Uncensored Wan 3.0: Video Limits, Prompt Templates and Open Alternatives

Key facts

Developer

Alibaba (served by Alibaba Cloud Model Studio)

Launch

August 24, 2026

Model id

wan3.0-video

Modes

Text, image (first/last frame) and reference to video

Inputs

Text, image, video, audio, PDF/PPT/DOCX, public URL

Output

480P, 720P, 1080P; up to 30 seconds

Open weights

No (Wan 2.2 and 2.1 are open)

Uncensored?

No, moderated hosted service

What is Wan 3.0?

Wan 3.0 (model id wan3.0-video) is the latest model in Alibaba's Wan video series, served by Alibaba Cloud Model Studio and announced in Alibaba's developer community on August 24, 2026. Alibaba describes it as an all-in-one multimodal model: one endpoint covers text, first-frame or first-and-last-frame, and reference generation. It accepts text, images, video, audio, documents (PDF, PPT, DOCX) or a public web link, and outputs 480P, 720P or 1080P clips of up to 30 seconds with native audio. Alibaba offers the model as a whole through its cloud; the API is one way to use it. Unlike the earlier open releases, its weights have not been published.

Why people look for an uncensored Wan 3.0

The model runs behind Alibaba Cloud's content moderation, so prompts, images or files that break its rules are rejected. Because the earlier open Wan models have been fine-tuned freely by the community, many people assume an uncensored version must exist too. It does not: there are no weights to modify, and community LoRAs labelled Wan are built on older releases. Spicy AI does not run it either. If you want uncensored generation, use Spicy AI's photo animation tool, which runs MiniMax H3 Turbo without an extra content filter for consenting adults.

Video length, 1080P and other limits

The headline spec is length: up to 30 seconds in a single generation, and first-frame generation accepts any whole number of seconds from 2 to 30. Resolutions are 480P, 720P and 1080P across several aspect ratios. The model page is just as clear about what is missing: no batch inference, no fine-tuning, no context cache, no prefix continuation, no web search, no function calling and no structured output. Every request is asynchronous, the rate limit is 300 requests per minute, and Alibaba's launch article says some advanced capabilities may still be invitation-only. Typical generation times for long 1080P clips have not been published, so expect queueing.

Document to video: PDF, PPT, DOCX and web pages

The feature that sets this model apart is that it reads a document or a public web page as reference material. You attach a PDF, PPT or DOCX file, or a public link, and the model uses it to shape the script, look and pacing of the clip. A file and a link cannot be combined in one request, the link must be publicly reachable, and parsing or moderation failures come back as errors. Dense charts and complex layouts can be misread, so a person should check every number, name and claim in the finished video before it is published anywhere.

What creators make with Alibaba's new model

Because the model combines long clips with document input, the scenarios Alibaba highlights are practical rather than purely artistic. Reception so far is mostly launch coverage and early developer tests; we have not found independent benchmarks or user reviews we could verify, so we do not quote any. These are the uses that fit the model best:

• E-commerce: turn a product slide deck or listing page into a 20 to 30 second ad with a consistent product look.
• Education and training: convert a lab manual or course PDF into a step-by-step demonstration clip.
• Brand marketing: feed a brand web page so colors and styling stay consistent across a series of promos.
• Character stories: use a first frame or reference images to keep one character identical across a multi-shot scene.
• Music and audio clips: add an audio track as a reference so motion and cuts follow the beat.

Using the model in Model Studio

Alibaba offers the model through Alibaba Cloud Model Studio, in its console and through a DashScope-compatible API. With the API you submit a task with the header X-DashScope-Async: enable, receive a task_id, and poll until the job returns a video URL. A request combines a prompt with optional first or last frames, reference images or videos, a document or link, and audio; check the API reference for exact field names. Usage is billed per second of output, with rates that vary by resolution and region, so check the official model page before planning volume. Plan for moderation rejections, file-parsing failures and timeouts.

Wan3 vs Seedance 2.5 vs MiniMax H3

All three are current-generation video models, but they suit different jobs. Alibaba's model is the only one of the three that reads documents and web pages, and it reaches 30 seconds at 1080P. ByteDance's Seedance 2.5 also makes 30-second clips and adds up to 50 references and timestamp-based editing. MiniMax H3 makes shorter clips but has an open-weights route with regional restrictions, which matters if you need to run or fine-tune a model yourself.

Compiled from Alibaba Cloud, BytePlus and MiniMax documentation.

Is it open source? Weights and local deployment

No official weights have been released, so local deployment is not possible. Wan3.0 repositories under the Wan-AI organization on Hugging Face return a not-found or private response, and the Wan-Video organization on GitHub lists only earlier projects such as Wan2.2, Wan2.1 and Wan-Animate. Downloads that use the 3.0 name for GGUF, LoRA or ComfyUI files are mislabelled older files or third-party naming. Check the base model on the model card before you download anything, and treat any file claiming to be the new release as suspect until Alibaba publishes an official repository.

New release vs Wan 2.2 and a ComfyUI workflow

Wan 2.2 and Wan 2.1 are the open predecessors, published on GitHub under Apache 2.0. Wan 2.2 offers text and image generation at 14B scale with a mixture-of-experts design, plus a lighter 5B model for a single consumer GPU; check the model cards for exact variants. Both run locally in ComfyUI: load the official example workflow, put the diffusion model, text encoder and VAE files in the matching folders, then set frame count and resolution. Compared with the hosted model, the older model makes shorter, lower-resolution clips and has no document input, but it is free to run, accepts LoRAs and keeps data on your machine.

How to write prompts for it

Alibaba has not published a dedicated prompt manual, so these are proven practices for long multimodal video prompts rather than official rules. Good prompts read like a short shot list: who is on screen, what they do, where, how the camera moves and how the scene is lit.

• Order each shot as subject, action, setting, camera move and lighting, in one or two sentences.
• For 20 to 30 second clips, number the shots so the model knows where each beat starts and ends.
• Use a first frame or reference image when one character must look the same across the whole clip.
• When you attach a document, name the page or section that matters; otherwise the model chooses for you.
• Describe sound explicitly (rain, crowd, a slow piano) because the model generates audio with the picture.
• Draft short and at low resolution, then rerun only the winning prompt at 1080P.

Prompt templates you can copy

Replace the bracketed parts with your own details. Each template works for Alibaba's model and also for the uncensored alternative on Spicy AI, where the photo you upload takes the place of the first frame.

• Product ad: "Shot 1: [product] rotates on a [surface] under soft studio light. Shot 2: close-up of [key detail]. Shot 3: the product in use in [setting], upbeat music."
• Document explainer: "Using section [name] of the attached file, show [process] in [number] clear steps, a narrator voice explains each step, clean white lab setting."
• Cinematic portrait: "A [adult character] in [outfit] [action] in [location], the camera slowly pushes in, [time of day] light, ambient [sound]."
• Multi-shot story: "Shot 1: [character] waits at [place]. Shot 2: close-up as [second character] arrives. Shot 3: they [action], wide shot, music swells."
• Brand promo: "Match the colors and style of the attached web page, show [product or service] in three quick scenes, end on a calm wide shot."

Is there an uncensored Wan 3.0?

No. The model exists only as Alibaba Cloud's moderated service, and nobody can remove filters from weights that have not been released. What turns up under that search is usually community LoRAs and fine-tunes of Wan 2.1 or 2.2, which are older models still bound by their license and local law. We do not share ways to bypass moderation. If your goal is uncensored video from your own adult photo, the uncensored Wan 3.0 alternative on Spicy AI runs MiniMax H3 Turbo without content filters, for adults only and never with real people who have not consented.

Prompt tips for Wan 3.0-style results

1. Keep each shot to one clear action; long clips work best as a sequence of simple beats.
2. Only one of a file or a web link can be attached to a single request.
3. Search model cards, not file names, to tell Wan 2.2 apart from mislabelled 3.0 files.
4. On Spicy AI, start from a sharp, well-lit photo; the video follows your image closely.
5. Every person in an image or prompt must be an adult (18+), and never use a real person without consent.

Sources

• Wan3.0-Video model page (Alibaba Cloud Model Studio)
• Wan3.0-Video launch article (Alibaba Cloud developer community)
• Wan2.2 repository (GitHub)
• Wan2.1 repository (GitHub)
• Wan-AI organization (Hugging Face)
• Seedance 2.5 documentation (BytePlus)
• MiniMax platform documentation

More uncensored AI Video models

• uncensored minimax h3
• Unlimited Seedance 2.5
• Uncensored Kling
• Uncensored Kling 2.6
• Uncensored Kling O1
• Uncensored Sora
• Uncensored Veo
• Uncensored Hailuo AI
• Uncensored MiniMax H3 Max
• Uncensored Wan
• Uncensored Seedance
• Uncensored Seedance 2.5
• Uncensored Seedance 2.0
• Uncensored Seedance 2.0 Mini
• Uncensored HunyuanVideo
• Uncensored LTX Video
• Uncensored Runway
• Uncensored Pika
• Uncensored Luma Dream Machine
• Uncensored Vidu
• Uncensored PixVerse
• Uncensored EdenLive
• Uncensored Grok Imagine