Note Introduces the idea of progressively distilling a diffusion model that requires fewer timesteps to sample from.

Note Combines the benefits of timestep distillation and faster ODE solver-based sampling. Custom diffusers pipeline: https://github.com/huggingface/diffusers/blob/main/examples/community/instaflow_one_step.py.

Note Introduces a new framework for distillation by learning to map any point in a probability flow ordinary differential equation (ODE )to its origin on the trajectory. CMs also allow for 1-4 step sampling. Play with them: https://huggingface.co/docs/diffusers/api/pipelines/consistency_models. Unconditional consistency model training: https://github.com/huggingface/diffusers/tree/main/examples/research_projects/consistency_training.

Note Extends Consistency Models by operating on the latent space. Also introduces Latent Consistency Fine-tuning for training on custom datasets. Play with them: https://huggingface.co/docs/diffusers/main/en/api/pipelines/latent_consistency_models Train your own: https://github.com/huggingface/diffusers/tree/main/examples/consistency_distillation

Note Builds on top of LCM and acts as a plugin for pre-trained Stable Diffusion models to enable faster few-step inference. Play with them: https://huggingface.co/docs/diffusers/main/en/using-diffusers/inference_with_lcm_lora Train your own: https://github.com/huggingface/diffusers/tree/main/examples/consistency_distillation

Note Enables one-step sampling through a GAN forward pass without running the reverse process. Community implementation: https://github.com/huggingface/diffusers/pull/6133.

Note Introduces a way to do GAN-style training to enable few-step inference. Combining GANs and Diffusion isn't new. Refer to this paper for references. This paper made the SD-Turbo and SDXL-Turbo possible. Play with them: https://huggingface.co/docs/diffusers/using-diffusers/sdxl_turbo. A training script is also being added here: https://github.com/huggingface/diffusers/pull/6303.

Note Successor of ADD (Adversarial Diffusion Distillation).

Note Popularly known as "BK-SDM". Up until now, we were mainly focusing on distilling to enable few-step inference. Those techniques necessarily don't concentrate on architectural compression. BK-SDM models (compatible with diffusers): https://huggingface.co/nota-ai

Note Builds on top of BK-SDM. This report presents an overview of the things that made SSD-1B and Vega family of architecturally compressed models work so well. Find the models: https://huggingface.co/Segmind

Note Combines progressive and adversarial distillation to achieve a balance between quality and mode coverage. Available checkpoints: https://huggingface.co/ByteDance/SDXL-Lightning (compatible with `diffusers`).

Note `diffusers` compatible implementation: https://hyper-sd.github.io/.

Note `diffusers` integrated implementation: https://research.nvidia.com/labs/toronto-ai/AlignYourSteps/.

Note It gradually merges redundant tokens, thereby accelerating inference. It supports diffusers: https://huggingface.co/docs/diffusers/main/en/optimization/tome.

Note It shows how the UNet features computed in the earlier reverse diffusion steps can be largely reused during the latter steps. This saves computation, thereby accelerating inference. It supports diffusers: https://huggingface.co/docs/diffusers/main/en/optimization/deepcache.

thibaud/sdxl_dpo_turbo

Text-to-Image • Updated Dec 20, 2023 • 233 • 83

Note What if you could combine two models coming from the same family? In this example, SDXL Turbo params and SDXL DPO params are averaged and we obtain a new model. This model enjoys alignment benefits from DPO and few-step sampling from Turbo.

Deci-early-access/DeciDiffusion-v2-0

Text-to-Image • Updated Jan 15 • 28 • 3

Note DeciDiffusion-v2 is a faster model than Stable Diffusion v1.5 while producing similar-caliber image quality. It benefits from an improved training recipe. Refer to the model card to know more.