license: cc-by-nc-nd-4.0
pipeline_tag: text-to-image
widget:
- text: '-'
output:
url: ComfyUI_00641_.png
- text: '-'
output:
url: ComfyUI_00637_.png
- text: '-'
output:
url: ComfyUI_00623_.png
- text: '-'
output:
url: ComfyUI_00617_.png
- text: '-'
output:
url: ComfyUI_00615_.png
- text: '-'
parameters:
negative_prompt: >
bad quality, bad anatomy, worst quality, low quality, low resolution,
extra fingers, blur, blurry, ugly, wrong proportions, watermark, image
artifacts, lowres, ugly, jpeg artifacts, deformed, noisy image,
embedding:ac_neg1,
output:
url: ComfyUI_00614_.png
description This model is a fine-tuned version of Stable Diffusion 1.5, specifically enhanced for generating high-quality images of people, hands, and text. It has been trained on 131,000 high-quality, captioned image pairs generated using DALL-E 3. The training was conducted on four NVIDIA 3090 GPUs with NVLink over 16 hours, spanning 8 epochs.
The model demonstrates notable proficiency in rendering human figures and intricate details like hand gestures and written text, although it shows less effectiveness with animal imagery. This specialization makes it well-suited for applications requiring precise human and text representations.
The fine-tuning process involved 13,100 unique examples, contributing to a total dataset size of 131,000 images. Each training epoch processed 31,000 examples, with a total train batch size of 40. The model underwent a total of 26,200 optimization steps, maintaining a gradient accumulation of 1 throughout the training period.
The enhancements in this version aim to minimize common image generation flaws such as blurriness, disproportion, noise, and low resolution, ensuring clear and anatomically accurate outputs.