---
base_model: THUDM/CogVideoX-5b
datasets: finetrainers/3dgs-dissolve
library_name: diffusers
license: other
license_link: https://huggingface.co/THUDM/CogVideoX-5b/blob/main/LICENSE
instance_prompt: 3D_dissolve A small tiger character in a colorful winter outfit appears in a 3D appearance, surrounded by a dynamic burst of red sparks. The sparks swirl around the penguin, creating a dramatic effect as they gradually evaporate into a burst of red sparks, leaving behind a stark black background.
widget:
- text: 3D_dissolve A small tiger character in a colorful winter outfit appears in a 3D appearance, surrounded by a dynamic burst of red sparks. The sparks swirl around the penguin, creating a dramatic effect as they gradually evaporate into a burst of red sparks, leaving behind a stark black background.
  output:
    url: "./assets/output_0.mp4"
- text: 3D_dissolve A small car, rendered in a 3D appearance, navigates through a swirling vortex of fiery particles. As it moves forward, the surrounding environment transforms into a dynamic display of red sparks that eventually evaporate into a burst of red sparks, creating a mesmerizing visual effect against the dark backdrop.
  output:
    url: "./assets/output_1.mp4"
tags:
- text-to-video
- diffusers-training
- diffusers
- cogvideox
- cogvideox-diffusers
- template:sd-lora
---

<Gallery />

This is a fine-tune of the [THUDM/CogVideoX-5b](https://huggingface.co/THUDM/CogVideoX-5b) model on the
[finetrainers/3dgs-dissolve](https://huggingface.co/datasets/finetrainers/3dgs-dissolve) dataset. We also provide
a LoRA variant of the params. Check it out [here](#lora).

Code: https://github.com/a-r-r-o-w/finetrainers

> [!IMPORTANT]
> This is an experimental checkpoint and its poor generalization is well-known.

Inference code:

```py
from diffusers import CogVideoXTransformer3DModel, DiffusionPipeline 
from diffusers.utils import export_to_video
import torch 

transformer = CogVideoXTransformer3DModel.from_pretrained(
    "finetrainers/3dgs-v0", torch_dtype=torch.bfloat16
)
pipeline = DiffusionPipeline.from_pretrained(
    "THUDM/CogVideoX-5b", transformer=transformer, torch_dtype=torch.bfloat16
).to("cuda")

prompt = """
3D_dissolve In a 3D appearance, a bookshelf filled with books is surrounded by a burst of red sparks, creating a dramatic and explosive effect against a black background.
"""
negative_prompt = "inconsistent motion, blurry motion, worse quality, degenerate outputs, deformed outputs"

video = pipeline(
    prompt=prompt, 
    negative_prompt=negative_prompt, 
    num_frames=81, 
    height=512,
    width=768,
    num_inference_steps=50
).frames[0]
export_to_video(video, "output.mp4", fps=25)
```

Training logs are available on WandB [here](https://wandb.ai/sayakpaul/finetrainers-cogvideox/runs/r39sv4do).

## LoRA

We extracted a 64-rank LoRA from the finetuned checkpoint 
(script [here](https://github.com/huggingface/diffusers/blob/main/scripts/extract_lora_from_model.py)). 
[This LoRA](./extracted_3dgs_lora_64.safetensors) can be used to emulate the same kind of effect:

<details>
<summary>Code</summary>

```py
from diffusers import DiffusionPipeline 
from diffusers.utils import export_to_video
import torch 

pipeline = DiffusionPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16).to("cuda")
pipeline.load_lora_weights("/fsx/sayak/finetrainers/cogvideox-crush/extracted_crush_smol_lora_64.safetensors", adapter_name="crush")
pipeline.load_lora_weights("/fsx/sayak/finetrainers/cogvideox-3dgs/extracted_3dgs_lora_64.safetensors", adapter_name="3dgs")
pipeline

prompts = ["""
In a 3D appearance, a small bicycle is seen surrounded by a burst of fiery sparks, creating a dramatic and intense visual effect against the dark background.
The video showcases a dynamic explosion of fiery particles in a 3D appearance, with sparks and embers scattering across the screen against a stark black background.
""",
"""
In a 3D appearance, a bookshelf filled with books is surrounded by a burst of red sparks, creating a dramatic and explosive effect against a black background.
""",
]
negative_prompt = "inconsistent motion, blurry motion, worse quality, degenerate outputs, deformed outputs, bad physique"
id_token = "3D_dissolve"

for i, prompt in enumerate(prompts):
    video = pipeline(
        prompt=f"{id_token} {prompt}", 
        negative_prompt=negative_prompt, 
        num_frames=81, 
        height=512,
        width=768,
        num_inference_steps=50,
        generator=torch.manual_seed(0)
    ).frames[0]
    export_to_video(video, f"output_{i}.mp4", fps=25)

```
  
</details>