Uncensored AI Video · HKUST & Huawei Research
Uncensored UniAVC: HKUST and Huawei's Audio-Video Model Explained (Not UNIVAC)
UniAVC is a unified audio-video model from the HKUST LongGroup lab, built with Huawei Research: one set of weights generates a short clip with synchronized sound, adds a soundtrack to a silent video, builds a picture around an audio track, or edits video and audio from a single instruction. It is not the 1950s UNIVAC computer. It ships as gated open weights under the LTX-2 Community License, needs an 80 GB GPU, and has no uncensored version. This guide covers what it does, how to run it, how it scores against MiniMax-H3, and how to write prompts; if you just want adult video from your own photo without a download, the H3 Turbo generator on Spicy AI runs in your browser.
Spicy AI is not UniAVC — this page currently runs MiniMax H3 Turbo (free tier) and Wan 2.6.
- Free to start
- No filters
- Runs in browser

UniAVC at a glance
- Developer
- HKUST LongGroup with Huawei Research
- Paper
- arXiv 2610.24791, Oct 17, 2026 (v2 Oct 20)
- Model size
- 19B DiT + Gemma-3 12B text encoder
- Tasks
- Text, audio or video to audio-video; joint editing
- Default output
- 704×1280, 24 fps, about 5 s with sound
- GPU
- Tested on a single 80 GB card
- Open weights
- Yes, gated; LTX-2 Community License
- Uncensored?
- No; no NSFW or community builds
UniAVC Alternative: H3 Turbo Uncensored Video Generator
Image to videoUpload a photo and describe the motion — Spicy AI animates it into a short clip.
Try a prompt
Example prompts are in English, which gives the most reliable results.
What is UniAVC?
UniAVC, short for Unified Audio-Video Creation, is a research model from the HKUST LongGroup lab at the Hong Kong University of Science and Technology, developed together with Huawei Research. The paper (arXiv 2610.24791) was first posted on October 17, 2026 and revised on October 20, with Kien T. Pham, Yuhao Liu, Qifeng Chen and Long Chen among the authors. The released checkpoint pairs a 19B diffusion transformer (DiT) in bf16 with a Gemma-3 12B text encoder; the Hugging Face sidebar shows 23B parameters, which likely counts extra modules such as the VAE and vocoder, so 19B refers to the DiT backbone only. The team publishes the code, weights and a Gradio demo together; the command-line scripts are just one way to run it. By default it renders 704×1280 video, 121 frames at 24 fps, about 5 seconds, with matching audio.
Why people look for an uncensored UniAVC
Searches for an uncensored UniAVC come from people who want a local audio-video model with no API filter in front of it. The honest answer is that the model has no uncensored version: the project page and model card offer no NSFW variant, and the LTX-2 Community License forbids bypassing content filters, non-consensual deepfakes and undisclosed machine-generated content. We found no community fine-tune or LoRA of it either; the only repository is the official one. Running an uncensored UniAVC locally is not an option either: local weights remove an API moderation layer, but that does not change what the license allows. If your goal is adult video from your own photos, Spicy AI's image-to-video tool runs MiniMax H3 Turbo without an extra moderation layer, for consenting adults only. It is a different model, without those editing features.
Want an uncensored UniAVC-style clip without a datacenter GPU? Animate a photo in your browser.
Open image-to-video →UniAVC is not UNIVAC: what the name means
Type uniavc into Google and it is usually corrected to univac, the UNIVAC I computer delivered in 1951 and the Univac company name. The two have nothing in common except letters. This is a 2026 AI model, and the name stands for Unified Audio-Video Creation, the title of its paper. To find the model rather than the computer, add a qualifier such as audio video model, HKUST or arXiv 2610.24791. The three primary sources are the paper, the GitHub repository HKUST-LongGroup/UniAVC and the Hugging Face repository TK3105/UniAVC; all are linked in the sources section below.
- UniAVC: 2026 audio-video generation and editing model, HKUST with Huawei Research.
- UNIVAC: early commercial computer line from the 1950s, unrelated to this model.
- Search tip: pair the name with "audio video", "HKUST", "github" or "huggingface".
How it works: learn alone first, then learn together
Most audio-video systems face two problems. Paired clips with good sound are scarce, and when video and audio share one noise schedule, tasks where one side is fixed and the other is generated are never really trained. The model answers with three ideas. Progressive training first teaches the video branch and the audio branch to generate and edit on their own large single-modality datasets, then freezes both and trains only the cross-attention bridges between them on paired data. Asymmetric noise scheduling keeps a shared timestep but gives each branch its own noise trajectory, because video needs to stay noisy longer to organize space and motion while audio settles earlier. Clean-branch subtasks pin one branch to a clean input during training, so conditioning one stream on the other is learned directly. The result is one checkpoint for every task, with no task-specific adapter: you switch tasks by changing the inputs and the script, not the weights.
| Training stage | Data used (paper) | What it teaches |
|---|---|---|
| Video branch | 34.3M samples from a 15.3M pool | Generating and editing video alone |
| Audio branch | 54.9M samples from a 14.9M pool | Generating and editing sound alone |
| Alignment | 0.67M paired samples from a 0.74M pool | Keeping picture and sound in sync |
The paired set is more than an order of magnitude smaller than the single-modality data, which is the point of the progressive design.
Supported tasks: generation, cross-modal and joint editing
It covers three generation modes and one editing mode, each with its own inference script. Text-to-audio-video creates a clip and its soundtrack from a prompt alone. Audio-to-video keeps an input track and builds matching images around it. Video-to-audio keeps an input video and writes a matching soundtrack. Joint editing changes picture and sound from one instruction; the official demos cover five edit types: changing a subject (for example swapping a T-shirt for a blue cycling jersey while the bar, beige wall and lighting stay put), adding a subject, removing one (such as a desk lamp, with the window and curtain filled in behind it), editing speech so a person says a new line with matching lip motion, and replacing the background. Each output is about 5 seconds; the team has not published a method for long videos, so longer pieces would have to be built from several clips.
| Task | Input | Output | Script |
|---|---|---|---|
| Text to audio-video (t2av) | Prompt only | New video and sound | infer_t2av.py |
| Audio to video (ta2v) | Prompt and audio | Video; audio kept | infer_ta2v.py |
| Video to audio (tv2a) | Prompt and video | Audio; video kept | infer_tv2a.py |
| Joint editing (tav2av) | Instruction, video and audio | Edited video and sound | infer_tav2av.py |
Intent-aware editing and self-refinement
Two features make the model practical for post-production rather than just demos. Intent-aware editing freezes one branch and uses it as the condition, so you can change only the audio while the picture stays identical, or change only the picture while the narration stays untouched. That is the workflow behind swapping a voice-over without re-rendering the shot, or replacing a background while keeping the original dialogue. Post-inference self-refinement handles the common case where one half of a result is good and the other is not: the model re-noises and re-denoises only the weak branch, using the accepted branch as a fixed reference. The README example refines audio with --post-refine audio and --refine-lambda 0.5, which runs about 15 steps by default.
- Change the dub, keep the shot: freeze video, edit audio.
- Change the scene, keep the narration: freeze audio, edit video.
- Fix only the weak half: refine audio or video after generation instead of starting over.
Hardware requirements: 80 GB GPU and VRAM
Hardware decides whether this model is usable for you. The 19B DiT and the Gemma-3 12B encoder together take about 65 GB, and the main model.safetensors file alone is about 46.1 GB. With the default staged offload, peak residency is about 39 GB for the DiT plus activations, and the team tested on a single 80 GB GPU. Turning offload off with --no-offload keeps every weight resident, which the README says needs far more than one 80 GB card at the default 720p size, and --no-tiled decoding raises memory use further. Runs on 24 GB or 48 GB consumer cards have not been shown, so do not plan on them. The team does not publish generation times either; 30 denoising steps on a 19B DiT is a heavy job, and real speed depends on your GPU and offload setting.
| Spec | Value |
|---|---|
| Weights | 19B DiT (bf16) + Gemma-3 12B text encoder |
| Download size | About 65 GB in total; model.safetensors about 46.1 GB |
| Peak VRAM (offload on) | About 39 GB plus activations |
| Tested GPU | Single 80 GB card |
| Default output | 704×1280, 121 frames, 24 fps, about 5 s |
| Size rules | Height and width multiples of 32; frames 8k+1 |
| Defaults | 30 steps, seed 42, --offload on |
| Software | Python 3.10, CUDA 12.8, PyTorch 2.8 |

UniAVC GitHub and Hugging Face download: setup steps
The code lives on GitHub at HKUST-LongGroup/UniAVC and the weights on Hugging Face at TK3105/UniAVC. The weight repository is gated, so you sign in and accept its conditions before the download works. Adoption is still small: shortly after release the GitHub repository had about 35 stars and the weights about 6 downloads and 8 likes. Those numbers show early interest from researchers, not search demand. Setup follows a standard Python workflow, summarized below; check the README for the exact arguments of each script.
- 1. Create a Python 3.10 environment (for example with conda) and install PyTorch 2.8 built for CUDA 12.8.
- 2. Clone the GitHub repository, then run pip install -r requirements.txt and pip install -e .
- 3. Log in to Hugging Face, accept the gated terms on TK3105/UniAVC, and run bash scripts/download_weights.sh; it also fetches the Gemma-3 text encoder.
- 4. Start with infer_t2av.py at the default size, then try infer_ta2v.py, infer_tv2a.py and infer_tav2av.py, or launch the bundled Gradio app.py.
- Common errors: download refused (gated access not accepted), missing Gemma-3 files, a frame count that is not 8k+1, a height or width that is not a multiple of 32, and out-of-memory after --no-offload or --no-tiled.
ComfyUI support: is there a node?
Not officially. The README offers only the command-line scripts and a Gradio app, there is no official ComfyUI node, and a Hugging Face Spaces search returns no hosted demo. Community nodes may appear later, but none could be confirmed from a reliable source, so treat any third-party node as unofficial and check what it actually loads. Because the model is initialized from LTX-2.3, LTX tooling may make a port easier, but that is an expectation, not a released feature. For now the realistic options are the Gradio app on your own large-GPU machine, or wrapping the CLI scripts in your own pipeline.
UniAVC vs MiniMax-H3: benchmark results
The paper compares the model with MiniMax-H3, InstructAV2AV, JAVEdit and JoVA. All numbers are self-reported by the authors and have not been reproduced independently, and the paper itself calls the results competitive, not a sweep. On editing, it leads on instruction compliance and video fidelity in JAVEditBench and has a much lower FVD on InsAVE-80K. On joint generation, H3 scores higher on perceptual quality, text consistency and lip sync in Verse-Bench, and on PEAVS in InsAVE-80K. A fair summary: the HKUST model is stronger as an editor, H3 is stronger as a generator.
| Benchmark and metric | UniAVC | MiniMax-H3 | InstructAV2AV |
|---|---|---|---|
| JAVEditBench IC ↑ | 4.42 | 4.11 | 2.88 |
| JAVEditBench VF ↑ | 4.48 | 4.17 | 3.93 |
| JAVEditBench AQ ↑ | 2.59 | 2.42 | 1.40 |
| JAVEditBench AVS ↑ | 0.2975 | 0.2307 | 0.1759 |
| InsAVE-80K FVD ↓ | 150.86 | 211.24 | 180.38 |
| InsAVE-80K FAD ↓ | 1.99 | 3.25 | 2.75 |
| InsAVE-80K PEAVS ↑ | 3.51 | 3.75 | 3.24 |
| Verse-Bench PQ ↑ | 6.26 | 6.87 | Not reported |
| Verse-Bench LSE-C ↑ | 6.72 | 6.96 | Not reported |
| Verse-Bench WER ↓ | 0.08 | 0.13 | Not reported |
↑ higher is better, ↓ lower is better. JAVEdit scores 0.3688 AVS on JAVEditBench, above all three. Cross-modal results: audio-to-video FVD 103.17, video-to-audio FAD 10.12.
Compared with LTX-2.3 and InstructAV2AV: which to pick
LTX-2.3 from Lightricks is the base model it starts from: a 22B open-weight DiT with a large ecosystem of community LoRAs. The HKUST team adds an audio branch and cross-modal bridges on top, aiming at joint editing; LTX-2.3 is the more mature choice for plain generation and has far more tooling. InstructAV2AV is the main editing baseline in the paper: the newer model beats it on instruction compliance, fidelity and FVD, while InstructAV2AV scores higher on some alignment metrics, such as TA-A on InsAVE-80K. MiniMax-H3 remains the stronger generator in the paper's own tables.
| UniAVC | LTX-2.3 | MiniMax-H3 | |
|---|---|---|---|
| Focus | Joint audio-video editing | Video generation | Joint generation and editing |
| Open weights | Yes, gated | Yes | See MiniMax |
| Ecosystem | Official repo only | Many community LoRAs | Official tools |
| Best for | Research and reproducible editing | Mature local workflows | Polished generation quality |
None of the three runs comfortably on a consumer graphics card.
Is UniAVC open source? Open weights and commercial use
The model is open-weight, not open source in the OSI sense. The code and weights are public and free to download after accepting the gate, but they come under the LTX-2 Community License from Lightricks, because it is derived from LTX-2.3 and derivatives must keep that license. Individuals, researchers and smaller companies can use, modify and host it, while entities with annual revenue of $10 million or more, counted together with affiliates, need a separate commercial license from Lightricks. Anyone redistributing it must include the full agreement, pass the acceptable use policy on to downstream users and keep attribution. Lightricks does not claim your outputs, but you are responsible for inputs, outputs and how they are used. Read the agreement itself before any commercial project.
Is there an uncensored UniAVC?
No. There is no uncensored UniAVC or NSFW release, and the license's acceptable use policy rules out bypassing content filters, deepfakes made without consent, undisclosed AI-generated media and anything involving minors. A Hugging Face search returns only the official repository, with no fine-tunes or LoRAs. LTX-2.3 has many community LoRAs, but they belong to the base model, not to UniAVC, and the same license applies. We do not provide prompts or scripts for removing filters. For adult video that you are allowed to make, a hosted tool is the simpler route: Spicy AI's image-to-video page animates your own photo with MiniMax H3 Turbo and no extra moderation layer, for consenting adults (18+) only.
Looking for an uncensored UniAVC alternative? Animate your own photo online, no local GPU required.
Animate a photo →Who it is for and how people use it
The model is closest to a research and post-production prototype: 5-second clips, 720p-class output and datacenter hardware put it out of reach of most hobbyists, but its editing tricks map onto real production chores. Short-drama and social video teams spend much of their time on re-dubbing, line fixes and background changes, exactly the edits shown in the official demos. Localization teams can try replacing a spoken line while keeping the shot, sound designers can generate a soundtrack for silent footage, and researchers get a reproducible baseline for joint audio-video editing with public code. Any edit involving a real person's face or voice needs their written consent and a clear AI label.
- Dialogue fixes: change one spoken line with matching lip motion and keep the room sound.
- Dubbing and localization tests: replace the voice track while the picture stays frozen.
- Foley for silent clips: run video-to-audio to add footsteps, rain or traffic that match the motion.
- Background swaps: move a scene from an office to a street without re-shooting the subject.
- Music-driven visuals: feed a short audio track and generate matching images around it.
- Research baselines: compare new joint-editing methods against the published numbers.
How to write UniAVC prompts that work
Because one prompt drives both picture and sound, write it like a short shot description with a separate sound line. Describe one shot with one main action in plain sentences, in this order: subject, action, setting, camera, light, then audio. For editing, write the instruction as a change plus what must stay the same; the official examples always name the parts to preserve, such as the background, the lighting or the other people. For speech edits, quote the new line exactly and say whose voice it is. Keep generation to roughly 5 seconds of action, since that is the default clip length, and remember sizes must be multiples of 32 with a frame count of 8k+1.
- Name the sound explicitly: "soft rain on the window, distant traffic" beats "realistic audio".
- Edits = change + keep: "Replace the lamp with a potted plant; keep the window, curtain and light."
- Speech edits: "She says 'See you tomorrow' in a calm voice; keep the background chatter."
- For audio-to-video, describe the visuals only; the track already sets the rhythm.
- For video-to-audio, list sound sources in the order they appear on screen.

Prompt templates you can copy
Replace the bracketed parts and keep the structure. The first rows target the model's own scripts; the last row is written for image-to-video, so you can also try it in the uncensored UniAVC alternative generator above.
| Goal | Template |
|---|---|
| Text to audio-video | [Adult subject] [one action] in [setting], [camera move], [light source]. Sound: [ambient], [voice or music]. |
| Subject edit | Change [item] to [new item]; keep [background], [lighting] and [other people] unchanged, keep the audio. |
| Remove an object | Remove the [object] in the foreground and fill in the [area behind it]; keep the soundtrack. |
| Speech edit | The [person] now says "[new line]" in the same voice with matching lips; keep picture and background sound. |
| Video to audio | Generate the sound for this clip: [source 1], then [source 2], [room tone], no music. |
| Image-to-video (alternative) | [Subject] turns toward the camera and says "[short line]", slow push-in, [light], [ambient sound]. |
Step by step: the uncensored UniAVC alternative on Spicy AI
If you lack datacenter hardware, the browser route skips the 65 GB download. It runs MiniMax H3 Turbo image-to-video, not UniAVC, so there is no joint audio-video editing; you animate a photo with a prompt.
- 1. Open the image-to-video tool from the button at the top of this page; the alternative model is preselected.
- 2. Upload a clear photo of an adult that you own or have consent to use.
- 3. Paste the image-to-video template above and fill in the action, camera, light and sound.
- 4. Generate a short clip, check the motion, then change one element per retry.
- 5. Download the result and label it as AI-generated when you share it.
What you can create
Photo-to-clip
Animate an adult portrait into a short clip in the browser, as a lighter uncensored UniAVC alternative.
Talking shots
A subject turns, smiles and says a short line, a format the HKUST model also targets with speech edits.
Mood scenes
Candlelit rooms, rain and neon streets for consenting adult characters.
Prompt tips for UniAVC-style results
- Write one shot with one action, and give sound its own sentence.
- For edits, always say what must stay the same.
- Plan for an 80 GB GPU; consumer cards are untested.
- Keep frame counts at 8k+1 and sizes at multiples of 32 to avoid errors.
- Only use faces and voices of adults who gave consent, and label AI output.
Self-hosting UniAVC vs generating online on Spicy AI
| Local deployment (official weights) | Spicy AI online | |
|---|---|---|
| Content filters | Set by HKUST & Huawei Research | No content filters |
| Open weights | Yes | Not needed — runs in the browser |
| Model | UniAVC | MiniMax H3 Turbo (free tier) and Wan 2.6 |
| Price to start | Varies by plan | Free tier available |
- Image-to-video in the browser
- Free queue at 0 credits (sign in, daily limit)
- Portrait (9:16) and landscape (16:9) clips
FAQ
What is UniAVC and who made it?
UniAVC is a unified audio-video generation and editing model from the HKUST LongGroup lab with Huawei Research, described in arXiv 2610.24791 (October 2026). One set of weights handles text-to-audio-video, audio-to-video, video-to-audio and joint editing.
Is UniAVC the same as UNIVAC?
No. UNIVAC is a 1950s computer line. Google often corrects uniavc to univac, so add "audio video model", "HKUST" or the arXiv number when searching.
What GPU do I need for UniAVC?
The team tested a single 80 GB GPU. Weights total about 65 GB, and default offload peaks around 39 GB plus activations. 24 GB and 48 GB cards are unverified.
How long are the generated clips?
The default is 704×1280, 121 frames at 24 fps, about 5 seconds. There is no official long-video mode.
Is there an uncensored UniAVC?
No. There is no NSFW release or community fine-tune, and the license forbids bypassing content filters. Spicy AI offers H3 Turbo image-to-video as a separate, adult-only alternative.
Is Spicy AI affiliated with HKUST & Huawei Research?
No. Spicy AI is an independent platform and is not affiliated with or endorsed by HKUST & Huawei Research. UniAVC is a trademark of its respective owner.
Is it free?
You can start free: sign in and Spicy AI's free video queue animates your photo with MiniMax H3 Turbo at 0 credits, with a daily limit. The fast queue, other video models and longer clips use credits.
Sources
Spicy AI is not affiliated with HKUST & Huawei Research. UniAVC and other product names are trademarks of their respective owners and are used here for identification only. All content must depict adults (18+).