Photon - v1
精選圖片
推薦提示詞
A simple sentence in natural language describing the image.
推薦反向提示詞
cartoon, painting, illustration, (worst quality, low quality, normal quality:2)
cartoon, painting, illustration, (worst quality, low quality, normal quality:2)
推薦參數
samplers
steps
cfg
resolution
推薦高解析度參數
upscaler
upscale
steps
denoising strength
提示
Avoid using negative embeddings unless absolutely necessary.
Experiment by adding positive and negative tags and adjusting the settings for best results.
Future updates aim to eliminate the need for negative prompts and fix hand generation issues.
版本亮點
This is the initial release focused on achieving photorealistic results. However, please note that the model's hands generation may have some limitations at this stage, and I'm actively exploring training techniques to enhance it.
Photon aims to generate photorealistic and visually appealing images effortlessly.
Recommendation for generating the first image with Photon:
Prompt: A simple sentence in natural language describing the image.
Negative: "cartoon, painting, illustration, (worst quality, low quality, normal quality:2)"
Sampler: DPM++ 2M Karras | Steps: 20 | CFG Scale: 6
Size: 512x768 or 768x512
Hires.fix: R-ESRGAN 4x+ | Steps: 10 | Denoising: 0.45 | Upscale x 2
(avoid using negative embeddings unless absolutely necessary)
From this initial point, experiment by adding positive and negative tags and adjusting the settings. Most of the sample images follow this format.
Development
The development process was somewhat chaotic but essentially:
It started from an old mix.
LORAs were trained on various topics using AI-generated photorealistic images.
These LORAs were mixed within the model using different weights.
In the midst of this mixing, hand generation broke.
LORAs were generated and remixed in an attempt to fix hand generation (not entirely successful).
In future versions, I will try to:
Completely eliminate the need for a negative prompt to generate high-quality images.
Fix the hand generation issue to minimize instances of poorly drawn hands.
Explore more automated training processes. I would love to have 5,000 or 50,000 high-quality AI-generated photorealistic images for training purposes.
I hope you enjoy using it as much as I enjoyed creating it! If you have any questions or suggestions, feel free to share. Have fun creating amazing images! 🎨
