Spaces:
Running
on
Zero
A newer version of the Gradio SDK is available:
5.39.0
title: FLUX Fast & Furious
emoji: πΌπ
colorFrom: purple
colorTo: red
sdk: gradio
sdk_version: 5.35.0
app_file: app.py
pinned: false
license: openrail++
short_description: FLUX 8 Step Fast & High Quality Mode
I'll create comprehensive documentation for this FLUX Fast & Furious image generation code in both English and Korean.
English Documentation
FLUX: Fast & Furious - Hyper-Speed Image Generation
This application implements an accelerated version of the FLUX.1-dev image generation model, optimized by ByteDance's AutoML team using their Hyper-SD technology to achieve high-quality image generation in just 8 steps instead of the typical 20-50 steps.
Key Features
Hyper-Speed Generation
- Utilizes Hyper-SD LoRA (Low-Rank Adaptation) technology from ByteDance
- Reduces inference steps from 20-50 to just 6-25 steps (default: 8)
- Maintains high image quality while dramatically reducing generation time
- Optimized for CUDA with TF32 precision enabled for maximum performance
Neon-Themed User Interface
- Custom cyberpunk-inspired design with glowing neon effects
- Animated hover effects and dynamic visual feedback
- Dark theme with blue, cyan, and magenta color accents
- Responsive layout optimized for both desktop and mobile devices
User-Friendly Features
- Example Prompts: Five pre-written creative prompts covering various genres:
- Cyberpunk cityscapes
- Fantasy fairy scenes
- Epic dragon imagery
- Sci-fi space stations
- Underwater ancient cities
- Click-to-Use Examples: Simply click any example to instantly populate the prompt field
- Advanced Settings: Collapsible panel for fine-tuning generation parameters
- Example Prompts: Five pre-written creative prompts covering various genres:
Customizable Generation Parameters
- Image Dimensions: Adjustable width and height (256-1152 pixels)
- Inference Steps: Control speed vs. quality trade-off (6-25 steps)
- Guidance Scale: Adjust prompt adherence (0.0-5.0)
- Seed Control: Reproducible results with manual seed input
Technical Implementation
The application leverages cutting-edge technologies:
- FLUX.1-dev: State-of-the-art diffusion model from Black Forest Labs
- Hyper-SD LoRA: ByteDance's acceleration technology achieving 5-10x speedup
- BFloat16 Precision: Reduced memory usage while maintaining quality
- Gradio Spaces: GPU-accelerated deployment with automatic resource management
- Custom CSS: Neon-themed styling with glow effects and animations
The generation pipeline:
- Loads the base FLUX.1-dev model in bfloat16 precision
- Applies Hyper-SD LoRA weights with 0.125 scaling factor
- Fuses LoRA weights for optimal performance
- Generates images using accelerated inference with custom parameters
- Outputs high-quality 1024x1024 images (default) in seconds
Performance Optimization
- GPU Acceleration: Automatic CUDA optimization with @spaces.GPU decorator
- Memory Efficiency: BFloat16 precision reduces VRAM usage by 50%
- Inference Mode: Torch inference mode and autocast for maximum speed
- TF32 Support: Enabled for compatible GPUs for additional speedup
- Cached Models: Local model caching to reduce loading times
Use Cases
Perfect for:
- Rapid prototyping of visual concepts
- Creative exploration with instant feedback
- Production of high-quality images for various projects
- Testing different artistic styles and compositions
- Educational purposes to understand AI image generation
νκΈ μ€λͺ μ
FLUX: Fast & Furious - μ΄κ³ μ μ΄λ―Έμ§ μμ±κΈ°
μ΄ μ ν리μΌμ΄μ μ ByteDanceμ AutoML νμ΄ κ°λ°ν Hyper-SD κΈ°μ μ νμ©νμ¬ FLUX.1-dev μ΄λ―Έμ§ μμ± λͺ¨λΈμ κ°μνν λ²μ μΌλ‘, κΈ°μ‘΄ 20-50λ¨κ³κ° νμνλ κ³Όμ μ λ¨ 8λ¨κ³λ‘ μ€μ¬ κ³ νμ§ μ΄λ―Έμ§λ₯Ό μμ±ν©λλ€.
μ£Όμ κΈ°λ₯
μ΄κ³ μ μμ±
- ByteDanceμ Hyper-SD LoRA(Low-Rank Adaptation) κΈ°μ νμ©
- μΆλ‘ λ¨κ³λ₯Ό 20-50λ¨κ³μμ 6-25λ¨κ³λ‘ λν μΆμ (κΈ°λ³Έκ°: 8λ¨κ³)
- μμ± μκ°μ νκΈ°μ μΌλ‘ λ¨μΆνλ©΄μλ λμ μ΄λ―Έμ§ νμ§ μ μ§
- μ΅λ μ±λ₯μ μν TF32 μ λ°λκ° νμ±νλ CUDA μ΅μ ν
λ€μ¨ ν λ§ μ¬μ©μ μΈν°νμ΄μ€
- λ°κ΄ λ€μ¨ ν¨κ³Όκ° μ μ©λ μ¬μ΄λ²νν¬ μ€νμΌμ λ§μΆ€ν λμμΈ
- μ λλ©μ΄μ νΈλ² ν¨κ³Όμ λμ μκ° νΌλλ°±
- νλμ, μ²λ‘μ, λ§μ ν μμ μ μΌνΈκ° μλ λ€ν¬ ν λ§
- λ°μ€ν¬ν±κ³Ό λͺ¨λ°μΌ κΈ°κΈ° λͺ¨λμ μ΅μ νλ λ°μν λ μ΄μμ
μ¬μ©μ μΉνμ κΈ°λ₯
- μμ ν둬ννΈ: λ€μν μ₯λ₯΄λ₯Ό λ€λ£¨λ 5κ°μ μ°½μμ μΈ ν둬ννΈ μ 곡:
- μ¬μ΄λ²νν¬ λμ νκ²½
- ννμ§ μμ μ₯λ©΄
- μ μ₯ν λλκ³€ μ΄λ―Έμ§
- SF μ°μ£Ό μ κ±°μ₯
- μμ€ κ³ λ λμ
- ν΄λ¦νμ¬ μ¬μ©: μμλ₯Ό ν΄λ¦νλ©΄ μ¦μ ν둬ννΈ νλμ μ λ ₯
- κ³ κΈ μ€μ : μμ± λ§€κ°λ³μ λ―ΈμΈ μ‘°μ μ μν μ μ μ μλ ν¨λ
- μμ ν둬ννΈ: λ€μν μ₯λ₯΄λ₯Ό λ€λ£¨λ 5κ°μ μ°½μμ μΈ ν둬ννΈ μ 곡:
λ§μΆ€ν μμ± λ§€κ°λ³μ
- μ΄λ―Έμ§ ν¬κΈ°: μ‘°μ κ°λ₯ν λλΉμ λμ΄ (256-1152 ν½μ )
- μΆλ‘ λ¨κ³: μλ λ νμ§ κ· ν μ‘°μ (6-25λ¨κ³)
- κ°μ΄λμ€ μ€μΌμΌ: ν둬ννΈ μ€μλ μ‘°μ (0.0-5.0)
- μλ μ μ΄: μλ μλ μ λ ₯μΌλ‘ μ¬ν κ°λ₯ν κ²°κ³Ό
κΈ°μ μ ꡬν
μ ν리μΌμ΄μ μ μ΅μ²¨λ¨ κΈ°μ μ νμ©ν©λλ€:
- FLUX.1-dev: Black Forest Labsμ μ΅μ νμ° λͺ¨λΈ
- Hyper-SD LoRA: 5-10λ°° μλ ν₯μμ λ¬μ±νλ ByteDanceμ κ°μ κΈ°μ
- BFloat16 μ λ°λ: νμ§μ μ μ§νλ©΄μ λ©λͺ¨λ¦¬ μ¬μ©λ κ°μ
- Gradio Spaces: μλ 리μμ€ κ΄λ¦¬κ° ν¬ν¨λ GPU κ°μ λ°°ν¬
- 컀μ€ν CSS: λ°κ΄ ν¨κ³Όμ μ λλ©μ΄μ μ΄ μλ λ€μ¨ ν λ§ μ€νμΌλ§
μμ± νμ΄νλΌμΈ:
- bfloat16 μ λ°λλ‘ κΈ°λ³Έ FLUX.1-dev λͺ¨λΈ λ‘λ
- 0.125 μ€μΌμΌλ§ ν©ν°λ‘ Hyper-SD LoRA κ°μ€μΉ μ μ©
- μ΅μ μ±λ₯μ μν LoRA κ°μ€μΉ μ΅ν©
- μ¬μ©μ μ μ λ§€κ°λ³μλ‘ κ°μνλ μΆλ‘ μ μ¬μ©νμ¬ μ΄λ―Έμ§ μμ±
- λͺ μ΄ λ§μ κ³ νμ§ 1024x1024 μ΄λ―Έμ§(κΈ°λ³Έκ°) μΆλ ₯
μ±λ₯ μ΅μ ν
- GPU κ°μ: @spaces.GPU λ°μ½λ μ΄ν°λ‘ μλ CUDA μ΅μ ν
- λ©λͺ¨λ¦¬ ν¨μ¨μ±: BFloat16 μ λ°λλ‘ VRAM μ¬μ©λ 50% κ°μ
- μΆλ‘ λͺ¨λ: μ΅λ μλλ₯Ό μν Torch μΆλ‘ λͺ¨λμ μλ μΊμ€νΈ
- TF32 μ§μ: νΈν GPUμμ μΆκ° μλ ν₯μμ μν΄ νμ±ν
- μΊμλ λͺ¨λΈ: λ‘λ© μκ° λ¨μΆμ μν λ‘컬 λͺ¨λΈ μΊμ±
μ¬μ© μ¬λ‘
λ€μκ³Ό κ°μ μ©λμ μ ν©ν©λλ€:
- μκ°μ 컨μ μ μ μν νλ‘ν νμ΄ν
- μ¦κ°μ μΈ νΌλλ°±μΌλ‘ μ°½μμ νμ
- λ€μν νλ‘μ νΈλ₯Ό μν κ³ νμ§ μ΄λ―Έμ§ μ μ
- λ€μν μμ μ μ€νμΌκ³Ό κ΅¬μ± ν μ€νΈ
- AI μ΄λ―Έμ§ μμ± μ΄ν΄λ₯Ό μν κ΅μ‘ λͺ©μ