comi-logic/Wan2GP
7
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Flux Kontext and more

Browse Source
This commit is contained in:
deepbeepmeep
2025-07-15 22:26:56 +02:00
parent 37f41804a6
commit 64c59c15d9
21 changed files with 734 additions and 392 deletions
Download Patch File Download Diff File Expand all files Collapse all files

71
docs/FINETUNES.md
View File

@@ -2,22 +2,30 @@
A Finetuned model is model that shares the same architecture of one specific model but has derived weights from this model. Some finetuned models have been created by combining multiple finetuned models.
As there are potentially an infinite number of finetunes, specific finetuned models are not known by default by WanGP, however you can create a finetuned model definition that will tell WanGP about the existence of this finetuned model and WanGP will do as usual all the work for you: autodownload the model and build the user interface.
As there are potentially an infinite number of finetunes, specific finetuned models are not known by default by WanGP. However you can create a finetuned model definition that will tell WanGP about the existence of this finetuned model and WanGP will do as usual all the work for you: autodownload the model and build the user interface.
WanGP finetune system can be also used to tweak default models : for instance you can add on top of an existing model some loras that will be always applied transparently.
Finetune models definitions are light json files that can be easily shared. You can find some of them on the WanGP *discord* server https://discord.gg/g7efUW9jGV
All the finetunes definitions files should be stored in the *finetunes/* subfolder.
Finetuned models have been tested so far with Wan2.1 text2video, Wan2.1 image2video, Hunyuan Video text2video. There isn't currently any support for LTX Video finetunes.
## Create a new Finetune Model Definition
All the finetune models definitions are json files stored in the **finetunes** sub folder. All the corresponding finetune model weights will be stored in the *ckpts* subfolder and will sit next to the base models.
WanGP comes with a few prebuilt finetune models that you can use as starting points and to get an idea of the structure of the definition file.
## Create a new Finetune Model Definition
All the finetune models definitions are json files stored in the **finetunes/** sub folder. All the corresponding finetune model weights when they are downloaded will be stored in the *ckpts/* subfolder and will sit next to the base models.
All the models used by WanGP are also described using the finetunes json format and can be found in the **defaults/** subfolder. Please don’t modify any file in the **defaults/** folder.
However you can use these files as starting points for new definition files and to get an idea of the structure of a definition file. If you want to change how a base model is handled (title, default settings, path to model weights, …) you may override any property of the default finetunes definition file by creating a new file in the finetunes folder with the same name. Everything will happen as if the two models will be merged property by property with a higher priority given to the finetunes model definition.
A definition is built from a *settings file* that can contains all the default parameters for a video generation. On top of this file a subtree named **model** contains all the information regarding the finetune (URLs to download model, corresponding base model id, ...).
You can obtain a settings file in several ways:
- In the subfolder **settings**, get the json file that corresponds to the base model of your finetune (see the next section for the list of ids of base models)
- From the user interface, go to the base model and click **export settings**
- From the user interface, select the base model for which you want to create a finetune and click **export settings**
Here are steps:
1) Create a *settings file*
@@ -26,45 +34,60 @@ Here are steps:
4) Restart WanGP
## Architecture Models Ids
A finetune is derived from a base model and will inherit all the user interface and corresponding model capabilities, here are Architecture Ids:
- *t2v*: Wan 2.1 Video text 2
- *i2v*: Wan 2.1 Video image 2 480p
- *i2v_720p*: Wan 2.1 Video image 2 720p
A finetune is derived from a base model and will inherit all the user interface and corresponding model capabilities, here are some Architecture Ids:
- *t2v*: Wan 2.1 Video text 2 video
- *i2v*: Wan 2.1 Video image 2 video 480p and 720p
- *vace_14B*: Wan 2.1 Vace 14B
- *hunyuan*: Hunyuan Video text 2 video
- *hunyuan_i2v*: Hunyuan Video image 2 video
Any file name in the defaults subfolder (without the json extension) corresponds to an architecture id.
Please note that weights of some architectures correspond to a combination of weight of a one architecture which are completed by the weights of one more or modules.
A module is a set a weights that are insufficient to be model by itself but that can be added to an existing model to extend its capabilities.
For instance if one adds a module *vace_14B* on top of a model with architecture *t2v* one gets get a model with the *vace_14B* architecture. Here *vace_14B* stands for both an architecture name and a module name. The module system allows you to reuse shared weights between models.
## The Model Subtree
- *name* : name of the finetune used to select
- *architecture* : architecture Id of the base model of the finetune (see previous section)
- *description*: description of the finetune that will appear at the top
- *URLs*: URLs of all the finetune versions (quantized / non quantized). WanGP will pick the version that is the closest to the user preferences. You will need to follow a naming convention to help WanGP identify the content of each version (see next section). Right now WanGP supports only 8 bits quantized model that have been quantized using **quanto**. WanGP offers a command switch to build easily such a quantized model (see below). *URLs* can contain also paths to local file to allow testing.
- *modules*: this a list of modules to be combined with the models referenced by the URLs. A module is a model extension that is merged with a model to expand its capabilities. So far the only module supported is Vace 14B (its id is *vace_14B*). For instance the full Vace model is the fusion of a Wan text 2 video and the Vace module.
- *modules*: this a list of modules to be combined with the models referenced by the URLs. A module is a model extension that is merged with a model to expand its capabilities. Supported models so far are : *vace_14B* and *multitalk*. For instance the full Vace model is the fusion of a Wan text 2 video and the Vace module.
- *preload_URLs* : URLs of files to download no matter what (used to load quantization maps for instance)
-*loras* : URLs of Loras that will applied before any other Lora specified by the user. These loras will be quite often Loras accelerator. For instance if you specified here the FusioniX Lora you will be able to reduce the number of generation steps to -*loras_multipliers* : a list of float numbers that defines the weight of each Lora mentioned above.
- *auto_quantize*: if set to True and no quantized model URL is provided, WanGP will perform on the fly quantization if the user expects a quantized model
-*visible* : by default assumed to be true. If set to false the model will no longer be visible. This can be useful if you create a finetune to override a default model and hide it.
-*image_outputs* : turn any model that generates a video into a model that generates images. In fact it will adapt the user interface for image generation and ask the model to generate a video with a single frame.
In order to favor reusability the properties of *URLs*, *modules*, *loras* and *preload_URLs* can contain instead of a list of URLs a single text which corresponds to the id of a finetune or default model to reuse.
For example let’s say you have defined a *t2v_fusionix.json* file which contains the URLs to download the finetune. In the *vace_fusionix.json* you can write « URLs » : « fusionix » to reuse automatically the URLS already defined in the correspond file.
Example of **model** subtree
```
"model":
{
"name": "Wan text2video FusioniX 14B",
"architecture" : "t2v",
"description": "A powerful merged text-to-video model based on the original WAN 2.1 T2V model, enhanced using multiple open-source components and LoRAs to boost motion realism, temporal consistency, and expressive detail. multiple open-source models and LoRAs to boost temporal quality, expressiveness, and motion realism.",
"URLs": [
"https://huggingface.co/DeepBeepMeep/Wan2.1/resolve/main/Wan14BT2VFusioniX_fp16.safetensors",
"https://huggingface.co/DeepBeepMeep/Wan2.1/resolve/main/Wan14BT2VFusioniX_quanto_fp16_int8.safetensors",
"https://huggingface.co/DeepBeepMeep/Wan2.1/resolve/main/Wan14BT2VFusioniX_quanto_bf16_int8.safetensors"
],
"model":
{
"name": "Wan text2video FusioniX 14B",
"architecture" : "t2v",
"description": "A powerful merged text-to-video model based on the original WAN 2.1 T2V model, enhanced using multiple open-source components and LoRAs to boost motion realism, temporal consistency, and expressive detail. multiple open-source models and LoRAs to boost temporal quality, expressiveness, and motion realism.",
"URLs": [
"https://huggingface.co/DeepBeepMeep/Wan2.1/resolve/main/Wan14BT2VFusioniX_fp16.safetensors",
"https://huggingface.co/DeepBeepMeep/Wan2.1/resolve/main/Wan14BT2VFusioniX_quanto_fp16_int8.safetensors",
"https://huggingface.co/DeepBeepMeep/Wan2.1/resolve/main/Wan14BT2VFusioniX_quanto_bf16_int8.safetensors"
],
"preload_URLs": [
],
"auto_quantize": true
},
"auto_quantize": true
},
```
## Finetune Model Naming Convention
If a model is not quantized, it is assumed to be mostly 16 bits (with maybe a few 32 bits weights), so *bf16* or *fp16* should appear somewhere in the name. If you need examples just look at the **ckpts** subfolder, the naming convention for the base models is the same.
If a model is quantized the term *quanto* should also be included since WanGP supports for the moment only *quanto* quantized model, most specically you should replace *fp16* by *quanto_fp16_int8* or *bf6* by *quanto_bf16_int8*.
If a model is quantized the term *quanto* should also be included since WanGP supports for the moment only *quanto* quantized model, most specically you should replace *fp16* by *quanto_fp16_int8* or *bf6* by *quanto_bf16_int8*.
Please note it is important than *bf16", "fp16* and *quanto* are all in lower cases letters.
@@ -82,4 +105,4 @@ If you launch the app with the *--save-quantized* switch, WanGP will create a qu
You need to create a quantized model specifically for *bf16* or *fp16* as they can not converted on the fly. However there is no need for a non quantized model as they can be converted on the fly while being loaded.
Wan models supports both *fp16* and *bf16* data types albeit *fp16* delivers in theory better quality. On the contrary Hunyuan and LTXV supports only *bf16*.
Wan models supports both *fp16* and *bf16* data types albeit *fp16* delivers in theory better quality. On the contrary Hunyuan and LTXV supports only *bf16*.

66
docs/LORAS.md
View File

@@ -6,18 +6,19 @@ Loras (Low-Rank Adaptations) allow you to customize video generation models by a
Loras are organized in different folders based on the model they're designed for:
### Text-to-Video Models
### Wan Text-to-Video Models
- `loras/` - General t2v loras
- `loras/1.3B/` - Loras specifically for 1.3B models
- `loras/14B/` - Loras specifically for 14B models
### Image-to-Video Models
### Wan Image-to-Video Models
- `loras_i2v/` - Image-to-video loras
### Other Models
- `loras_hunyuan/` - Hunyuan Video t2v loras
- `loras_hunyuan_i2v/` - Hunyuan Video i2v loras
- `loras_ltxv/` - LTX Video loras
- `loras_flux/` - Flux loras
## Custom Lora Directory
@@ -64,7 +65,7 @@ For dynamic effects over generation steps, use comma-separated values:
## Lora Presets
Presets are combinations of loras with predefined multipliers and prompts.
Lora Presets are combinations of loras with predefined multipliers and prompts.
### Creating Presets
1. Configure your loras and multipliers
@@ -95,17 +96,37 @@ WanGP supports multiple lora formats:
- **Replicate** format
- **Standard PyTorch** (.pt, .pth)
## Safe-Forcing lightx2v Lora (Video Generation Accelerator)
Safeforcing Lora has been created by Kijai from the Safe-Forcing lightx2v distilled Wan model and can generate videos with only 2 steps and offers also a 2x speed improvement since it doesnt require classifier free guidance. It works on both t2v and i2v models
## Loras Accelerators
Most Loras are used to apply a specific style or to alter the content of the output of the generated video.
However some Loras have been designed to tranform a model into a distilled model which requires fewer steps to generate a video.
You will find most *Loras Accelerators* here:
https://huggingface.co/DeepBeepMeep/Wan2.1/tree/main/loras_accelerators
### Setup Instructions
1. Download the Lora:
```
https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan21_T2V_14B_lightx2v_cfg_step_distill_lora_rank32.safetensors
```
2. Place in your `loras/` directory
1. Download the Lora
2. Place it in your `loras/` directory if it is a t2v lora or in the `loras_i2v/` directory if it isa i2v lora
## FusioniX (or FusionX) Lora
If you need just one Lora accelerator use this one. It is a combination of multiple Loras acelerators (including Causvid below) and style loras. It will not only accelerate the video generation but it will also improve the quality. There are two versions of this lora whether you use it for t2v or i2v
### Usage
1. Select a Wan t2v model (e.g., Wan 2.1 text2video 13B or Vace 13B)
2. Enable Advanced Mode
3. In Advanced Generation Tab:
- Set Guidance Scale = 1
- Set Shift Scale = 2
4. In Advanced Lora Tab:
- Select CausVid Lora
- Set multiplier to 1
5. Set generation steps from 8-10
6. Generate!
## Safe-Forcing lightx2v Lora (Video Generation Accelerator)
Safeforcing Lora has been created by Kijai from the Safe-Forcing lightx2v distilled Wan model and can generate videos with only 2 steps and offers also a 2x speed improvement since it doesnt require classifier free guidance. It works on both t2v and i2v models
You will find it under the name of *Wan21_T2V_14B_lightx2v_cfg_step_distill_lora_rank32.safetensors*
### Usage
1. Select a Wan t2v or i2v model (e.g., Wan 2.1 text2video 13B or Vace 13B)
2. Enable Advanced Mode
@@ -118,17 +139,10 @@ Safeforcing Lora has been created by Kijai from the Safe-Forcing lightx2v distil
5. Set generation steps to 2-8
6. Generate!
## CausVid Lora (Video Generation Accelerator)
CausVid is a distilled Wan model that generates videos in 4-12 steps with 2x speed improvement.
### Setup Instructions
1. Download the CausVid Lora:
```
https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan21_CausVid_14B_T2V_lora_rank32.safetensors
```
2. Place in your `loras/` directory
### Usage
1. Select a Wan t2v model (e.g., Wan 2.1 text2video 13B or Vace 13B)
2. Enable Advanced Mode
@@ -149,25 +163,10 @@ CausVid is a distilled Wan model that generates videos in 4-12 steps with 2x spe
*Note: Lower steps = lower quality (especially motion)*
## AccVid Lora (Video Generation Accelerator)
AccVid is a distilled Wan model that generates videos with a 2x speed improvement since classifier free guidance is no longer needed (that is cfg = 1).
### Setup Instructions
1. Download the AccVid Lora:
- for t2v models:
```
https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan21_AccVid_T2V_14B_lora_rank32_fp16.safetensors
```
- for i2v models:
```
https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan21_AccVid_I2V_480P_14B_lora_rank32_fp16.safetensors
```
2. Place in your `loras/` directory or `loras_i2v/` directory
### Usage
1. Select a Wan t2v model (e.g., Wan 2.1 text2video 13B or Vace 13B) or Wan i2v model
@@ -268,6 +267,7 @@ In the video, a man is presented. The man is in a city and looks at his watch.
--lora-dir-hunyuan path # Path to Hunyuan t2v loras
--lora-dir-hunyuan-i2v path # Path to Hunyuan i2v loras
--lora-dir-ltxv path # Path to LTX Video loras
--lora-dir-flux path # Path to Flux loras
--lora-preset preset # Load preset on startup
--check-loras # Filter incompatible loras
```

26
docs/MODELS.md
View File

@@ -2,6 +2,8 @@
WanGP supports multiple video generation models, each optimized for different use cases and hardware configurations.
Most models can combined with Loras Accelerators (check the Lora guide) to accelerate the generation of a video x2 or x3 with little quality loss
## Wan 2.1 Text2Video Models
Please note that that the term *Text2Video* refers to the underlying Wan architecture but as it has been greatly improved overtime many derived Text2Video models can now generate videos using images.
@@ -65,6 +67,12 @@ Please note that that the term *Text2Video* refers to the underlying Wan archite
## Wan 2.1 Specialized Models
#### Multitalk
- **Type**: Multi Talking head animation
- **Input**: Voice track + image
- **Works on**: People
- **Use case**: Lip-sync and voice-driven animation for up to two people
#### FantasySpeaking
- **Type**: Talking head animation
- **Input**: Voice track + image
@@ -82,7 +90,7 @@ Please note that that the term *Text2Video* refers to the underlying Wan archite
- **Requirements**: 81+ frame input videos, 15+ denoising steps
- **Use case**: View same scene from different angles
#### Sky Reels v2
#### Sky Reels v2 Diffusion
- **Type**: Diffusion Forcing model
- **Specialty**: "Infinite length" videos
- **Features**: High quality continuous generation
@@ -107,22 +115,6 @@ Please note that that the term *Text2Video* refers to the underlying Wan archite
<BR>
## Wan Special Loras
### Safe-Forcing lightx2v Lora
- **Type**: Distilled model (Lora implementation)
- **Speed**: 4-8 steps generation, 2x faster (no classifier free guidance)
- **Compatible**: Works with t2v and i2v Wan 14B models
- **Setup**: Requires Safe-Forcing lightx2v Lora (see [LORAS.md](LORAS.md))
### Causvid Lora
- **Type**: Distilled model (Lora implementation)
- **Speed**: 4-12 steps generation, 2x faster (no classifier free guidance)
- **Compatible**: Works with Wan 14B models
- **Setup**: Requires CausVid Lora (see [LORAS.md](LORAS.md))
<BR>
## Hunyuan Video Models