DynamiCrafter that works natively with ComfyUI's nodes, optimizations, ControlNet, and more.
 |
 |
Note
While this is still considered WIP (or beta), everything should be fully functional and adaptable to various workflows.
Go to your custom_nodes directory in ComfyUI, and install by:
git clone https://github.com/ExponentialML/ComfyUI_Native_DynamiCrafter.git
Important
This is a rapid release project. If there are any issues installing from main, the last stable branch is here. If everything is working fine, you can ignore this, but will miss out on the latest features.
The pruned UNet checkpoints have been uploaded to HuggingFace. Each variant is working and fully functional.
https://huggingface.co/ExponentialML/DynamiCrafterUNet
You will also need a VAE, The CLIP model used with Stable Diffusion 2.1, and the Open CLIP Vision Model. All of the necessary model downloads are at that link.
If you aready have the base SD models, you do not need to download them (just use the CheckpointSimpleLoader without the model part).
Place the DynamiCrafter models inside ComfyUI_Path/models/dynamicrafter_models
If you are downloading the CLIP and VAE models separately, place them under their respective paths in the ComfyUI_Path/models/ directory.
-
model: The loaded DynamiCrafter model.
-
clip_vision: The CLIP Vision Checkpoint.
-
vae: A Stable Diffusion VAE. If it works with < SD 2.1, it will work with this.
-
image_proj_model: The Image Projection Model that is in the DynamiCrafter model file.
-
images: The input images necessary for inference. If you are doing interpolation, you can simply batch two images together, check the toggle (see below), and everything will be handled automatically.
-
use_interpolation: Use the interpolation mode with the interpolation model variant. You can interpolate any two frames (images), or predict the rest using one input.
-
fps: Controls the speed of the video. If you're using a 256 based model, the highest plausible value is 4
-
frames: The amount of frames to use. If you're doing interpolation, the max is 16. This is strictly enforced as it doesn't work properly (blurry results) if set higher.
-
model (output): The output into the a Sampler.
-
empty_latent: An empty latent with the same size and frames as the processed ones.
-
latent_img: If you're doing Img2Img based workflows, this is the necessary one to use.
You can now use DynamiCrafter by applying ControlNet to the Spatial (image) portion to guide video generations in various ways. The ControlNets are based on 2.1, so you must download them at the link below (Thanks @thibaud !) .
ControlNet 2.1: https://huggingface.co/thibaud/controlnet-sd21
After you download them, you can use them as you would with any other workflow.
Tip
You don't have to use the latent outputs. As long as you use the same frame length (as your batch size) and same height and with as your image inputs, you can use your own latents. This means that you can experiment with inpainting and so on.
Tip
You can choose which frame you use as init by using VAE Encode Inpaint or Set Latent Noise Mask. You set the beginning batch mask to full black, while the rest are at full white. This also means you can do interpolation with regular models. As these workflows are more advanced, examples will arrive at a future date.
- Add various workflows.
- Add advanced workflows.
- Add support for Spatial Transformer options.
- Add ControlNet support.
- Ensure attention optimizations are working properly.
- Add autoregressive nodes (this may be a separate repository)
- Add examples. (For more, check here).
Thanks to @Doubiiu for for open sourcing DynamiCrafter! Please support their work, and please follow any license terms they may uphold.