street at night, neon reflections, she turns her head and smiles, handheld camera."
  • Cartoon: "A cheerful cartoon fox dancing in a cozy vintage room with floral wallpaper, bright colors, smooth motion."
  • Landscape: "Aerial drone shot gliding over misty mountains at sunrise, slow forward motion, golden light."
  • Image to video: first frame plus "The character gently blinks and her hair moves in the wind, camera slowly pushes in."
  • Glamour (18+): "An adult model in a satin evening gown walks down a candlelit hallway, slow tracking shot, warm light."
  1. Write one subject and one action per clip; eight seconds goes fast.
  2. Name the camera move (push-in, orbit, tracking, drone) in every prompt.
  3. Add one light source, such as neon, golden hour or rim light.
  4. Draft with the 4-step model, then re-render the best prompt with the full model.
  5. State that every person shown is an adult (18+).

Known limits and early community buzz

NVIDIA is upfront about the weak spots of SANA-Video 2.0: crowded or very dynamic scenes, human anatomy, on-screen text and physics can wobble, long complicated prompts are followed less closely, and image-to-video can drift from the source picture. The VBench score is self-reported. Early interest is real, with a few thousand downloads of the checkpoint on Hugging Face in its first month, but large independent reviews have not appeared yet. A 14B model rendering at 480p is listed as coming soon, with no date announced.

Is there an uncensored SANA-Video 2.0?

Hosted video apps check every prompt and frame, and they often refuse swimwear, lingerie or romance even between consenting adults. Open weights are different: when you run SANA-Video 2.0 720p on your own machine, no moderation service sits in the loop, so in that sense the local model is already uncensored. The honest catch is that it was not trained or tuned for adult content, and no NSFW fine-tune or LoRA had been found as of November 2026, so explicit detail is not its strength. Treat it as a fast draft engine for tasteful adult scenes, and keep every person an adult (18+), never a real person without consent.

No H100? Make adult videos in your browser

Most people do not have a data-center GPU or the patience to set up Python scripts. Spicy AI offers similar uncensored video generation online: upload a photo, describe the motion and get a short clip for adults (18+), with nothing to install. Only use photos you own or have permission to use.

Turn one photo into an uncensored short video, no GPU or setup needed.

Open image to video →

Self-hosting SANA-Video 2.0 vs generating online on Spicy AI

Local deployment (official weights)Spicy AI online
Content filtersNone in your own setupNo content filters for adults
Open weightsYesNot needed, runs in the browser
ModelSANA-Video 2.0 5B 720p or 4-stepSimilar uncensored video model, hosted online
Price to startFree download, your own GPUFree tier available

FAQ

What is SANA-Video 2.0?

It is an open-weight video model from NVIDIA's NVlabs team with partners such as MIT and HKU. The 5B 720p checkpoint, released in August 2026, makes clips of about eight seconds at 24 fps from text or from a first-frame image.

How long and how sharp are the videos?

The 720p model renders 1280×736 at 193 frames and 24 fps, about eight seconds. The 4-step preview offers 5 seconds at 16 fps or 8 seconds at 24 fps.

How much VRAM does it need?

NVIDIA has not published requirements yet. The official speed numbers were all measured on a single H100.

Does it work in ComfyUI?

Native ComfyUI or Diffusers support for version 2.0 has not been confirmed. For now, use the scripts in the NVlabs/Sana repository or the Hugging Face demo.

Is the 14B version out?

Not yet. The 14B model at 480p is marked as coming soon on the project page, with no release date.

Is it really uncensored?

Run locally, there is no moderation service filtering prompts. It was not trained for adult content, though, and no NSFW fine-tune has been found, so explicit detail is limited and all content must involve adults only.

Is Spicy AI affiliated with NVIDIA?

No. Spicy AI is an independent platform and is not affiliated with or endorsed by NVIDIA. NVIDIA and SANA are trademarks of their respective owners.

Sources

Sign in

Welcome to Spicy AI

or