segmind/tiny-sd-mxfinetune

Text-to-image Distillation

This pipeline was distilled from SG161222/Realistic_Vision_V4.0 on a Subset of recastai/LAION-art-EN-improved-captions dataset. Below are some example images generated with the tiny-sd model.

This Pipeline is based upon the paper. Training Code can be found here.

Pipeline usage

You can use the pipeline like so:

from diffusers import DiffusionPipeline
import torch

pipeline = DiffusionPipeline.from_pretrained("segmind/tiny-sd-mxtune", torch_dtype=torch.float16)
prompt = "Portrait of a pretty girl"
image = pipeline(prompt).images[0]
image.save("my_image.png")

Training info

These are the key hyperparameters used during training:

• Steps: 125000
• Learning rate: 1e-4
• Batch size: 32
• Gradient accumulation steps: 4
• Image resolution: 512
• Mixed-precision: fp16

Finetune info

These are the key hyperparameters used during training:

• Steps: 8000 / 100000
• Learning rate: 1e-5
• Batch size: 24
• Gradient accumulation steps: 1
• Image resolution: 768
• Mixed-precision: fp16

Speed Comparision

We have observed that the distilled models are upto 80% faster than the Base SD1.5 Models. Below is a comparision on an A100 80GB.

Below is the code for benchmarking the speeds