CommonArt-PoC

CommonArt is a text-to-image generation model with authorized images only. The architecture is based on DiT that is used by Stable Diffusion 3 and Sora.

How to Get Started with the Model

You can use this model by diffusers library.

import torch
from diffusers import Transformer2DModel, PixArtSigmaPipeline

device = "cpu"
weight_dtype = torch.float32

transformer = Transformer2DModel.from_pretrained(
    "alfredplpl/CommonArt-PoC", 
    torch_dtype=weight_dtype,
    use_safetensors=True,
)

pipe = PixArtSigmaPipeline.from_pretrained(
    "PixArt-alpha/pixart_sigma_sdxlvae_T5_diffusers",
    transformer=transformer,
    torch_dtype=weight_dtype,
    use_safetensors=True,
)

pipe.to(device)

prompt = " A picturesque photograph of a serene coastline, capturing the tranquility of a sunrise over the ocean. The image shows a wide expanse of gently rolling sandy beach, with clear, turquoise water stretching into the horizon. Seashells and pebbles are scattered along the shore, and the sun's rays create a golden hue on the water's surface. The distant outline of a lighthouse can be seen, adding to the quaint charm of the scene. The sky is painted with soft pastel colors of dawn, gradually transitioning from pink to blue, creating a sense of peacefulness and beauty."
image = pipe(prompt,guidance_scale=4.5,max_squence_length=512).images[0]
image.save("beach.png")

Model Details

Model Description

• Developed by: alfredplpl
• Funded by : alfredplpl
• Shared by : alfredplpl
• Model type: Diffusion transformer
• Language(s) (NLP): English
• License: Apache-2.0

Model Sources

• Repository: Pixart-Sigma
• Paper: PixArt-Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation

Uses

• Any purpose

Direct Use

• To develop commercial text-to-image generation.
• To research non-commercial text-to-image generation.

Out-of-Scope Use

• To generate misinformation.

Bias, Risks, and Limitations

• limited represantation

Training Details

Training Data

I used these dataset to train the transformer.

• CommonCatalog CC BY
• CommonCatalog CC BY Extention

Training Hyperparameters

• Training regime:

_base_ = ['../PixArt_xl2_internal.py']
data_root = "/mnt/my_raid/pixart"
image_list_json = ['data_info.json']

data = dict(
    type='InternalDataSigma', root='InternData', image_list_json=image_list_json, transform='default_train',
    load_vae_feat=False, load_t5_feat=False,
)
image_size = 256

# model setting
model = 'PixArt_XL_2'
mixed_precision = 'fp16'  # ['fp16', 'fp32', 'bf16']
fp32_attention = True
#load_from = "/mnt/my_raid/pixart/working/checkpoints/epoch_1_step_17500.pth"  # https://huggingface.co/PixArt-alpha/PixArt-Sigma
#resume_from = dict(checkpoint="/mnt/my_raid/pixart/working/checkpoints/epoch_37_step_62039.pth", load_ema=False, resume_optimizer=True, resume_lr_scheduler=True)
vae_pretrained = "output/pretrained_models/pixart_sigma_sdxlvae_T5_diffusers/vae"  # sdxl vae
multi_scale = False  # if use multiscale dataset model training
pe_interpolation = 0.5

# training setting
num_workers = 10
train_batch_size = 64  # 64 as default
num_epochs = 200  # 3
gradient_accumulation_steps = 1
grad_checkpointing = True
gradient_clip = 0.2
optimizer = dict(type='CAMEWrapper', lr=2e-5, weight_decay=0.0, betas=(0.9, 0.999, 0.9999), eps=(1e-30, 1e-16))
lr_schedule_args = dict(num_warmup_steps=1000)

#visualize=True
#train_sampling_steps = 3
#eval_sampling_steps = 3
log_interval = 20
save_model_epochs = 1
#save_model_steps = 2500
work_dir = 'output/debug'

# pixart-sigma
scale_factor = 0.13025
real_prompt_ratio = 0.5
model_max_length = 512
class_dropout_prob = 0.1

How to resume training

1. Download the model.
2. Set the model as "resume_from" model.

Environmental Impact

• Hardware Type: A6000x2
• Hours used: 700
• Compute Region: Japan
• Carbon Emitted: Not so much

Technical Specifications

Model Architecture and Objective

Diffusion Transformer

Compute Infrastructure

Desktop PC

Hardware

A6000x2

Software

Pixart-Sigma repository

Model Card Contact

alfredplpl