Model Description for V1 Release
Version V1.0 - Official Release
I am excited to announce the release of V1.0 of the F-22 Raptor model! After extensive testing and refinement, this version incorporates valuable feedback and improvements based on the previous iterations.
Key Features:
Enhanced Dataset: This model is trained on a carefully curated dataset of 150 diverse images of the F-22 Raptor, providing a rich variety of contexts and angles to improve the quality of generated images.
Optimized for Quality: The model supports various aspect ratios, including non-default formats like 16:9, allowing for versatile image outputs. This is particularly beneficial for users looking to create high-quality visuals for posters or presentations.
Improved Text Generation: Incorporates posters and text images to enhance text generation capabilities, making it suitable for creative applications that require textual elements in the visuals.
Flexible Tokenization: Utilizes the default Flux tokenizer, which has shown to produce better results in my testing, ensuring that the model effectively understands prompts and generates accurate imagery.
Rich Style Variability: The model includes various styles and logos, enhancing its flexibility and enabling users to create unique designs tailored to their needs.
Quality Control Insights
To ensure the highest standards in this release, I have implemented rigorous quality control measures throughout the development process. For a detailed overview of the quality control protocols and testing methods used, I invite you to read the accompanying article. It provides insights into the evaluation criteria and the steps taken to enhance the model's performance.
Article click here: how i conduct flux lora quality control before release
Performance Note:
While this model demonstrates significant improvements, please remember that it is still in an early release phase. Your feedback is crucial for ongoing enhancements, so I encourage you to share your experiences in the comments.
Contact Information
For inquiries or to discuss commissioning model training, please reach out:
Discord Username: theaiking.2024
Email: [email protected]
Description
Not Recommended by the author its still in BETA V1 releases in a WEEK
the original model was saved in "float" and because of it the lora came out 3GB because of that i resized it which based on my testing has reduced the quality to a hit or miss point if you want the float version until V1 comes out comment and i will upload it for you
this was a test run on the new dataset and the new release of Kohya
76 image dataset
different aspect ratios included for better non-default AR generations like 16X9
Posters and text images included for better text generations
not captioned and only uses the default Flux tokenizer which in theory should provide better results and through my testing it has
different styles and even logo is included again for better flexibility
overall its a good model but again its beta so dont expect too much and please leave me your feedbacks in the comments