WanGP by DeepBeepMeep : The best Open Source Video Generative Models Accessible to the GPU Poor
WanGP supports the Wan (and derived models) but also Hunyuan Video, Flux, Qwen, Z-Image, LongCat, Kandinsky, LTXV, LTX-2, Qwen3 TTS, Chatterbox, HearMula, ... with:
- Low VRAM requirements (as low as 6 GB of VRAM is sufficient for certain models)
- Support for old Nvidia GPUs (RTX 10XX, 20xx, ...)
- Support for AMD GPUs (RDNA 4, 3, 3.5, and 2), instructions in the Installation Section Below.
- Very Fast on the latest GPUs
- Easy to use Full Web based interface
- Support for many checkpoint Quantized formats: int8, fp8, gguf, NV FP4, Nunchaku
- Auto download of the required model adapted to your specific architecture
- Tools integrated to facilitate Video Generation : Mask Editor, Prompt Enhancer, Temporal and Spatial Generation, MMAudio, Video Browser, Pose / Depth / Flow extractor, Motion Designer
- Plenty of ready to use Plug Ins: Gallery Browser, Upscaler, Models/Checkpoints Manager, CivitAI browser and downloader, ...
- Loras Support to customize each model
- Queuing system : make your shopping list of videos to generate and come back later
- Headless mode: launch the generation of multiple image / videos / audio files using a command line
Discord Server to get Help from the WanGP Community and show your Best Gens: https://discord.gg/g7efUW9jGV
Follow DeepBeepMeep on Twitter/X to get the Latest News: https://x.com/deepbeepmeep
- Vista 4D: Vista4D allows a Video Reshooting of a Dynamic scene from novel camera trajectories and viewpoints. In other words this Wan 2.1 model will let you relive from a different (moving) perspective a scene with moving people or objects. The sequences are quite short (usually 49 frames, max around 97 frames) but it is a lot of fun as for once this really works.
In real life, there is no chance you should have been able to run this model (it requires x3 the amount of VRAM than what is usally required for equivalent output res and the preprocessing needs 24 GB of VRAM to build a 4D map). But once again thanks to WanGP magic VRAM requirements have been reduced to 10 GB of VRAM or less.
It is highly recommended to apply the Lightx2v 4 steps lora profile. Also for best efficiency, you must list all the dynamic objects / people in the Dynamic object keywords input.
-
Magic Mask: generating a Video Mask or Image Mask has never been easier and faster. No need to get into the Video Mask Generator tab, just click the Magic Wand next to Mask field and enter a few keywords like blue car or lady to the right and a high quality mask powered by SAM3 will be generated automatically. You will appreciate the very good Temporal Consistency brought by SAM3.
-
Video Mask Generator with SAM3 support: if you still need to generate complex masks you can combine the good old point and click masks with the SAM3 / Magic Mask masks. You need to enable this feature in the Config / Extensions tab.
More nice goodies for LTX-2:
-
HDR Control Video support: you can now provide an HDR Control Video it will be automatically converted to SDR if model doesnt support HDR
-
LTX 2.3 SDR to HDR: thanks to a new HDR Ic lora, you can now convert SDR Videos to HDR using LTX 2.3. This feature is available as a new Control Video process and also in the Process Full Video plugin. Please note that the embedded Gradio Gallery video player converts automatically any HDR content to SDR, so if you want to enjoy the full HDR content you will need an external media player (for instance MPC-BE)
-
LTX 2.3 Control Video Injection in Phase 2: up to now even if you picked 2 phases, the Control Video was only injected in Phase 1 (Phase 2 was only used for upsampling). Now if you have chosen for at least one Ic Lora, a non null mutiplier for phase 2, the control video will be injected also for phase 2. This will increase output quality with 2 phases but will require more VRAM for phase 2.
-
Process Full Video Custom Settings: you can now reuse your own presaved settings in the plugin. As you will link the plugin to your settings any change to the saved settings will be immediatly available in the plugin. If you find some great combination of loras / model / settings to be used with this Plugin please share them on the discord server so that I can add them in the official list.
update 11.41: added Process Full Video Custom Settings
Lots of nice goodies for LTX-2:
-
LTX-2.3 Distilled 1.1: new version of the Distilled model released by LTX team, it should offer better audio and visuals. You will find also a Dev 1.1 version which uses Distilled 1.1 for Phase 2.
-
VBVR Lora Preset: This LoRA enhances the base LTX-2 for Enhanced Complex Prompt Understanding, Improved Motion Dynamics & Temporal Consistency. You can select it in the Settings list at the top.
-
Phase 1/2 Choice: you can now either you go for a good old 2 Phases Gen (1st Phase Low Res, 2nd shorter Phase High res) or go straight to a single High Res Phase (needs more VRAM and slower, but potentially higher quality). Please note that Outpainting mode and Pose/Edge/Depth extractors are always using 1 phase.
-
Improved Sliding Window: transition between windows should be less noticable, Sliding Windows overlapped Frames carry now also the audio of the overlapped frames, so the higher the number of overlapped frames the higher the chance that the sound / voice used in the previous window will be used in the new one.
-
Video Length not Limited by Audio: if you provide an Audio input, WanGP will no longer stops when the audio is consumed. It will continue the Video/Audio Gen based on the content of your Text prompt, and guess what ? it may reuse the same voice/sound used up to now ! This is an option, you need to check the checkbox Video Length not Limited by Audio.
-
Silent Movie Mode: if for some reason you want video with not only no sound but that takes into account that there is no sound (you dont want people to open their mouth for instance), just now leave the Control Audio empty
~~ - LTX2/2.3 Loras Split: as LTX2.0 Loras work badly with LTX2-3 and were getting on the way, now each version of LTX2 has its own lora folder. Loras will be moved automatically at startup using a lora migration script. I invit you to verify that the loras landed in the right folder.~~
-
System Loras Multipliers Overrides: WanGP adds automatically and transparently loras (that is they are loaded although they are not visible) if needed by a feature (distilled lora, id lora, outpaint lora, union control lora). You can now override the default multipliers used by WanGP by selecting the target lora in the Activated Loras input and by specifiying the corresponding Loras Multipliers.
-
Transfer Human Motion With Pose Alignment: you are trying to transfer a human motion from a control video, but you use a start image with a person who has a different body shape (larger, taller, ...) and stands in a different location in the frame. This is not going to work well as you start image wil end up distorted. This is a past issue, as now the control video pose can be aligned with the start image if you pick Transfer Human Motion With Pose Alignment. This feature is also supported by Wan Vace, start image must be the Background ref image.
-
Injected Frames & Sliding Windows: injected frames were not properly injected starting from window no 2. This is now supported.
-
Process Process Full Video Plugin: this bundled PlugIn which needs to be enabled first in the PlugIn tab, right now supports only Outpainting. It relies on LTX2 Lora outpainting. It is more or less a Super Sliding Windows mode but without the RAM restrictions and no risk to explode the Video Gallery with huge files. If you are patient enough you can change the Aspect Ratio of a few hours movie (check out below the 1 min sample). Behold how Sliding Windows transitions are almost invisible !
-
NEW Processes for Full Video Plugin: Refocus (remove blur), Ungrade (remove stylized color grading) and Uncompress (remove compression artifacts) have been added. Many thanks to Oumoumad Mohamed who created the Ic Loras (including the Outpainting lora ) that power these processes. If you have found some Ic Loras that are useful and dont cause glitches with Sliding Windows, let me know and I will add them.
-
WanGP API Video Gen: Plugin Developers can now Queue a Gen directly from a plugin. This opens the possibility of plugins that place various gen orders and then combine the results (hint: we could have our very own version of LTX-Destop inside WanGP).
-
New One Click Install / Update Scripts: We have to thank Tophness / @steve_Jabz for that one. Huge Kudos to him! The scripts will not only install WanGP but also all the Kernels (among Triton, Sage, Flash, GGuf, Lightx2v, Nunchaku) supported by your GPU. Please have a look at the instructions further down. Dont't hesitate to share feedback or report any issue.
update 11.31: fixed phase 1 forced incorrectly in some cases
update 11.32: bugs fixes, Process Full Video now supports Distilled 1.1 & accepts video without audio
update 11.33: Separated LTX2 & LTX2.3 loras in different folders, added easy loras multipliers override
update 11.34: Reverted split as not popular
update 11.35: added Aligned Pose Transfer, Injected Frames & Sliding Windows support, new processes for Process Full Video Plugin
- LTX-2 Ic Lora Rebooted: Ic Loras behave like Control Nets and can do Video to Video by applying an effect specific to the Ic Lora for instance Pose Extraction, Upsampling, Transfer Camera Movement, ... More and More Ic Loras are available nowadays. Until now WanGP Ic Lora implementation was based on the official LTX-2 github implementation (which a 2 phases process where the Ic Lora is only applied during the first low res phase). However I have just discovered that all the Ic Loras around expect in fact the ComfyUI implementation which is one phase only process at full res.
So from then on WanGP Ic Lora will work this way too. The downside is that a single Full Res pass is much more GPU intensive. But all is good in WanGP world, as the LTX2 VRAM optimisations will allow you to use Ic Loras at resolutions impossible anywhere else.
As a bonus I have tuned Sliding Windows for Ic Loras, and if you set Overlap Size to a single frame, transitions between windows when using Ic Lora will be almost invisible.
-
Outpaint Ic Lora: this new impressive Ic Lora will be loaded automatically if you select the Control Video for Ic Lora option and enable Outpainting. If you use Sliding Windows with Outpainting you will be able to outpaint a full movie (assuming you have enough RAM).
-
New Outpainting Auto Change Aspect Ratio: As a reminder WanGP let you define manually where an Outpainting should happen. Alternatively you can now ask WanGP to use outpainting to change the Width/ Height Aspect ratio of the Control Video. For instance you can turn any 16/9 video into a 4/3 video by generating new details instead of adding black bars. The Top/Bottom/Left/Right Sliders in this new mode will be used to define which area should be expanded in priority to meet the requested aspect ratio..
update 11.26: fixed outpainting ignored with if Manual Expansion was selected
- Magi Human: this is a newly Talking Head model that accepts either a custom soundrack or can generate the audio speech that comes with the video.
- The bad news :it is VRAM hungry (targets RTX 5090+) and very res picky, that is the ouput res must be either 256p or 1080p (using a 2 stage pipeline with upsampling). There is also a 540p version (using also an upsampler) but it is not included as I found it unpractical (ghosting guaranteed if your output is not exactly the right height/width ratio),
- The good news : now that it is WanGP optimized, 101 frames at 1080p requires "only" 16 GB of VRAM. If you dont have that much VRAM I recommend to still go for 1080p but set a 45 frames Sliding Window (not too low to avoid artifacts) as Sliding Windows sometime works well with this model.
I have spent a lot of time optimizing Magi Human, but I am not yet sure it is worth keeping it given all the constraints to run this model. So this is where I need YOU. Please share your experience using Magi Human on the Discord server and you shall decide its fate. Should we keep it or send it to the model graveyard ?
-
Ace 1.5 Turbo XL: the best open source song generator has now a big brother XL that delivers better audio quality and sticks closer to the requested lyrics.
-
LTX 2 Id Lora: due to a huge popular demand I have added this one (it is a new Generate Video option). You can provide a voice audio sample, a start image and text script and it will turn LTX 2/2.3 into talking heads. Cost is high to get this feature as Id Lora works only with LTX2/2.3 DEV. By chance it seems it can produce decent results in only 10 inference steps. To get the best results it is recommended to use prefix tags [VISUAL], [SPEECH] & [SOUND]. Alternatively you can use WanGP Prompt Enhancer that has been to tuned to generate a prompt following this syntax.
-
LTX 2 NAG: you can now inject a Negative Prompt even if you use the Distilled Model thanks to NAG support for LTX 2
-
LTX 2 DEV HQ Mode: this High Quality mode should produce better output at higher res. You can turn it on using the new HQ (res2s) Sampler and set 15 steps and guidance rescaler to 0.45. It is compatible with Id Loras. Note that a HQ steps is twice as slow as a vanilla Dev step, so it is going to be as slow as Dev if not slower.
-
LTX2 DEV Presets: Vanilla Dev mode & HQ Mode have lots of tunable settings. To make your life easier I have added selectionable presets in the Settings Drop Downbox
-
More Deepy :
- UI Improvements: you can queue requests by inserting empty lines between two requests, get the last turn by clicking the Down Arrow
- More Responsive: Deepy should execute much more quickly consecutive actions
- More Reliable: fast full context compaction (when deepy ran out of tokens), Deepy will remember what you stopped / aborted
- More Capabilities: you can ask Deepy to specifiy a guidance, denoising strength, ... value (the value defined in the tool template will be overridden)
As a reminder beside writting huge essays about how great you are, Deepy can generate Video, Image & Audio, extract / transcribe / trim / resize (when applicable) video or audio clip, inspect the content of an image or a video frame, generate black frames, ... Deepy used Tool templates but you can specify for one task the loras, number of frames, dimensions, ... There is also a CLI version of Deepy quite useful for remote use. Please check the fulldoc docs/DEEPY.md.
- Multi Multilines Prompts: check new options in "How to Process each Line of the Text Prompt", you can now have multiple multi lines prompts. They just need to be separated by an empty line.
update 11.21: added Ace Step 1.5 Turbo XL
update 11.22: added LTX2 NAG
Meet Deepy your friendly WanGP Agent.
It works offline with as little of 8 GB of VRAM and won't divulge your secrets. It is 100% free (no need for a ChatGPT/Claude subscription).
You can ask Deepy to perform for you tedious tasks such as:
generate a black frame, crop a video, extract a specific frame from a video, trim an audio, ...
Deepy can also perform full workflows:
1) Generate an image of a robot disco dancing on top of a horse in a nightclub.
2) Now edit the image so the setting stays the same, but the robot has gotten off the horse and the horse is standing next to the robot.
3) Verify that the edited image matches the description; if it does not, generate another one.
4) Generate a transition between the two images.
or
Create a high quality image portrait that you think represents you best in your favorite setting. Then create an audio sample in which you will introduce the users to your capabilities. When done generate a video based on these two files.
Deepy can also transcribe the audio content of a video (new to WanGP 11.11)
extract the video from the moment it says "Deepy changed my life"
Deepy reuses the Qwen3VL Abliterated checkpoints and it is highly recommended to install the GGUF kernels (check docs/INSTALLATION.md) for low VRAM / fast inference. now available with Linux!
Please install also flash attention 2 and triton to enable vllm and get x2/x3 speed gain and lower VRAM usage.
You can customize Deepy to use the settings of your choice when generating a video, image, ... (please check docs/DEEPY.Md).
Go the Config > Prompt Enhancer / Deep tab to enable Deepy (you must first choose a Qwen3.5VL Prompt Enhancer)
Important: in order to save Deepy from learning all the specificities of each model to generate image, videos or audio, Deepy uses Predefined Settings Templates for its six main tools (Generate Video, Generate Image, ...). You can change the templates used in a session or even add your own settings. Just have a look at the doc.
With WanGP 11.11 you can ask Deepy to generate a Video or an Image in specific dimensions and also a number of frames for a video. You can also specify an optional number of inference of steps or loras to use with multipliers. If you don't mention any of these to Deepy, Deepy Default settings or the current Templated Settings will be used instead.
WanGP 11 addresses a long standing Gradio issue: Queues keep being processed even if your Web Browser is in the background. Beware this feature may drain more battery, so you can disable it in the Config / General tab.
You have maybe also noticed the new option Keep Intermediate Sliding Windows in the Config / Outputs tab that allows you to discard intermediate Sliding Windows
See full changelog: Changelog
The 1-click automated scripts for both Windows (.bat) and Linux/macOS (.sh) make installation, environment management, and updates as seamless as possible. These scripts will not only install WanGP but also best acceleration kernels (Triton, Sage, Flash, GGuf, Lightx2v, Nunchaku) available for your config.
👉 Windows Users: Double-click the .bat files. Linux Users: Run the .sh files in your terminal.
Choose Installation Type
- Auto Install
- Manual Install
Manual Install
If you selected Manual Install, you will be guided through:
- Choose your package manager
- Name your environment
- Select your Install Mode
Once installed, use this script to launch the application. It runs WAN2GP using your active environment.
If you want to pass extra command-line flags to the WAN2GP launcher (like enabling advanced UI features or automatically opening your browser), create an args.txt file in your scripts folder.
Example args.txt:
--advanced --open-browser
Use this script to get the latest updates for WAN2GP and upgrade dependencies.
- 1. Update: Fetches the latest code from GitHub (
git pull) and updates requirements (pip install -r requirements.txt). - 2. Upgrade: Allows you to manually individually upgrade heavy backend components (like PyTorch, Triton, Sage Attention) based on your hardware profile.
Use this script to manage and switch between your sandboxed environments safely.
- Example Scenario: Let's say you have an environment named
env_stablethat works perfectly, but you want to try the new "Use Latest" combo. Instead of risking your working setup, you can runinstall.bat, create a new environment calledenv_testing, and select "Use Latest". - If the testing environment breaks or gives you errors, you can simply open
manage.bat, select Set Active Environment, and switch back toenv_stable. You are back up and running instantly.
Get started instantly with Pinokio App
It is recommended to use in Pinokio the Community Scripts wan2gp or wan2gp-amd by Morpheus rather than the official Pinokio install.
git clone https://github.com/deepbeepmeep/Wan2GP.git
cd Wan2GP
conda create -n wan2gp python=3.11.14
conda activate wan2gp
pip install torch==2.10.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txtgit clone https://github.com/deepbeepmeep/Wan2GP.git
cd Wan2GP
conda create -n wan2gp python=3.10.9
conda activate wan2gp
pip install torch==2.7.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cu128
pip install -r requirements.txtpython wgp.pyFirst time using WanGP ? Just check the Guides tab, and you will find a selection of recommended models to use.
If using Pinokio use Pinokio to update otherwise: Get in the directory where WanGP is installed and:
git pull
conda activate wan2gp
pip install -r requirements.txtUpgrade from Python 3.10, Pytorch 2.7.1, Cuda 12.8 to Python 3.11, Pytorch 2.10, Cuda 13/13.1 (for non GTX10xx users)
I recommend renaming first the old conda environment to avoid bad surprises when installing a different config in this old environment.
conda rename -n wan2gp old_wan2gpGet in the directory where WanGP is installed and:
git pull
conda create -n wa2gp python=3.11.9
conda activate wan2gp
pip install torch==2.10.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txtOnce you are done you will have to reinstall Sage Attention, Triton, Flash Attention. Check the Installation Guide -
if you get some error messages related to git, you may try the following (beware this will overwrite local changes made to the source code of WanGP):
git fetch origin && git reset --hard origin/main
conda activate wan2gp
pip install -r requirements.txtWhen you have the confirmation it works well you can then delete the old conda env:
conda uninstall -n old_wan2gp --all Process saved queues without launching the web UI:
# Process a saved queue
python wgp.py --process my_queue.zipCreate your queue in the web UI, save it with "Save Queue", then process it headless. See CLI Documentation for details.
For Debian-based systems (Ubuntu, Debian, etc.):
./run-docker-cuda-deb.shThis automated script will:
- Detect your GPU model and VRAM automatically
- Select optimal CUDA architecture for your GPU
- Install NVIDIA Docker runtime if needed
- Build a Docker image with all dependencies
- Run WanGP with optimal settings for your hardware
Docker environment includes:
- NVIDIA CUDA 12.4.1 with cuDNN support
- PyTorch 2.6.0 with CUDA 12.4 support
- SageAttention compiled for your specific GPU architecture
- Optimized environment variables for performance (TF32, threading, etc.)
- Automatic cache directory mounting for faster subsequent runs
- Current directory mounted in container - all downloaded models, loras, generated videos and files are saved locally
Supported GPUs: RTX 40XX, RTX 30XX, RTX 20XX, GTX 16XX, GTX 10XX, Tesla V100, A100, H100, and more.
For detailed installation instructions for different GPU generations:
- Installation Guide - Complete setup instructions for RTX 10XX to RTX 50XX
For detailed installation instructions for different GPU generations:
- Installation Guide - Complete setup instructions for RDNA 4, 3, 3.5, and 2
- Getting Started Guide - First steps and basic usage
- Models Overview - Available models and their capabilities
- Prompts Guide - How WanGP interprets prompts, images as prompts, enhancers, and macros
- Deepy Assistant - Enable Deepy, configure its tool presets, use selected media and frames, and run Deepy from the CLI
- Loras Guide - Using and managing Loras for customization
- Finetunes - Add manually new models to WanGP
- VACE ControlNet - Advanced video control and manipulation
- Command Line Reference - All available command line options
- Changelog - Latest updates and version history
- Troubleshooting - Common issues and solutions
- Nice Video that explain how to use Vace:
https://www.youtube.com/watch?v=FMo9oN2EAvE - Another Vace guide:
https://www.youtube.com/watch?v=T5jNiEhf9xk
- HuanyuanVideoGP - One of the best open source Text to Video generators
- Hunyuan3D-2GP - Image to 3D and text to 3D tool
- FluxFillGP - Inpainting/outpainting tools based on Flux
- Cosmos1GP - Text to world generator and image/video to world
- OminiControlGP - Flux-derived application for object transfer
- YuE GP - Song generator with instruments and singer's voice
Made with ❤️ by DeepBeepMeep