Cakeify V0

🚀 蛋糕化視頻微調模型

本項目是基於 THUDM/CogVideoX - 5b 模型，在 finetrainers/cakeify - smol 數據集上進行微調的成果。我們將日常物品轉化為逼真蛋糕道具的創意視頻生成，帶來意想不到的驚喜。同時，還提供了參數的 LoRA 變體，你可以在 此處 查看詳情。

代碼倉庫：https://github.com/a-r-r-o-w/finetrainers

🚀 快速開始

推理代碼示例

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

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

prompt = """
PIKA_CAKEIFY On a gleaming glass display stand, a sleek black purse quietly commands attention. Suddenly, a knife appears and slices through the shoe, revealing a fluffy vanilla sponge at its core. Immediately, it turns into a hyper-realistic prop cake, delighting the senses with its playful juxtaposition of the everyday and the extraordinary.
"""
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)

訓練日誌可在 WandB 此處 查看。

✨ 主要特性

• 創意視頻生成：將日常物品通過視頻展示轉化為逼真的蛋糕道具，帶來創意十足的視覺體驗。
• LoRA 變體支持：提供 LoRA 變體參數，方便在不同場景下使用。

📦 模型信息

💻 使用示例

基礎用法

# 上述推理代碼即為基礎使用示例
from diffusers import CogVideoXTransformer3DModel, DiffusionPipeline 
from diffusers.utils import export_to_video
import torch 

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

prompt = """
PIKA_CAKEIFY On a gleaming glass display stand, a sleek black purse quietly commands attention. Suddenly, a knife appears and slices through the shoe, revealing a fluffy vanilla sponge at its core. Immediately, it turns into a hyper-realistic prop cake, delighting the senses with its playful juxtaposition of the everyday and the extraordinary.
"""
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)

高級用法

# 使用 LoRA 變體的代碼示例
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_cakeify_lora_64.safetensors")

prompt = """
PIKA_CAKEIFY On a gleaming glass display stand, a sleek black purse quietly commands attention. Suddenly, a knife appears and slices through the shoe, revealing a fluffy vanilla sponge at its core. Immediately, it turns into a hyper-realistic prop cake, delighting the senses with its playful juxtaposition of the everyday and the extraordinary.
"""
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)

🔧 LoRA

我們從微調後的檢查點中提取了一個秩為 64 的 LoRA（腳本 此處）。這個 LoRA 可用於模擬相同的效果：

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_cakeify_lora_64.safetensors")

prompt = """
PIKA_CAKEIFY On a gleaming glass display stand, a sleek black purse quietly commands attention. Suddenly, a knife appears and slices through the shoe, revealing a fluffy vanilla sponge at its core. Immediately, it turns into a hyper-realistic prop cake, delighting the senses with its playful juxtaposition of the everyday and the extraordinary.
"""
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)

以下是 LoRA 和非 LoRA 輸出的對比（相同設置和種子）：

⚠️ 重要提示

這是一個實驗性的檢查點，其泛化能力較差是已知的情況。