Add SDTurbo pipeline

[cyrildiagne]

Adds the SD Turbo pipeline (as demoed here):

• Adds a new pipeline SDTurboPipeline
• Adds a new scheduler EulerDiscreteScheduler
• Adds Tensor.sqrt & interp tensor utils
• Adds the pipeline to the react example

TODO:

• Find a better way to identify the SD Turbo model config

Note:
This PR doesn't include the usage of TAESD which greatly increases speed for a small quality downgrade (as applied in this demo). But it's not part of the original pipeline so it makes more sense to add it in a separate PR.

[jdp8]

Hello @cyrildiagne,

Sorry for taking so long but I was trying to implement the Img2Img pipeline of your SD-Turbo but I'm not getting the expected output. The input image that I'm using is the following:

This is more or less what I'm getting:

I added the addNoise() function (taken from here) to the scheduler like so:

addNoise (originalSamples: Tensor, noise: Tensor, timestep: number) {
    const sigma = this.sigmas.data[timestep]
    return originalSamples.add(noise.mul(sigma))
}

Also added code similar to the img2img pipeline that already exists in the Stable Diffusion pipeline, before the denoising loop like so:

if (input.img2imgFlag) {
      const inputImage = input.inputImage || new Float32Array()
      const strength = input.strength || 0.8

      await dispatchProgress(input.progressCallback, {
        status: ProgressStatus.EncodingImg2Img,
      })

      let imageLatent = await this.encodeImage(inputImage, input.width, input.height) // Encode image to latent space
      imageLatent = imageLatent.mul(this.scheduler.initNoiseSigma)

      // Taken from https://towardsdatascience.com/stable-diffusion-using-hugging-face-variations-of-stable-diffusion-56fd2ab7a265#2d1d
      const initTimestep = Math.round(input.numInferenceSteps * strength)
      const timestep = initTimestep
      latents = this.scheduler.addNoise(imageLatent, latents, timestep)
      // Computing the timestep to start the diffusion loop
      const tStart = Math.max(input.numInferenceSteps - initTimestep, 0)
      timesteps = timesteps.slice(tStart)
}

I've kept trying but I'm not completely sure what I'm doing wrong or what's missing to implement the SD-Turbo Img2Img pipeline. Do you have any idea what it could be?

I would appreciate any assistance with this. Thank you!