The best uncensored AI image and video generators you can run on your own PC (2026)
By Listicles research · Updated · Downloads, licences and file sizes pulled from Hugging Face on that date
The best uncensored AI image generator you can run locally is Z-Image-Turbo, and for video it's Wan 2.2. Both are free, open-weight and Apache 2.0, so you can use what you make commercially. Their 4-bit files are just 3.9 GB and 3.4 GB, small enough for an 8 GB graphics card. Run them in ComfyUI and there is no credit system, no account and no moderation filter between your prompt and the result.
That last part is why people search for "uncensored". Online generators run every prompt and every image through a safety filter, and they refuse, blur or bill you for the privilege. A model running on your own machine has no filter to hit: what it can draw is limited by its training data and by the law, not by a service's content policy. The real trade-offs are hardware and licences, and several of the most popular "uncensored" models can't legally be used for anything commercial.
The short version
Images, any use: Z-Image-Turbo (Apache 2.0, 3.9 GB at 4-bit).
Images, personal use only: Qwen-Image 2.1 Uncensored is the most downloaded uncensored image model (2.17M in 30 days), but its licence is research and evaluation only.
Video: Wan 2.2 (Apache 2.0). The 5B model fits an 8 GB card; the 14B image-to-video model wants 24 GB.
The app to run them: ComfyUI (free, 143,000+ GitHub stars). Invoke is friendlier for beginners.
The line that doesn't move: sexual images of real people without consent, or of anyone under 18, are illegal whatever tool makes them.
Uncensored image and video models compared
Downloads are Hugging Face's 30-day count on 29 October 2026. File size is the smallest practical quantised version (4-bit unless noted); you need roughly that much GPU memory plus a few GB of headroom.
We left out online "uncensored generator" websites entirely: they still decide what you can make, they charge credits, and your images sit on their servers. This list is open-weight models you download once. We ranked them on output quality and current relevance (30%), how much hardware they need, favouring 8 to 24 GB consumer GPUs (25%), whether the licence lets you use the output commercially (20%), how widely used they are, from 30-day Hugging Face downloads and likes (15%), and how easy they are to run in the common local apps (10%).
All figures came from the Hugging Face and GitHub APIs on 29 October 2026. We read each licence ourselves; several popular models are non-commercial, which most roundups don't mention.
The 8 best local uncensored image and video options, ranked
Best image model overall: small, fast and fully commercial.
Z-Image-Turbo is a 6.2-billion-parameter image model from Alibaba's Tongyi lab, released under Apache 2.0. It had 5,522 likes on Hugging Face, among the most of any image model, and 746,000 downloads in 30 days. It is a "turbo" model, built to produce an image in a handful of steps, so it's quick even on modest cards.
The 4-bit GGUF version is 3.9 GB and the 8-bit is 7.2 GB, so it runs on an 8 GB GPU. Because the licence is Apache 2.0, you can sell what you make and build on it, which is not true of the more famous "uncensored" model at #3.
Best video model: Apache 2.0, from 8 GB cards upwards.
Wan 2.2 is the most downloaded video model on Hugging Face: ComfyUI's repackaged version alone had 8.69 million downloads in 30 days. It comes in two useful sizes. The 5B text-and-image-to-video model is 3.4 GB at 4-bit and runs on an 8 GB card. The 14B image-to-video model gives much better motion, but it uses two 9.7 GB files (a "high noise" and a "low noise" stage), so plan for a 24 GB GPU.
Both are Apache 2.0, the most permissive licence among the video models here. Expect generation to take minutes per clip rather than seconds on consumer hardware.
The most popular uncensored image model, for personal use only.
This community build of Alibaba's Qwen-Image 2.1 is the most downloaded uncensored image model on Hugging Face: 2.17 million downloads in 30 days and 5,290 likes, more likes than the official model it's based on. It comes in versions for every kind of hardware: a 4.0 GB 4-bit file for Apple Silicon (MLX), a 4.2 GB NVFP4 file for recent Nvidia cards, 6.6 GB FP8, and a 14.2 GB full-precision GGUF.
The catch is the licence. Qwen-Image 2.1 is released under the Qwen Research License, which defines its permitted use as "for research or evaluation purposes only". Uncensoring the model doesn't change that. It's fine for experimenting at home, but don't use it for client work or anything you sell.
The strongest open video model, if you have the GPU.
MiniMax released H3's weights in July 2026, and it has quickly become one of the most downloaded models on the site: 5.58 million downloads in 30 days and 6,430 likes. It's large, at 33 billion parameters, but the community GGUF conversion has a pruned 4-bit version at 11.4 GB (18.8 GB unpruned), so a 16 to 24 GB card can run it.
Its community licence allows commercial use, but you need MiniMax's written permission if your products earn more than $20 million a year, and commercial products must display "MiniMax H3". Its model card says MiniMax's own hosted version applies automated moderation; the downloaded weights don't include that service layer, but the licence's lawful-use terms still apply.
Lightricks' LTX-2.5 had the most likes of any model in this list, 9,786, and 2.27 million downloads in 30 days. The LTX-2 line generates video with synchronised audio. The previous release, LTX-2.3, is 22 billion parameters and its 4-bit GGUF is 14.2 GB (10.6 GB at 3-bit), which gives a good idea of what LTX-2.5 needs.
LTX uses its own community licence rather than a standard open-source one, and we couldn't read the full licence file on our check. Read it on the model page before using LTX output commercially.
Old, but the biggest library of community models and styles.
Stable Diffusion XL is from 2023, yet it is still the most downloaded text-to-image model on Hugging Face (4.31 million in 30 days, 8,408 likes). The reason is its ecosystem: thousands of community fine-tunes and LoRA add-ons for specific styles, such as RealVisXL V5.0 for photorealism (974,000 downloads) and Animagine XL 4.0 for anime (799,000). Many of those fine-tunes were trained without the base model's content restrictions.
It's small (2.6 billion parameters), runs comfortably on 8 GB, and the OpenRAIL++ licence allows commercial use with a list of prohibited uses. Image quality and prompt-following trail the 2025–26 models above.
Cost
Free. CreativeML OpenRAIL++.
Hardware
6–8 GB GPU.
Weak spot
A generation behind on quality and text in images.
Black Forest Labs' FLUX models are known for image quality, but most are non-commercial. FLUX.2 klein 4B is the exception: 3.9 billion parameters and Apache 2.0, with 437,000 downloads in 30 days. Its bigger siblings, klein 9B and FLUX.2 dev, use the FLUX non-commercial licence. If you want FLUX quality and commercial rights on a modest GPU, this is the one.
Not a model: the free app that runs all of the above.
ComfyUI is the standard way to run local image and video models: 143,909 GitHub stars, GPL-3.0, updated daily. It works as a node graph, which looks intimidating, but it ships ready-made workflows for each major model, and the ComfyUI-GGUF extension (4,210 stars) loads the small quantised files this list recommends. If ComfyUI feels like too much, Invoke (28,912 stars, Apache 2.0) is a more traditional canvas app, and SwarmUI (MIT) puts a simple interface on top of ComfyUI. Forge, once popular, hadn't been updated since July 2025 on our check.
Cost
Free. GPL-3.0.
Runs on
Windows, Linux, macOS (Apple Silicon).
Weak spot
Steep learning curve.
How to set up a local uncensored generator
Check your GPU memory. 8 GB covers Z-Image-Turbo, Wan 2.2 5B, SDXL and FLUX.2 klein 4B. For 14B+ video models, plan on 16 to 24 GB. On a Mac, use the MLX or GGUF versions and count your total memory.
Install ComfyUI (or Invoke if you want a simpler interface), then add the ComfyUI-GGUF extension for quantised files.
Download the model file from its Hugging Face page. Pick the 4-bit (Q4_K_M or NVFP4/MLX) file first; move up to 8-bit if you have memory to spare.
Load a template workflow for that model, type a prompt and generate. Nothing is uploaded anywhere.
Check the licence before you sell anything. Apache 2.0 (Z-Image, Wan 2.2, FLUX.2 klein 4B) is the safe choice for commercial work.
What running it locally doesn't change
A local model has no filter, but the law still applies to what you make and share. Sexual imagery of anyone under 18 is illegal everywhere, including AI-generated images. In the US, the TAKE IT DOWN Act of 2025 makes it a federal crime to publish non-consensual intimate images of real people, including AI-generated "deepfakes", and many other countries have similar laws. Licences add their own prohibited uses on top: OpenRAIL++, the MiniMax H3 licence and others list them. Creative freedom over subject matter, style and violence in fiction is what local generation is genuinely for.
Which should you use?
Images on an 8 GB card, any use: Z-Image-Turbo.
Images, personal experiments, the most popular uncensored build: Qwen-Image 2.1 Uncensored.
A specific art style: SDXL with a community fine-tune.
Video on an 8 GB card: Wan 2.2 5B. On 24 GB: Wan 2.2 14B or MiniMax H3.
Video with sound: LTX-2.5, after reading its licence.
How we made this list
Scope
Open-weight image and video generation models you can download and run locally with free software. Hosted "uncensored generator" websites were left out.
Criteria
Quality and relevance (30%), hardware needs (25%), commercial licence (20%), 30-day downloads and likes (15%), ease of running locally (10%).
Sources
Hugging Face model API, model cards and licence files; GitHub API for apps. All checked on 29 October 2026.
Not tested
We did not benchmark image quality or generation speed. Hardware needs are estimated from file sizes. The LTX licence file could not be read on our check.
Updated
Questions
What is the best uncensored AI image generator to run locally?
Z-Image-Turbo, as of 29 October 2026. It is Apache 2.0, so you can use the output commercially, and its 4-bit file is 3.9 GB, small enough for an 8 GB GPU. Qwen-Image 2.1 Uncensored is more downloaded, but its licence allows research and evaluation use only.
Is there an uncensored AI video generator I can run for free?
Yes. Wan 2.2 is free and Apache 2.0: its 5B model is 3.4 GB at 4-bit and runs on an 8 GB GPU, and its 14B image-to-video model needs about 24 GB. MiniMax H3 and LTX-2.5 are larger alternatives with their own community licences.
How much VRAM do I need for local AI image generation?
8 GB runs most current image models in 4-bit form: Z-Image-Turbo is 3.9 GB, Qwen-Image 2.1 about 4 GB and SDXL about 7 GB. Video needs more: 8 GB for Wan 2.2 5B, and 16 to 24 GB for 14B-plus models such as Wan 2.2 14B or MiniMax H3.
Can I use uncensored AI images commercially?
It depends on the model's licence, not on whether it is uncensored. Z-Image-Turbo, Wan 2.2 and FLUX.2 klein 4B are Apache 2.0 and allow commercial use. Qwen-Image 2.1 is research and evaluation only, FLUX.2 dev is non-commercial, and MiniMax H3 requires permission above $20 million in annual revenue.
Is it legal to run an uncensored image generator?
Running open-weight models on your own computer is legal. What you create is not exempt from the law: sexual images of minors are illegal everywhere, and US law since 2025 makes publishing non-consensual intimate images of real people, including AI deepfakes, a federal crime. Model licences add their own prohibited uses.
Sources
Hugging Face model API: downloads, likes, licences and parameter counts for each model, 29 Oct 2026. huggingface.co
Quantised file sizes: unsloth Z-Image-Turbo and Qwen-Image 2.1 GGUF, QuantStack Wan 2.2 GGUF, unsloth LTX-2.3 GGUF, leejet MiniMax H3 GGUF, 29 Oct 2026. huggingface.co/unsloth