ViralCut Agent is a fine-tuned Qwen2.5-3B-Instruct model trained with QLoRA SFT on tool-calling trajectories for video editing, social media optimization, and content strategy.

What It Does

Capability	How
🎬 Video Analysis	Analyze raw footage, find best moments, detect scenes
✂️ Professional Editing	Trim, transitions, effects, text overlays, color grading via FFmpeg
🎵 Audio Production	Search & add trending royalty-free music, sound effects, audio mixing
📊 Viral Optimization	Score content for TikTok/Instagram/YouTube, optimize for algorithms
🔍 Trend Research	Search current trends, hooks, sounds via web search
🚫 AI Slop Detection	Filter out AI-generated junk content
✍️ Caption Generation	Platform-optimized captions, hashtags, posting strategy

Tools

The agent was trained to call these tools autonomously:

python

# 1. FFmpeg for video processing
ffmpeg_cmd(command="ffmpeg -y -i input.mp4 -vf 'eq=saturation=1.3' output.mp4", 
           description="Boost color saturation")

# 2. Web search for assets and trends  
web_search(query="trending TikTok sounds food 2025", search_type="trending_content")
web_search(query="royalty free lo-fi beat", search_type="royalty_free_music")

# 3. Video analysis
analyze_video(video_path="raw.mp4", analysis_type="full")

# 4. Virality scoring
score_virality(video_path="edit.mp4", platform="tiktok", niche="food")

# 5. Caption generation
generate_caption(video_description="...", platform="tiktok", tone="casual")

# 6. AI content detection
detect_ai_slop(content_path="broll.mp4", check_type="video")

Quick Start

Install

pip install transformers torch peft bitsandbytes duckduckgo-search

Use as Agent (with real tools)

bash

# Clone the repo
git clone https://huggingface.co/ryu34/viralcut-agent
cd viralcut-agent

# Edit a video
python agent.py --video raw_footage.mp4 --platform tiktok --niche food

# Get a content plan (no video needed)
python agent.py --plan --niche "coffee shop" --platform tiktok

# Check files for AI slop
python agent.py --check-slop clip1.mp4 clip2.mp4

# Interactive mode
python agent.py

Use as Model (inference only)

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ryu34/viralcut-agent", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("ryu34/viralcut-agent")

messages = [
    {"role": "system", "content": "You are ViralCut Agent..."},
    {"role": "user", "content": "Edit my beach video into a TikTok with trending music and effects"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:]))

Training

Data

Mixed dataset of ~2,800 examples:

10 synthetic video editing trajectories — multi-turn conversations showing full edit pipelines (analyze → search → edit → score → caption)
~1,300 TikTok hooks/captions — real viral content data from benxh/tiktok-hooks-finetune
~1,200 general function-calling — tool-use backbone from NousResearch/hermes-function-calling-v1

Full dataset: ryu34/viralcut-agent-data

Method

Base model: Qwen/Qwen2.5-3B-Instruct
Method: QLoRA SFT (4-bit quantization, rank 16, alpha 32)
Training: 3 epochs, lr=2e-4, cosine schedule, assistant-only loss
Hardware: T4 16GB GPU (free tier compatible)
Framework: TRL v1.3+ SFTTrainer

Train It Yourself

bash

# Option 1: Google Colab (free T4 GPU)
# Open: https://huggingface.co/datasets/ryu34/viralcut-agent-data/blob/main/train_colab.ipynb

# Option 2: Direct script
wget https://huggingface.co/datasets/ryu34/viralcut-agent-data/resolve/main/train.py
pip install transformers trl torch datasets accelerate peft bitsandbytes
python train.py

Architecture

User Request ("Edit my raw footage into a viral TikTok")
    │
    ▼
┌─────────────────────────────────┐
│   ViralCut Agent (Qwen2.5-3B)  │
│   Fine-tuned for tool-calling   │
│                                 │
│   Thinks → Plans → Calls Tools  │
└──────────┬──────────────────────┘
           │
    ┌──────┼──────────────────────┐
    │      │      │      │        │
    ▼      ▼      ▼      ▼        ▼
 FFmpeg  Web    Video  Viral   AI Slop
  Edit  Search  Anal.  Score  Detect
    │      │      │      │        │
    └──────┴──────┴──────┴────────┘
           │
           ▼
    Final edited video + caption + strategy

Example Output

Input: "I have 8 minutes of raw ramen footage from Tokyo. Make a TikTok."

Agent actions:

📊 analyze_video(raw_ramen.mp4, "full") → Found 8 scenes, best: noodle pull at 0.9 energy
🔍 web_search("trending TikTok sounds food ASMR 2025") → Lo-fi city pop trending
🎵 web_search("royalty free lo-fi Japanese beat") → Found "Tokyo Nights" CC BY 4.0
✂️ ffmpeg_cmd(...) → Extracted hook shot with color boost
✂️ ffmpeg_cmd(...) → Speed-ramped broth prep
✂️ ffmpeg_cmd(...) → Assembled with fadeblack + slideright transitions
🎵 ffmpeg_cmd(...) → Mixed lo-fi music at 70% with ambient
📝 ffmpeg_cmd(...) → Added text hook + location overlay
📈 score_virality(...) → 82/100
🚫 detect_ai_slop(...) → Authentic ✅
✍️ generate_caption(...) → "This man has been making ramen by hand for 30 years"

Output: 17s vertical TikTok with professional transitions, trending music, text overlays. Score: 82/100.

Limitations

Model is 3B parameters — for complex creative decisions, larger models (7B+) would perform better
FFmpeg commands may need adjustment for specific file formats
Virality scoring is heuristic-based, not ML-based
Web search requires duckduckgo-search package
No actual video generation — this is an editing agent that works with your existing footage

Citation

bibtex

@misc{viralcut-agent-2025,
  title={ViralCut Agent: Autonomous Video Editing for Social Media},
  author={ryu34},
  year={2025},
  url={https://huggingface.co/ryu34/viralcut-agent}
}