Inspire AI
  • πŸ“–INTRODUCTION
    • What is Inspire AI?
    • Getting Started
    • Prompting
  • πŸ“£COMMAND REFERENCE
    • General
      • /commands
      • /styles
    • Text to Image
      • /square
      • /wide
      • /tall
      • /ultrawide
      • /phone
      • /reroll
      • /upscale
    • Styles
      • -inspire
      • -pixel
      • -lego
      • -ghibli
      • -manga
      • -coloringbook
      • -pixar
      • -anime
    • Text to GIF
      • /gifsquare
      • /gifwide
      • /giftall
  • βš™οΈTECHNICAL
    • Adversarial Diffusion Distillation
    • Hardware
  • πŸ—ΊοΈROADMAP
    • 🟒Phase 0
    • 🟒Phase 1
    • 🟑Phase 2
    • 🟑Phase 3
    • 🟑Phase 4
    • 🟑Phase 5
Powered by GitBook
On this page
  1. TECHNICAL

Adversarial Diffusion Distillation

PreviousTECHNICALNextHardware

Last updated 1 year ago

Inspire AI utilizes the latest and greatest in open-source technology. This is what grants us the unparalleled speed and accurate imagery it generates. Adversarial Diffusion Distillation (ADD) is a novel training approach introduced to efficiently sample large-scale foundational image diffusion models in just 1-4 steps while maintaining high image quality. This technique leverages large-scale off-the-shelf image diffusion models as a teacher signal, combined with an adversarial loss to ensure high image fidelity even in the low-step regime, such as one or two sampling steps.

The analyses demonstrate that ADD outperforms existing few-step methods, including GANs and Latent Consistency Models, in a single step and reaches the performance of state-of-the-art diffusion models, such as SDXL, in only four steps. ADD represents the first method capable of unlocking single-step, real-time image synthesis with foundation models​​.

The ADD-student is trained as a denoiser that receives diffused input images xs and outputs samples and optimizes two objectives: a) adversarial loss: the model aims to fool a discriminator which is trained to distinguish the generated samples from real images b) distillation loss: the model is trained to match the denoised targets xΛ†Οˆ of a frozen DM teacher.

Inspire AI's software backbone is in this technology. As it's continually worked and improved upon by the open-source community, we'll continually update our systems to utilize the most advanced technologies possible.

βš™οΈ
Source