Awesome

SegMoE: Segmind Mixture of Diffusion Experts

image/png

SegMoE is a powerful framework for dynamically combining Stable Diffusion Models into a Mixture of Experts within minutes without training. The framework allows for creation of larger models on the fly which offer larger knowledge, better adherence and better image quality. It is inspired by mergekit's mixtral branch but for Stable Diffusion models.

Installation

pip install segmoe

Usage

Load Checkpoint from Hugging Face

We release 3 merges on Hugging Face,

SegMoE 2x1 has two expert models.
SegMoE 4x2 has four expert models.
SegMoE SD 4x2 has four Stable Diffusion 1.5 expert models.

They can be loaded as follows:

from segmoe import SegMoEPipeline

pipeline = SegMoEPipeline("segmind/SegMoE-4x2-v0", device = "cuda")

prompt = "cosmic canvas, orange city background, painting of a chubby cat"
negative_prompt = "nsfw, bad quality, worse quality"
img = pipeline(
    prompt=prompt,
    negative_prompt=negative_prompt,
    height=1024,
    width=1024,
    num_inference_steps=25,
    guidance_scale=7.5,
).images[0]
img.save("image.png")

Comparison

The Prompt Understanding seems to improve as shown in the images below. From Left to Right SegMoE-2x1-v0, SegMoE-4x2-v0, Base Model (RealVisXL_V3.0)

<div align="center">three green glass bottles</div> <br>

<div align="center">panda bear with aviator glasses on its head</div> <br>

<div align="center">the statue of Liberty next to the Washington Monument</div>

Creating your Own Model

Create a yaml config file, config.yaml, with the following structure:

base_model: Base Model Path, Model Card or CivitAI Download Link
num_experts: Number of experts to use
moe_layers: Type of Layers to Mix (can be "ff", "attn" or "all"). Defaults to "attn"
num_experts_per_tok: Number of Experts to use 
type: Type of the individual models (can be "sd" or "sdxl"). Defaults to "sdxl"
experts:
  - source_model: Expert 1 Path, Model Card or CivitAI Download Link
    positive_prompt: Positive Prompt for computing gate weights
    negative_prompt: Negative Prompt for computing gate weights
  - source_model: Expert 2 Path, Model Card or CivitAI Download Link
    positive_prompt: Positive Prompt for computing gate weights
    negative_prompt: Negative Prompt for computing gate weights
  - source_model: Expert 3 Path, Model Card or CivitAI Download Link
    positive_prompt: Positive Prompt for computing gate weights
    negative_prompt: Negative Prompt for computing gate weights
  - source_model: Expert 4 Path, Model Card or CivitAI Download Link
    positive_prompt: Positive Prompt for computing gate weights
    negative_prompt: Negative Prompt for computing gate weights

Any number of models can be combined, An Example config can be found here. For detailed information on how to create a config file, please refer to the Config Parameters

Note Both Huggingface Models and CivitAI Models are supported. For CivitAI models, paste the download link of the model, For Example: "https://civitai.com/api/download/models/239306"

Then run the following command:

segmoe config.yaml segmoe_v0

This will create a folder called segmoe_v0 with the following structure:

├── model_index.json
├── scheduler
│   └── scheduler_config.json
├── text_encoder
│   ├── config.json
│   └── model.safetensors
├── text_encoder_2
│   ├── config.json
│   └── model.safetensors
├── tokenizer
│   ├── merges.txt
│   ├── special_tokens_map.json
│   ├── tokenizer_config.json
│   └── vocab.json
├── tokenizer_2
│   ├── merges.txt
│   ├── special_tokens_map.json
│   ├── tokenizer_config.json
│   └── vocab.json
├── unet
│   ├── config.json
│   └── diffusion_pytorch_model.safetensors
└──vae
    ├── config.json
    └── diffusion_pytorch_model.safetensors

Alternatively, you can also use the following command to create a mixture of experts model:


from segmoe import SegMoEPipeline

pipeline = SegMoEPipeline("config.yaml", device="cuda")

pipeline.save_pretrained("segmoe_v0")

Push to Hub

The Model can be pushed to the hub via the huggingface-cli

huggingface-cli upload segmind/segmoe_v0 ./segmoe_v0

Detailed usage can be found here

SDXL Turbo

To use SDXL Turbo style models, just change the scheduler to DPMSolverMultistepScheduler. Example config can be found here

Usage:


from segmoe import SegMoEPipeline

pipeline = SegMoEPipeline("segomoe_config_turbo.yaml", device = "cuda")
pipeline.pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.pipe.scheduler.config)

prompt = "cosmic canvas, orange city background, painting of a chubby cat"

image = pipe(prompt=prompt, num_inference_steps=6, guidance_scale=2).images[0]  

image.save("image.png")

Stable Diffusion 1.5 Support

Stable Diffusion 1.5 Models are also supported and work natively. Example config can be found here

Note: Stable Diffusion 1.5 Models can be combined with other SD1.5 Models only.

Other Tasks

Our Framework is tightly integrated with the Diffusers package which allows the use of AutoPipelineForImage2Image, AutoPipelineForInpainting and any other pipeline which supports the from_pipe method.

Image to Image

Here is example code for Image to Image generation:

from segmoe import SegMoEPipeline
from diffusers import AutoPipelineForImage2Image
t2i = SegMoEPipeline("segmind/SegMoE-SD-4x2-v0")

prompt = "cosmic canvas,  orange city background, painting of a chubby cat"
negative_prompt = "nsfw, bad quality, worse quality"
img = t2i(
    prompt=prompt,
    negative_prompt=negative_prompt,
    height=1024,
    width=1024,
    num_inference_steps=25,
    guidance_scale=7.5,
).images[0]
img.save("base_image.png")
pipeline = AutoPipelineForImage2Image.from_pipe(t2i.pipe)
prompt = "cosmic canvas,  orange city background, painting of a dog"
image = pipeline(prompt, img).images[0]
image.save("changed_image.png")

Inpainting

Here is example code for Inpainting:

from segmoe import SegMoEPipeline
from diffusers import AutoPipelineForInpainting

t2i = SegMoEPipeline("segmind/SegMoE-SD-4x2-v0")
pipeline = AutoPipelineForInpainting.from_pipe(t2i.pipe)
image = pipeline(prompt, image=init_image, mask_image=mask_image).images[0]
image.save("inpainted_image.png")

Memory Requirements

SDXL 2xN : 19GB
SDXL 4xN : 25GB
SD1.5 4xN : 7GB

Advantages

Benefits from The Knowledge of Several Finetuned Experts
Training Free
Better Adaptability to Data
Model Can be upgraded by using a better finetuned model as one of the experts.

Limitations

Though the Model improves upon the fidelity of images as well as adherence, it does not be drastically better than any one expert without training and relies on the knowledge of the experts.
This is not yet optimized for speed.
The framework is not yet optimized for memory usage.

Research Roadmap

Config Parameters

Base Model

The base model is the model that will be used to generate the initial image. It can be a Huggingface model card, a CivitAI model download link or a local path to a safetensors file.

Number of Experts

The number of experts to use in the mixture of experts model. The number of experts must be greater than 1. The Number of experts can be anything greater than 2 as long as the GPU fits it.

MOE Layers

The type of layers to mix. Can be "ff", "attn" or "all". Defaults to "attn". "ff" merges only the feedforward layers, "attn" merges only the attention layers and "all" merges all layers.

Type

The type of the models to mix. Can be "sd" or "sdxl". Defaults to "sdxl".

Experts

The Experts are the models that will be used to generate the final image. Each expert must have a source model, a positive prompt and a negative prompt. The source model can be a Huggingface model card, a CivitAI model download link or a local path to a safetensors file. The positive prompt and negative prompt are the prompts that will be used to compute the gate weights for each expert and impact the quality of the final model, choose these carefully.

Citation

@misc{segmoe,
  author = {Yatharth Gupta, Vishnu V Jaddipal, Harish Prabhala},
  title = {SegMoE: Segmind Mixture of Diffusion Experts},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/segmind/segmoe}}
}