finetrainers/crush-smol-v0

Prompt

DIFF_crush A red candle is placed on a metal platform, and a large metal cylinder descends from above, flattening the candle as if it were under a hydraulic press. The candle is crushed into a flat, round shape, leaving a pile of debris around it.

Prompt

DIFF_crush A bulb is placed on a wooden platform, and a large metal cylinder descends from above, crushing the bulb as if it were under a hydraulic press. The bulb is crushed into a flat, round shape, leaving a pile of debris around it.

Prompt

DIFF_crush A thick burger is placed on a dining table, and a large metal cylinder descends from above, crushing the burger as if it were under a hydraulic press. The bulb is crushed, leaving a pile of debris around it.

This is a fine-tune of the THUDM/CogVideoX-5b model on the finetrainers/crush-smol dataset. We also provide a LoRA variant of the params. Check it out here.

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

This is an experimental checkpoint and its poor generalization is well-known.

Inference code:

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

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

prompt = """
DIFF_crush A thick burger is placed on a dining table, and a large metal cylinder descends from above, crushing the burger as if it were under a hydraulic press. The bulb is crushed, leaving a pile of debris around it.
"""
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.

LoRA

We extracted a 64-rank LoRA from the finetuned checkpoint (script here). This LoRA can be used to emulate the same kind of effect:

Code

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("finetrainers/cakeify-v0", weight_name="extracted_crush_smol_lora_64.safetensors")

prompt = """
DIFF_crush A thick burger is placed on a dining table, and a large metal cylinder descends from above, crushing the burger as if it were under a hydraulic press. The bulb is crushed, leaving a pile of debris around it.
"""
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_lora.mp4", fps=25)