AI Art Studio

Local AI image and video training + generation, optimized for your RTX 4070 Laptop GPU (8GB VRAM).

Quick Start

# 1. Run automated setup (creates venv, installs everything)
python setup.py

# 2. Launch the app
run.bat
# or manually:
venv\Scripts\python.exe main.py

Architecture

ai_art_studio/
├── main.py              # Entry point
├── setup.py             # Automated installer
├── run.bat              # Windows launcher
├── requirements.txt     # Dependencies
├── core/
│   ├── config.py        # Central config with auto-save
│   └── dataset.py       # Dataset management (all formats)
├── captioning/
│   └── auto_caption.py  # WD Tagger + BLIP-2/Florence-2
├── training/
│   └── trainer.py       # LoRA training engine
├── generation/
│   ├── image_gen.py     # Image generation (SD1.5/SDXL/FLUX)
│   └── video_gen.py     # Video generation (WAN2.1/AnimateDiff)
└── gui/
    ├── main_window.py   # Main application window
    ├── widgets.py        # Custom reusable widgets
    └── theme.py          # Dark theme stylesheet

Features

Dataset Tab

• Drag folders or pick individual files
• Supports every image format: PNG, JPG, GIF, WebP, HEIC, AVIF, BMP, TIFF, PSD, DNG, CR2, NEF, TGA, EXR, HDR, JXL, and dozens more
• Supports every video format: MP4, AVI, MOV, MKV, WebM, FLV, MTS, and more
• Animated GIF/APNG/WebP frame extraction
• Video keyframe extraction
• Auto-converts everything to training-compatible format
• Preview with caption display

Auto-Captioning (Dataset Coherence System)

This is what makes your diverse dataset work together:

1. WD Tagger v3 — Identifies specific visual features, character details, poses, compositions using booru-style tags
2. BLIP-2 / Florence-2 — Generates natural language descriptions for scene context
3. Content Analyzer — Categorizes images (solo, multi-character, comic, scene) for format-appropriate captions
4. Combined Mode — Merges tags + natural language so the model understands:
   • What characters look like
   • What's happening in the scene
   • How different images in your dataset relate to each other

Training Tab

• SD 1.5: Full LoRA training, ~6GB VRAM
• SDXL: LoRA with fused backward pass, ~8-10GB VRAM
• FLUX: LoRA with split_mode + FP8, fits in 8GB at 512x512
• Real-time loss/LR/VRAM monitoring
• Full hyperparameter control
• Multiple optimizers: Adafactor, AdamW8bit, Prodigy, Lion
• Automatic checkpoint saving

Generation Tab

• Image: SD 1.5, SDXL, FLUX
• Video: WAN 2.1, AnimateDiff
• 12+ samplers with Karras variants
• Hi-res fix (upscale + refine)
• Full LoRA loading with adjustable weight
• Generation history with thumbnails
• Auto-save with metadata

Settings Tab

• 4 offloading presets tuned for 8GB VRAM:
  • none — All on GPU (needs 12GB+)
  • balanced — Smart CPU offload (recommended)
  • aggressive — Sequential offload (slow but fits anything)
  • cpu_only — Maximum VRAM savings
• Granular transformer/text encoder offload percentages
• VAE slicing, tiling, attention slicing toggles
• FP8 base model support
• xformers toggle
• All settings auto-save on change

Hardware Optimization Notes (RTX 4070 Laptop 8GB)

Your setup: i7-13620H, 32GB DDR5 5600, RTX 4070 Laptop 8GB, CUDA 12.4

Training recommendations:

• SD 1.5 LoRA: 512x512, batch 1, Adafactor — fits easily
• SDXL LoRA: 512x512, batch 1, Adafactor, fp8_base, cache_latents — tight but works
• FLUX LoRA: 512x512, batch 1, split_mode + fp8 — ~7.5GB VRAM
• Always use cache_latents_to_disk and cache_text_encoder_outputs_to_disk
• 32GB RAM is plenty for aggressive CPU offloading

Generation recommendations:

• SDXL: 768x768 with balanced offloading works well
• FLUX: 512x512-768x768 with model CPU offload
• Video (WAN 2.1): 480x320, 49 frames with aggressive offload
• Use balanced offload mode as default, switch to aggressive if OOM

Config Location

All settings, models, and outputs are stored in:

%USERPROFILE%\.ai_art_studio\
├── settings.json    # All app settings (auto-saved)
├── models/          # Trained LoRAs
├── outputs/         # Generated images/videos
├── datasets/        # Processed datasets
├── captions/        # Caption logs
├── cache/           # Latent cache
└── logs/            # Application logs