SDXL 0.9 with Emad, Applied ML Team & Amli on Discord!
1.8 هزار بار بازدید -
12 ماه پیش
-
This week's Stable Stage with
This week's Stable Stage with Emad Mostaque, Joe Penna, Alex Goodwin, Yam Levi & Amli of @Stability_AI
🚀 Summary: SDXL 0.9 Release and Updates:
- Releasing on API, Dream Studio, and Clip Drop
- Research weights sent to selected researchers
- Targeting full 1.0 release mid-July
- Creative ML license like original SD
- Continuous improvement after initial release
🌈 Differences from Previous Models:
- Better at handling prompts and following instructions
- Higher quality and more creative outputs
- Easier to communicate with in natural language
- Minimum effort produces equal or better results
💻 Hardware Requirements:
- 8GB VRAM sufficient for image generation and training
- Train full model on RTX 2070 (8GB VRAM) or better
- Goal: Optimize for accessibility on less powerful hardware
🔧 Workflow and Model Building Process:
-Experiment with different models and architectures
-Combine successful elements from "ugly" models
-Explore multiple directions, converge on solutions
-24/7 GPU testing for continuous progress
-Analyze loss graphs, samples, and try various techniques
-Team collaboration for experimentation and improvement
🌟 Plans after SDXL 1.0:
- Video models, new SD architectures
- Specialized models for photo-realistic, digital art, etc.
- Further optimization for efficiency and accessibility
- Provide tools and resources for building on the models
- Goal: Democratization of this groundbreaking technology
✨ Don't miss out on the future of generative AI imagery! Subscribe to our channel for exciting updates, artist spotlight, streams and more around open source AI ✨
🚀 Summary: SDXL 0.9 Release and Updates:
- Releasing on API, Dream Studio, and Clip Drop
- Research weights sent to selected researchers
- Targeting full 1.0 release mid-July
- Creative ML license like original SD
- Continuous improvement after initial release
🌈 Differences from Previous Models:
- Better at handling prompts and following instructions
- Higher quality and more creative outputs
- Easier to communicate with in natural language
- Minimum effort produces equal or better results
💻 Hardware Requirements:
- 8GB VRAM sufficient for image generation and training
- Train full model on RTX 2070 (8GB VRAM) or better
- Goal: Optimize for accessibility on less powerful hardware
🔧 Workflow and Model Building Process:
-Experiment with different models and architectures
-Combine successful elements from "ugly" models
-Explore multiple directions, converge on solutions
-24/7 GPU testing for continuous progress
-Analyze loss graphs, samples, and try various techniques
-Team collaboration for experimentation and improvement
🌟 Plans after SDXL 1.0:
- Video models, new SD architectures
- Specialized models for photo-realistic, digital art, etc.
- Further optimization for efficiency and accessibility
- Provide tools and resources for building on the models
- Goal: Democratization of this groundbreaking technology
✨ Don't miss out on the future of generative AI imagery! Subscribe to our channel for exciting updates, artist spotlight, streams and more around open source AI ✨
12 ماه پیش
در تاریخ 1402/04/06 منتشر شده
است.
1,894
بـار بازدید شده