Rapid text-to-image generation on-device – Google Research Blog

Text-to-image diffusion fashions have proven distinctive capabilities in producing high-quality photos from textual content prompts. However, main fashions characteristic billions of parameters and are consequently costly to run, requiring highly effective desktops or servers (e.g., Stable Diffusion, DALL·E, and Imagen). While current developments in inference options on Android through MediaPipe and iOS through Core ML have been made previously yr, fast (sub-second) text-to-image generation on cellular gadgets has remained out of attain.

To that finish, in “MobileDiffusion: Subsecond Text-to-Image Generation on Mobile Devices”, we introduce a novel strategy with the potential for fast text-to-image generation on-device. MobileDiffusion is an environment friendly latent diffusion mannequin particularly designed for cellular gadgets. We additionally undertake DiffusionGAN to realize one-step sampling throughout inference, which fine-tunes a pre-trained diffusion mannequin whereas leveraging a GAN to mannequin the denoising step. We have examined MobileDiffusion on iOS and Android premium gadgets, and it may possibly run in half a second to generate a 512×512 high-quality picture. Its comparably small mannequin measurement of simply 520M parameters makes it uniquely suited to cellular deployment.

Background

The relative inefficiency of text-to-image diffusion fashions arises from two main challenges. First, the inherent design of diffusion fashions requires iterative denoising to generate photos, necessitating a number of evaluations of the mannequin. Second, the complexity of the community structure in text-to-image diffusion fashions includes a considerable variety of parameters, recurrently reaching into the billions and leading to computationally costly evaluations. As a consequence, regardless of the potential advantages of deploying generative fashions on cellular gadgets, similar to enhancing person expertise and addressing rising privateness issues, it stays comparatively unexplored inside the present literature.

The optimization of inference effectivity in text-to-image diffusion fashions has been an energetic analysis space. Previous research predominantly focus on addressing the primary problem, in search of to cut back the variety of operate evaluations (NFEs). Leveraging superior numerical solvers (e.g., DPM) or distillation strategies (e.g., progressive distillation, consistency distillation), the variety of vital sampling steps have considerably diminished from a number of lots of to single digits. Some current strategies, like DiffusionGAN and Adversarial Diffusion Distillation, even scale back to a single vital step.

However, on cellular gadgets, even a small variety of analysis steps could be gradual because of the complexity of mannequin structure. Thus far, the architectural effectivity of text-to-image diffusion fashions has obtained comparatively much less consideration. A handful of earlier works briefly touches upon this matter, involving the elimination of redundant neural community blocks (e.g., SnapFusion). However, these efforts lack a complete evaluation of every element inside the mannequin structure, thereby falling wanting offering a holistic information for designing extremely environment friendly architectures.

MobileDiffusion

Effectively overcoming the challenges imposed by the restricted computational energy of cellular gadgets requires an in-depth and holistic exploration of the mannequin’s architectural effectivity. In pursuit of this goal, our analysis undertakes an in depth examination of every constituent and computational operation inside Stable Diffusion’s UNet structure. We current a complete information for crafting extremely environment friendly text-to-image diffusion fashions culminating within the MobileDiffusion.

The design of MobileDiffusion follows that of latent diffusion fashions. It accommodates three parts: a textual content encoder, a diffusion UNet, and a picture decoder. For the textual content encoder, we use CLIP-ViT/L14, which is a small mannequin (125M parameters) appropriate for cellular. We then flip our focus to the diffusion UNet and picture decoder.

Diffusion UNet

As illustrated within the determine under, diffusion UNets generally interleave transformer blocks and convolution blocks. We conduct a complete investigation of those two elementary constructing blocks. Throughout the research, we management the coaching pipeline (e.g., knowledge, optimizer) to review the consequences of various architectures.

In basic text-to-image diffusion fashions, a transformer block consists of a self-attention layer (SA) for modeling long-range dependencies amongst visible options, a cross-attention layer (CA) to seize interactions between textual content conditioning and visible options, and a feed-forward layer (FF) to post-process the output of consideration layers. These transformer blocks maintain a pivotal position in text-to-image diffusion fashions, serving as the first parts accountable for textual content comprehension. However, in addition they pose a big effectivity problem, given the computational expense of the eye operation, which is quadratic to the sequence size. We comply with the concept of UViT structure, which locations extra transformer blocks on the bottleneck of the UNet. This design alternative is motivated by the truth that the eye computation is much less resource-intensive on the bottleneck as a consequence of its decrease dimensionality.

Our UNet structure incorporates extra transformers within the center, and skips self-attention (SA) layers at increased resolutions.

Convolution blocks, particularly ResNet blocks, are deployed at every degree of the UNet. While these blocks are instrumental for characteristic extraction and knowledge stream, the related computational prices, particularly at high-resolution ranges, could be substantial. One confirmed strategy on this context is separable convolution. We noticed that changing common convolution layers with light-weight separable convolution layers within the deeper segments of the UNet yields comparable efficiency.

In the determine under, we evaluate the UNets of a number of diffusion fashions. Our MobileDiffusion reveals superior effectivity by way of FLOPs (floating-point operations) and variety of parameters.

Comparison of some diffusion UNets.

Image decoder

In addition to the UNet, we additionally optimized the picture decoder. We educated a variational autoencoder (VAE) to encode an RGB picture to an 8-channel latent variable, with 8× smaller spatial measurement of the picture. A latent variable could be decoded to a picture and will get 8× bigger in measurement. To additional improve effectivity, we design a light-weight decoder structure by pruning the unique’s width and depth. The ensuing light-weight decoder results in a big efficiency increase, with almost 50% latency enchancment and higher high quality. For extra particulars, please seek advice from our paper.

VAE reconstruction. Our VAE decoders have higher visible high quality than SD (Stable Diffusion).

One-step sampling

In addition to optimizing the mannequin structure, we undertake a DiffusionGAN hybrid to realize one-step sampling. Training DiffusionGAN hybrid fashions for text-to-image generation encounters a number of intricacies. Notably, the discriminator, a classifier distinguishing actual knowledge and generated knowledge, should make judgments primarily based on each texture and semantics. Moreover, the price of coaching text-to-image fashions could be extraordinarily excessive, significantly within the case of GAN-based fashions, the place the discriminator introduces further parameters. Purely GAN-based text-to-image fashions (e.g., StyleGAN-T, GigaGAN) confront comparable complexities, leading to extremely intricate and costly coaching.

To overcome these challenges, we use a pre-trained diffusion UNet to initialize the generator and discriminator. This design allows seamless initialization with the pre-trained diffusion mannequin. We postulate that the interior options inside the diffusion mannequin comprise wealthy data of the intricate interaction between textual and visible knowledge. This initialization technique considerably streamlines the coaching.

The determine under illustrates the coaching process. After initialization, a loud picture is shipped to the generator for one-step diffusion. The result’s evaluated in opposition to floor reality with a reconstruction loss, much like diffusion mannequin coaching. We then add noise to the output and ship it to the discriminator, whose result’s evaluated with a GAN loss, successfully adopting the GAN to mannequin a denoising step. By utilizing pre-trained weights to initialize the generator and the discriminator, the coaching turns into a fine-tuning course of, which converges in lower than 10K iterations.

Illustration of DiffusionGAN fine-tuning.

Results

Below we present instance photos generated by our MobileDiffusion with DiffusionGAN one-step sampling. With such a compact mannequin (520M parameters in complete), MobileDiffusion can generate high-quality various photos for numerous domains.

Images generated by our MobileDiffusion

We measured the efficiency of our MobileDiffusion on each iOS and Android gadgets, utilizing totally different runtime optimizers. The latency numbers are reported under. We see that MobileDiffusion could be very environment friendly and might run inside half a second to generate a 512×512 picture. This lightning pace probably allows many fascinating use instances on cellular gadgets.

Latency measurements (s) on cellular gadgets.

Conclusion

With superior effectivity by way of latency and measurement, MobileDiffusion has the potential to be a really pleasant choice for cellular deployments given its functionality to allow a fast picture generation expertise whereas typing textual content prompts. And we’ll guarantee any software of this expertise will probably be in-line with Google’s accountable AI practices.

Acknowledgments

We prefer to thank our collaborators and contributors that helped deliver MobileDiffusion to on-device: Zhisheng Xiao, Yanwu Xu, Jiuqiang Tang, Haolin Jia, Lutz Justen, Daniel Fenner, Ronald Wotzlaw, Jianing Wei, Raman Sarokin, Juhyun Lee, Andrei Kulik, Chuo-Ling Chang, and Matthias Grundmann.