import torch
from diffusers import DiffusionPipeline
from diffusers.utils import load_image, export_to_video
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Lightricks/LTX-2.3", dtype=torch.bfloat16, device_map="cuda")
pipe.load_lora_weights("WepeNerd/Obscura_Remova")
prompt = "Remove the window curtains from the foreground."
input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png")
image = pipe(image=input_image, prompt=prompt).frames[0]
export_to_video(output, "output.mp4")Obscura Remova
Obscura Remova is an LTX 2.3 video-to-video LoRA for occlusion removal, haze cleanup, foreground-object removal, and temporal scene reconstruction.
- Prompt
- Remove the window curtains from the foreground.
- Prompt
- Remove the wooden planks from the foreground.
- Prompt
- Remove the windows from the foreground.
- Prompt
- Remove the windows from the foreground.
- Prompt
- Remove the objects from the foreground.
- Prompt
- Remove the green plants from the foreground.
- Prompt
- Remove the computer circuits from the foreground.
- Prompt
- Remove the plants from the foreground.
- Prompt
- Remove the bubbles from the foreground.
- Prompt
- Remove the fork and knife from the foreground.
Overview
This v2v LoRA removes anything obscuring the scene. It was made to "reconstruct" ground truth video from sparse data (when details of the scene can only partially be seen or the scene has semi-occlusion like smoke or haze.) If the data is completely invisible from any of the frames then it will have to be hallucinated, but if it is present anywhere, even semi-transparently, then it will be retained. Adjust the strength of the LoRA to remove more layers or go "deeper" into the scene. Start at a strength of about 1.25 and adjust to your liking. More complex removal requires higher strength.
Example prompt: "Remove the green plant from the foreground." is all that is needed to perform the task, but if details of the scene need to be hallucinated, it will be best to describe the scene in the background.
LTX still struggles with fast movement, but if you interpolate the input to 50fps and run inference at 50fps then it should help with fast moving scenes.
See examples for what this LoRA can do.
Recommended strength
Start at a LoRA strength of about 1.3 and adjust from there.
More complex removal usually requires higher strength.
- Downloads last month
- 463
Model tree for WepeNerd/Obscura_Remova
Base model
Lightricks/LTX-2.3