Mann-E 4 Revision 0.1
Mann-E is a text to image model which has been developed by Muhammadreza Haghiri in order to be part of the Cognitive Web movement and projects. This is revision 0.1 of the 4th version of the model.
What does Mann-E mean?
It's a play with the name Mani, who was a Persian religious leader at the early Sassanian era and also a painter and he's famous for both his religious and artistic works. His artistic side was more considered for naming the model of course.
How to use the model
Colab
You can access Web UI colab through this link
Code
The following code is written for CUDA supported devices. If you use UI's or inference tools on other devices, you may need to tweak them in order to get them to the work. Otherwise, it will be fine.
First, you need to install required libraries:
pip3 install diffusers transformers scipy ftfy accelerate
NOTE: installation of accelerate
library makes the inference process amazingly faster. but it's totally optional.
Then, you need to import required libraries:
from diffusers import StableDiffusionPipeline, EulerDiscreteScheduler, DiffusionPipeline, DPMSolverMultistepScheduler
import torch
and then, create a pipeline (this pipeline is made with Euler Scheduler):
model_id = "mann-e/mann-e_4_rev-0-1"
scheduler = EulerDiscreteScheduler.from_pretrained(model_id, subfolder="scheduler")
pipe = StableDiffusionPipeline.from_pretrained(model_id, scheduler=scheduler, torch_dtype=torch.float16)
pipe = pipe.to("cuda")
and of course, since you may get NSFW filteration warnings even on simplest prompts, you may consider disabling it:
def dummy(images, **kwargs):
return images, False
pipe.safety_checker = dummy
NOTE: Please consider consequences of disabling this filter as well. we do not want people to get any sort of damage or injury from the image generation results.
And after that, you easily can start inference:
prompt = "Concept art of a hostile alien planet with unbreathable purple air and toxic clouds, sinister atmosphere, deep shadows, sharp details"
negative_prompt = "low quality, blurry"
width = 768
height = 512
then:
image = pipe(prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=100, width=width, height=height, guidance_scale=10).images[0]
image.save("My_image.png")