Transparency Maker

A Canva app turns text prompts into high-quality transparent images of various styles, eliminating the need for manual background removal and saving time for designers, marketers, and content creators

Comment

๐ŸŽจ Inspiration

  • We found transparent images are vital for logos, icons, and web elements in graphic design.
  • Removing background apps are popular but do it manually is tedious, even with AI.
  • We aimed to streamline this by generating transparent images directly from text prompts with the latest diffusion model enhancement.

๐Ÿš€ What it does

  • AI-powered app generates high-quality transparent images from text prompt.
  • Eliminates the need for manual background removal.
  • Seamlessly integrates with Canva for easy use in design projects.
  • Align with Canva's content policy and industry standard when it comes to AI safety

๐Ÿ› ๏ธ How we built it

  • Used Latent Diffusion Models (LDMs) with SDXL/SD15 and LoRA for alpha channel encoding.
  • ComfyUI manages AI workflows with LayerDiffusion model.
  • Employed WebSocket for real-time communication between front-end and back-end.
  • AWS Lambda & SQS handles serverless scalable processing.
  • AWS Cloudfront CDN ensures efficient content delivery to global users.
  • Gen AI template comes with Canva SDK

๐Ÿšง Challenges

  • Converting ComfyUI workflow into a production-ready backend service.
  • Filtering NSFW content and ensuring generated images match prompts (The strict content safety rules of Canva review almost make us give up. And yes, even the official Canva Gen AI template on NPM wonโ€™t pass its own review standard for AI safety).
  • Implemented WebSocket API to avoid excessive API polling during long generation times.

๐ŸŽ‰ Accomplishments

  • Achieved professional-grade transparent image generation and efficient workflow.
  • Overcame strict Canva content moderation with advanced safety tools (OpenAI's toolkit, etc.).
  • Enhanced productivity and creativity while complying with Canvaโ€™s Gen AI Safety standards.

๐Ÿ“š What we learned

  • Mastered encoding transparency in image layers and developed robust NSFW filtering.
  • Learned to streamline ComfyUI workflows into production, speeding up future AI app development.

๐Ÿ”ฎ Whatโ€™s next

  • Integrate more powerful diffusion models to further improve image quality and style versatility.
  • Cost of keep the ComfyUI workflow running on GPU servers is still high, need to save budget on that

Built With

Share this project:

Updates