VDN-H3 Β· OpenVDN Β· open weights

Uncensored VDN MiniMax H3: AI Video That Renders Faster Than It Plays

Waiting minutes for a few seconds of AI video kills the fun. VDN MiniMax H3 turns that around: on a powerful server, a 14-second clip with matching sound is finished in about nine seconds, so the video is ready before you could even watch it. It is a faster take on the open MiniMax H3 video model, made by the OpenVDN research team, and it keeps the same look and voice. Even better, a desktop app called FreeVideo lets ordinary gaming PCs run it. Below you will find what the speed means for you, what people create with it, how to try it at home and prompts you can paste right away.

  • Faster than playback
  • Video with sound
  • Runs on home PCs
  • Open weights
Turn a photo into video now

A finished clip before you could finish watching it

Here is the promise in one line: the video takes less time to make than to play. On a server with eight top-end graphics cards, the fast mode of VDN-H3 turns a prompt into a 14.4-second widescreen clip, sound included, in about nine seconds. The research paper behind it says that is roughly 14.5 times quicker than the original H3 on the same machines.

You do not need a server room to feel the difference. On a single high-end card, the original model needs close to 14 minutes for one clip. The new full-quality mode brings that down to about five minutes, and the fast mode to under a minute. That is the gap between one test shot per coffee break and a whole scene sketched in the same time.

VersionWhere it runsTime for one 14-second clip
Original H3One high-end server cardAbout 14 minutes
VDN, full-quality modeThe same cardAbout 5 minutes
VDN, fast modeThe same cardUnder 1 minute
VDN, fast mode (8 cards)Eight server cardsAbout 9 seconds

Times come from the developers' own tests on NVIDIA B200 cards, published on GitHub, the project blog and in the paper.

What the speed changes for you

Speed sounds like a number for engineers, but it changes how creating feels. When a draft arrives in seconds, you stop guessing what a prompt might do and simply look. Here is what this model brings to everyday creators.

Ten ideas instead of one

Try different outfits, moods, camera moves and lines of dialogue in the time one clip used to take, then keep the take you love.

Sound arrives with the picture

Voices, footsteps, rain and room tone are generated together with the image in stereo, so there is no separate dubbing step.

The H3 look, kept

The developers describe the results as nearly identical to regular H3: the same cinematic picture, the same lip sync, the same respect for your prompt.

Your own PC is enough

Thanks to the FreeVideo app, a gaming computer with an 8GB graphics card can give it a go. No studio hardware needed.

Stories longer than a few seconds

The model remembers what happened earlier in the clip, which makes videos of a minute or more a realistic goal.

Open for everyone

The weights and code are public, so the community keeps adding apps, ComfyUI nodes and demos around it.

What creators make with near real-time video

Fast, sound-on video suits anyone who needs many clips rather than one perfect shot. These are the scenes where it shines, from client work to pure fun.

Short dramas, shot by shot

Fix a character with a start and end picture, chain several shots together, preview each in moments and polish the best ones in full-quality mode.

Vertical ads and product clips

Test many hooks, voice lines and product angles side by side, then post the version that feels strongest for your feed.

Talking portraits and living art

Start from a single still, add a line of dialogue and some background sound, and watch a portrait, illustration or AI photo speak and move.

Group shots from separate photos

The official demo places two people, a Samoyed and a cafe from different reference pictures into one moving scene with cafe chatter.

Storyboards for bigger projects

Filmmakers and game writers can block out a scene in an afternoon and show the team a moving draft instead of a page of notes.

Loops for virtual hosts and streams

The paper is literally about livestream video. Back-to-back clips suit virtual presenters and choose-your-path stories, though this is still early and worth testing yourself.

Romance and glamour scenes

Beach walks in swimwear, candlelit dates, slow dances: with original adult characters (18+), quick drafts make it easy to find the right mood.

Memes and greetings

Turn a selfie or a pet photo into a short clip with a spoken punchline or a birthday wish, ready to share in a chat.

Look closely, remember the rest: why it is so quick

VDN MiniMax H3 was released on September 6, 2026 by OpenVDN, a group with researchers from UC Berkeley, UT Austin and Impossible, Inc. Their paper, called Video DeltaNet, starts from a simple finding: the original model spends more than 85% of its time comparing every frame with every other frame. That is like a film editor re-watching the entire movie before each cut.

The new design works the way a good editor really does. It looks closely only at the frames right next to each other, and keeps short running notes about everything else, so the story still makes sense from start to finish. The first and last frames act as anchors, which keeps the opening and the ending consistent. Words and sound are still handled with full care, so prompts are followed and lips match the voice. A fast mode then cuts the drawing process from 50 passes to 8.

Row of video frames linked by short local beams and one long glowing memory ribbon, showing how hybrid attention speeds up the model

Fast drafts or final cut: picking a version

There are two versions to choose from. The fast one draws a clip in 8 passes and is what the team recommends for everyday use. The full-quality one keeps the classic 50 passes for final renders where every detail counts. A smart habit is to explore ideas in fast mode, then redo your favorite in full quality.

How does VDN-H3 compare with other ways of speeding up H3? The developers say its picture is almost the same as the original and that it beats MiniMax FastH3 on both quality and following instructions. The community Turbo LoRA only reduces the number of passes, while this model also changes how frames are compared. These are the team's own impressions from side-by-side examples; no test scores have been published, so trust your own eyes.

VDN fast modeVDN full qualityH3 Turbo LoRAMiniMax FastH3
Best forDrafts and daily useFinal rendersSpeeding up plain H3MiniMax's own quick option
How it saves timeSmarter frame comparison + 8 passesSmarter frame comparisonFewer passes onlySeparate fast variant
SoundYes, stereoYes, stereoYesYes
Extra downloadβ‰ˆ 5.1GB on top of H3β‰ˆ 4.3GB on top of H3One LoRA fileSeparate weights

One-minute videos and endless story chains

A single run gives you about 14 seconds. Because the model keeps notes instead of re-checking every frame, going longer no longer blows up the waiting time. FreeVideo version 0.4.0, released on October 13, 2026, says it can produce videos of more than a minute.

A simple recipe works in any tool: write your story in chapters of roughly 14 seconds, and repeat the character's look and the place at the start of each chapter, so your hero does not change coats halfway through. You can also use the last picture of one clip as the first picture of the next. The scenes then flow into each other and can be cut together like a real episode.

Run it at home on an 8GB gaming GPU with FreeVideo

The official code expects a 24GB card such as an RTX 4090. FreeVideo, an open-source desktop app released on October 2, 2026, lowers the bar a lot: it lists 8GB of graphics memory plus 16GB of system memory as the minimum, and picks the right lightweight settings for your card by itself. That puts the model within reach of an RTX 3060.

A word of honesty: nobody has published how long a clip takes on a small card, so expect it to be much slower than the server times above. FreeVideo runs on Windows and Linux, has a macOS preview, and is also available as a ComfyUI plugin.

  1. 1Download FreeVideo for your system from its GitHub page, or install the ComfyUI plugin.
  2. 2Let it fetch the model files and keep about 82GB of disk space ready.
  3. 3Add the helper files it asks for: the text encoder, the upscaler and the turbo LoRA.
  4. 4Choose the fast mode, paste a prompt and make a short test clip before trying long ones.
  5. 5Drop community LoRAs into the app's LoRA folder if you want a special style.
Your hardwareEasiest routeWhat to expect
8–12GB card (e.g. RTX 3060)FreeVideo appMeets the listed minimum; speed not published
24GB card (e.g. RTX 4090)FreeVideo or the official codeFits, using about 20–22GB of memory
One server cardOfficial codeUnder a minute per clip in fast mode
Eight server cardsSGLang serverAbout 9 seconds per clip
Home creator desk with a consumer gaming PC rendering a cafe scene video locally on a monitor

Start and end pictures, plus a ComfyUI workflow

Besides plain text, VDN-H3 accepts pictures as guides: a first frame, a last frame, both together (known as FL2VA) and, since September 17, 2026, several reference photos in one scene. The first-and-last-frame mode is the storyteller's favorite. You decide where a shot begins and where it lands, and the model fills in the movement and the sound in between.

ComfyUI users have two community options: the ready-made t8star/Vdn-Minimax-H3-Comfy files and the FreeVideo plugin, often paired with an upscaler for a sharper picture. Load your start and end images, describe what happens between them, and use the same words for clothes and lighting as in both pictures. The transition then looks natural.

Where to download it and how teams serve it

The model lives on Hugging Face under OpenVDN/vdn-minimax-h3, with the code on GitHub. You need the MiniMax H3 base files, about 72GB, plus a small VDN add-on: 5.1GB for the fast version or 4.3GB for the full-quality one. On the first start the software prepares itself for a few minutes; after that it is quick.

Studios that want the nine-second result run it on eight GPUs with SGLang, a serving tool the project supports out of the box. That setup currently handles text-to-video and first-and-last-frame jobs. The team also suggests letting a prompt helper such as H3-Context-IR polish rough ideas before you generate.

Prompt templates you can paste today

This model likes prompts that read like a director's note: who is in the picture, what they do in order, where the camera goes and what we hear. The templates below follow that pattern and use original adult characters.

  • Describe actions in order with first, then and finally; the model follows sequences well.
  • Put sound last with audio: and dialogue: so voices and background match the scene.
  • For long videos, write 14-second chapters and repeat the character's look and the place in each one.
  • Draft in fast mode, then redo your favorites in full quality.

All-purpose video with sound

[shot type], [subject and look], [action: first…, then…, finally…], [setting and light], [camera move], audio: [ambience], dialogue: "[line]"

Short drama

Medium close-up, a young woman in a red trench coat stands under a neon-lit awning in the rain, she turns toward the camera and whispers, slow push-in, cinematic teal-orange grade, audio: rain on metal, distant traffic, dialogue: "You came back."

First and last frame

From the first frame to the last frame, the character walks from the doorway to the window, camera tracks left smoothly, consistent outfit and lighting, soft room tone.

Group shot from several photos

[character A] and [character B] sit at the cafe table from reference 4, the Samoyed from reference 3 rests at their feet, they laugh and clink coffee cups, handheld warm look, cafe chatter.

Poolside glamour (18+)

Wide shot, an adult woman in a navy swimsuit walks along a sunlit pool, she turns toward the camera and smiles, slow dolly-in, golden hour light, audio: water splashing, soft summer music.

What the community is saying

Interest grew quickly. When this page was researched, the model had around 360 likes on Hugging Face, about 3,300 downloads in a month and 22 community demos built on it, including the popular minimax-h3-ultra-fast Space. Its GitHub page had about 670 stars, and FreeVideo had passed 2,800.

The mood around VDN-H3 is excitement: drafts in under a minute on one card, and real hope for long videos made at home. The usual complaint is the setup, with big downloads, several helper files and early-version bugs, as FreeVideo's long list of open issues shows. Reports on picture quality are positive, but they are personal impressions rather than measured scores.

Uncensored VDN MiniMax H3: what open weights really give you

Many people ask whether the fast version comes with a filter. There is no special uncensored or NSFW edition, and the developers do not describe any built-in content filter. The weights are open and run on your own machine, so no online moderation sits in between, although nobody knows exactly what the model learned in training. FreeVideo can load community LoRAs, but because the inner design changed, it is untested whether existing adult LoRAs for H3 still work. If you would rather skip the setup, an uncensored video generator runs in your browser for original adult (18+) characters.

No big download, no graphics card: animate a still of your original adult character online.

Open image to video

Uncensored Alternative Video Generator

Upload one still image, describe the movement, camera and sound you want, and get a short clip in the browser. Original adult characters (18+) only.

Start generating

Self-hosting VDN MiniMax H3 vs generating online on Spicy AI

Running it yourself gives you full control and the raw speed of the fast mode; the online route trades that for zero setup.

Local deployment (official weights)Spicy AI online
Content filtersYou configure your own setupRelaxed filters for adult content (18+)
Open weightsYes, about 82GB in totalNot needed, runs in the browser
Model in useVDN fast or full-quality versionA similar hosted video model
Getting startedDownload, install and a capable graphics cardOpen the page and upload a picture

FAQ

What is VDN MiniMax H3?

It is OpenVDN's faster version of the open MiniMax H3 video model. It makes widescreen 768p video with sound from text or pictures, using a smarter way of comparing frames called Video DeltaNet and a fast 8-pass mode.

How fast is it really?

The developers report about 9 seconds for a 14.4-second clip on eight server cards and under a minute on one. Their paper puts it at 14.5 times quicker than the original 50-pass H3.

Does the picture get worse?

The team calls it nearly lossless and better than MiniMax FastH3, based on side-by-side examples. No test scores have been published yet.

Can my graphics card run it?

The official code wants a 24GB card and uses about 20 to 22GB. FreeVideo lists 8GB of graphics memory and 16GB of system memory as the minimum, without published speeds for small cards.

Can it animate a photo or use first and last frames?

Yes. It works from text, a first frame, a last frame, both together (FL2VA) or several reference photos. The eight-GPU SGLang setup supports text and first-and-last-frame only.

How big is the download?

About 82GB in total: roughly 72GB for the H3 base model plus the VDN add-on and helper files.

Is there an uncensored version?

There is no dedicated uncensored build. Running the open weights at home has no online filter, and you need to test yourself whether community H3 LoRAs still work.

Sources

Browse all uncensored AI models β†’

Product and project names mentioned here belong to their respective owners. All characters must be adults (18+); do not create content of real people without their consent.

Sign in

Welcome to Spicy AI

or