SD15-768 Open-Source Image Generation Model - Free Deployment, Stable Output in High Resolution, Create Images in Multiple Ratios Freely

SD15 768

Developed by panopstor

An image generation model fine-tuned based on the Stable Diffusion 1.5 framework, optimized for high-resolution output stability and supports multiple aspect ratio image generation

Text-to-Image EnglishOpen Source License:Openrail #High-resolution image generation #Multi-aspect ratio adaptation #Specialized for art photography and gaming

Downloads 43

Release Time : 11/6/2023

Model Overview

This model is specifically optimized for generating standard 768x768 resolution images or images with equivalent total pixels in various aspect ratios, suitable for content creation in fields such as art, photography, and video games

Model Features

High-resolution optimization

Optimized for high-resolution outputs like 768x768 through multi-stage training strategies, reducing subject repetition issues

Batch aspect ratio adaptation

Utilizes EveryDream2 trainer's batch aspect ratio adaptation technology to significantly reduce cropping artifacts

Multi-domain training data

Trained on 30,000 curated images covering diverse themes including art, photography, and video games

Training optimization strategy

Adopts phased training schedules (40 epochs) and cosine annealing learning rate strategies, combined with 8bit optimizer for enhanced training efficiency

Model Capabilities

Text-to-image generation

High-resolution image generation

Multi-aspect ratio image generation

Art style transfer

Use Cases

Digital art creation

Concept art generation

Generate game/film concept art based on text descriptions

High-quality 768x768 resolution concept art

Commercial design

Advertising material generation

Quickly generate visual materials that meet marketing needs

Commercial-grade images in various aspect ratios

🚀 Fine-Tuned Stable Diffusion 1.5

This project is a fine - tuned version of Stable Diffusion 1.5, generating high - quality text - to - image outputs at a nominal size of 768x768.

🚀 Quick Start

This model is a fine - tuned version of Stable Diffusion 1.5, leveraging the MSE VAE from RunwayML. It's designed to generate images at a nominal size of 768x768 or various aspects with a similar total pixel count.

✨ Features

High - Quality Training: Fine - tuned using EveryDream2 (https://github.com/victorchall/EveryDream2trainer) for 40 epochs across 4 sessions on 30,000 handpicked images from fine art, photography, and video games.
Reduced Cropping Artifacts: EveryDream2 trainer implements aspect ratio batch fitting, significantly reducing cropping artifacts.
Consistent High - Resolution Outputs: Compared to the original SD1.5 checkpoint, it provides more consistent high - resolution outputs.
Resumable Training: Optimizer states for the text encoder and Unet are provided in .pt form, allowing resumption in the EveryDream2 trainer software.

🔧 Technical Details

Training Process

Training Time: Approximately 60 hours on an RTX 6000 Ada 48GB.
Training Sizes and Batch Sizes: Initially trained at 512 nominal size with a batch size of 12, then 640 at batch size 12, and finally 768 at batch size 8 with 4 gradient accumulation steps.
Unet Weights Tuning: Tuned with bitsandbytes AdamW8bit with a learning rate of 1e - 6. A constant learning rate was used for the first 30 epochs and a cosine learning rate for the final 10.
Text Encoder Tuning: Tuned via backpropagation through the Unet using the same AdamW8bit optimizer with a learning rate of 2e - 7 cosine for each session and a weight decay of 0.040 to account for lower normal.

Output Comparison

The higher - resolution outputs of this fine - tuned model are more consistent compared to the original SD1.5 checkpoint, which often duplicates subject matter beyond the trained 512x512 resolution.

Resumable Training

Optimizer states for the text encoder and Unet are provided in .pt form. These can be loaded along with the Diffusers weights for resuming training in the EveryDream2 trainer software.

📄 License

This project is licensed under the creativeml - openrail - m license.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご