Logo Generator

AI-powered logo design tool that creates unique brand icons through interactive conversation or one-shot generation. Uses multiple image generation models via OpenRouter for diverse results.

Warning: This tool uses paid API credits via OpenRouter for both LLM prompt enhancement and image generation. Every chat message and every generated image costs credits. There is currently no free tier or local generation option. Make sure your OpenRouter account has sufficient credits before use.

Screenshots

Chat Designer

Interactive brand consultation with AI-generated logo results:

Settings

Configure LLM and image generation models:

Features

• Chat Designer - Conversational brand consultation: describe your brand, the AI asks clarifying questions, then generates optimized prompts and images
• Quick Generate - One-shot mode: describe a concept, get 4 diverse prompt variations generated across multiple models
• Multi-Model Generation - Same prompt runs through GPT-5 Image, Gemini Flash, and more for varied interpretations
• Seed Variation - Multiple seeds per model for even more diversity
• LLM Prompt Enhancement - Claude Sonnet 4 transforms simple descriptions into detailed, optimized image generation prompts
• Icon-Only Designs - Focused on unique symbols and abstract marks, no text/typography
• Upscaling - 4 methods: Lanczos, Bicubic, Real-ESRGAN Anime (best for flat logos), Real-ESRGAN General (best for complex logos). Supports 2x, 4x, 8x scales with GPU acceleration
• Background Removal - 5 rembg models (U2-Net, ISNet, BiRefNet, etc.) with alpha matting support. Reads from both generated and upscaled folders

Quick Start

# Clone and setup
git clone <repo-url>
cd logo-gen

# Create .env with your OpenRouter API key
echo "OPENROUTER_KEY=sk-or-v1-your-key-here" > .env

# Install and run
uv sync
uv run logo-gen

Opens at http://localhost:7860

Configuration

Edit settings in the UI (Settings tab) or via environment variables:

Variable	Default	Description
OPENROUTER_KEY	-	Your OpenRouter API key (required)
LLM_MODEL	anthropic/claude-sonnet-4	LLM for prompt enhancement
IMAGE_MODELS	GPT-5 Image Mini, Gemini 2.5/3.1 Flash	Image generation models
IMAGES_PER_MODEL	2	Seed variations per model per prompt
OUTPUT_DIR	output	Where generated images are saved

Available Image Models (OpenRouter)

• openai/gpt-5-image - Best quality, most expensive
• openai/gpt-5-image-mini - Good quality, cheaper
• google/gemini-2.5-flash-image - Gemini 2.5 Flash
• google/gemini-3.1-flash-image-preview - Gemini 3.1 Flash
• google/gemini-3-pro-image-preview - Gemini 3 Pro

Project Structure

src/logo_gen/
  config.py              # Settings (pydantic-settings, reads .env)
  prompt_engine.py       # LLM prompt enhancement & chat session
  generator.py           # Multi-model generation orchestrator
  upscaler.py            # Image upscaling (Lanczos, Bicubic, Real-ESRGAN)
  postprocess.py         # Background removal (rembg)
  app.py                 # Gradio web UI (5 tabs)
  clients/
    openrouter.py        # OpenRouter API client (LLM + image gen)

Output Structure

output/
  generated/   # AI-generated logos
  upscaled/    # Upscaled versions
  cleaned/     # Background-removed versions

How It Works

1. Prompt Enhancement - Your concept is sent to an LLM (Claude Sonnet 4) which generates 4-6 diverse, detailed image generation prompts, each exploring a different visual direction
2. Multi-Model Generation - Each prompt is sent to multiple image models with different random seeds
3. Upscaling - Optionally upscale to 2x/4x/8x using AI (Real-ESRGAN) or classical (Lanczos/Bicubic) methods
4. Background Removal - Remove backgrounds with your choice of 5 neural network models
5. Results - All generated logos are displayed in galleries for comparison

Requirements

• Python 3.12+
• uv package manager
• OpenRouter API key with credits

GPU / CPU Setup

The project works on both CPU and GPU. GPU is only needed for AI upscaling (Real-ESRGAN); everything else runs on CPU.

CPU-only (default)

uv sync
uv run logo-gen

This installs the CPU-only PyTorch build. AI upscaling methods (Real-ESRGAN) will fall back to Lanczos automatically. Lanczos and Bicubic upscaling work without GPU.

GPU (NVIDIA CUDA) - Recommended for upscaling

The project is configured to install CUDA-enabled PyTorch from the cu124 index. If uv sync installs the CPU version instead, force the CUDA build:

# Verify GPU is detected
uv run python -c "import torch; print('CUDA:', torch.cuda.is_available())"

# If False, reinstall torch with CUDA:
uv pip install torch --index-url https://download.pytorch.org/whl/cu124 --force-reinstall

Real-ESRGAN models are downloaded automatically on first use (~17MB for anime, ~67MB for general) and cached in ~/.cache/logo-gen/models/.

VRAM Requirements

Method	VRAM	Notes
Lanczos / Bicubic	0	CPU only
Real-ESRGAN Anime 6B	~1.5 GB	Best for flat/illustrative logos
Real-ESRGAN General x4	~2.5 GB	Best for complex logos

Images larger than 512x512 are processed in tiles to avoid VRAM overflow.