This LoRA is trained on the Wan2.1 14B I2V 480p model and allows you to squish any object in an image. The effect works on a wide variety of objects, from animals to vehicles to people!
- Transform any image into a video of it being squished
- Trained on the Wan2.1 14B 480p I2V base model
- Consistent results across different object types
- Simple prompt structure that's easy to adapt
- Generate videos with 100+ Camera Control and VFX LoRAs on the Remade Canvas.
- Discord: Join our community to generate videos with this LoRA for free
- squish_18.safetensors - LoRA Model File
- wan_img2video_lora_workflow.json - Wan I2V with LoRA Workflow for ComfyUI
pip install git+https://github.com/huggingface/diffusers.git
import torch
from diffusers.utils import export_to_video, load_image
from diffusers import AutoencoderKLWan, WanImageToVideoPipeline
from transformers import CLIPVisionModel
import numpy as np
model_id = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"
image_encoder = CLIPVisionModel.from_pretrained(model_id, subfolder="image_encoder", torch_dtype=torch.float32)
vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
pipe = WanImageToVideoPipeline.from_pretrained(model_id, vae=vae, image_encoder=image_encoder, torch_dtype=torch.bfloat16)
pipe.to("cuda")
pipe.load_lora_weights("Remade/Squish")
pipe.enable_model_cpu_offload() #for low-vram environments
prompt = "In the video, a miniature cat toy is presented. The cat toy is held in a person's hands. The person then presses on the cat toy, causing a sq41sh squish effect. The person keeps pressing down on the cat toy, further showing the sq41sh squish effect."
image = load_image("https://huggingface.co/datasets/diffusers/cat_toy_example/resolve/main/1.jpeg")
max_area = 480 * 832
aspect_ratio = image.height / image.width
mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
image = image.resize((width, height))
output = pipe(
image=image,
prompt=prompt,
height=height,
width=width,
num_frames=81,
guidance_scale=5.0,
num_inference_steps=28
).frames[0]
export_to_video(output, "output.mp4", fps=16)
For best results, use this prompt structure:
In the video, a miniature [object] is presented. The [object] is held in a person's hands. The person then presses on the [object], causing a sq41sh squish effect. The person keeps pressing down on the [object], further showing the sq41sh squish effect.
Simply replace [object] with whatever you want to see squished!
This LoRA works with a modified version of Kijai's Wan Video Wrapper workflow. The main modification is adding a Wan LoRA node connected to the base model.
See the Downloads section above for the modified workflow.
The model weights are available in Safetensors format. See the Downloads section above.
- Base Model: Wan2.1 14B I2V 480p
- Training Data: 1.5 minutes of video (20 short clips of things being squished)
- Epochs: 18
Training was done using Diffusion Pipe for Training
35/F,Tencent Building,Kejizhongyi Avenue,Nanshan District,Shenzhen
