SVGDreamer: Text-Guided SVG Generation with Diffusion Model

SVGDreamer is an advanced text-to-SVG generation model that creates high-quality vector graphics using a multi-particle optimization approach. It generates multiple SVG variants simultaneously, allowing for diverse and creative outputs.

Model Description

SVGDreamer leverages Stable Diffusion to guide the generation of vector graphics through a novel multi-particle system. The model optimizes multiple SVG representations in parallel, enabling exploration of different artistic interpretations of the same text prompt.

Key Features

• Multi-Particle Generation: Creates multiple SVG variants simultaneously
• Style Control: Supports different artistic styles (iconography, pixel art, sketch, painting)
• High Quality: Produces detailed and aesthetically pleasing vector graphics
• Flexible Parameters: Extensive customization options for fine-tuning output

Usage

Direct API Call

import requests

API_URL = "https://api-inference.huggingface.co/models/jree423/svgdreamer"
headers = {"Authorization": "Bearer YOUR_HF_TOKEN"}

def query(payload):
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()

output = query({
    "inputs": "a majestic eagle soaring through clouds",
    "parameters": {
        "n_particle": 6,
        "num_iter": 1000,
        "guidance_scale": 7.5,
        "style": "iconography",
        "width": 224,
        "height": 224,
        "seed": 42
    }
})

Using the Inference Client

from huggingface_hub import InferenceClient

client = InferenceClient("jree423/svgdreamer")
result = client.post(
    json={
        "inputs": "a cyberpunk cityscape at night",
        "parameters": {
            "n_particle": 4,
            "style": "pixel_art",
            "guidance_scale": 8.0
        }
    }
)

Parameters

• n_particle (int, default: 6): Number of SVG particles to generate. Each particle represents a different interpretation of the prompt.
• num_iter (int, default: 1000): Number of optimization iterations. More iterations improve quality but take longer.
• guidance_scale (float, default: 7.5): Controls how closely the generation follows the text prompt.
• width (int, default: 224): Output SVG width in pixels.
• height (int, default: 224): Output SVG height in pixels.
• seed (int, default: 42): Random seed for reproducible results.
• style (string, default: "iconography"): Style of the generated SVG. Options: "iconography", "pixel_art", "sketch", "painting".

Output Format

The model returns a list of JSON objects, one for each particle, containing:

• particle_id: Unique identifier for the particle
• svg: The generated SVG content as a string
• svg_base64: Base64 encoded SVG for easy transmission
• prompt: The input text prompt
• style: The style used for generation
• parameters: The parameters used for generation

Styles

Iconography

Clean, minimalist vector graphics suitable for icons and logos.

• Example: "a simple house icon"

Pixel Art

Retro-style graphics with pixelated aesthetics.

• Example: "a pixel art character"

Sketch

Hand-drawn style with organic lines and artistic flair.

• Example: "a sketch of a mountain landscape"

Painting

Rich, painterly style with complex color gradients.

• Example: "an oil painting of a sunset"

Examples

Nature Scenes

• "a forest with tall pine trees"
• "ocean waves crashing on rocks"
• "a field of sunflowers under blue sky"

Characters and Objects

• "a friendly robot character"
• "a vintage bicycle"
• "a magical wizard casting spells"

Abstract Art

• "geometric patterns in bright colors"
• "flowing organic shapes"
• "mandala design with intricate details"

Technical Details

• Base Model: Stable Diffusion 2.1
• Framework: PyTorch + Diffusers
• Vector Rendering: DiffVG (differentiable vector graphics)
• Optimization: Multi-particle VPSD (Vector Particle-based Score Distillation)
• Parallel Processing: Simultaneous optimization of multiple SVG representations

Citation

@inproceedings{xing2024svgdreamer,
  title={SVGDreamer: Text Guided SVG Generation with Diffusion Model},
  author={Xing, XiMing and others},
  booktitle={CVPR},
  year={2024}
}

License

This model is released under the MIT License.