The refiner is unnecessary, and VAE is included.
Leaving the negative prompt empty generally brings about the best quality.
GOAL
Stable Diffusion XL has 6.6 billion parameters, which is about 6.6 times more than the SD v1.5 version. I believe that this is not just a number, but a number that can lead to a significant improvement in performance.
It has been a while since we realized that the overall performance of SD v1.5 has improved beyond imagination thanks to the explosive contributions of our community. Therefore, I am working on completing this AlbedoBase XL model in order to optimally reproduce the performance improvement that occurred in v1.5 in this XL version as well.
My goal is to directly test the performance of all Checkpoints and LoRAs that are publicly uploaded to Civitai, and merge only the resources that are judged to be optimal after passing through several filters. This will surpass the performance of image-generating AI of companies such as Midjourney.
As of now, AlbedoBase XL v1.3 has merged exactly 141 selected checkpoints and 251 LoRAs.
LOG
v2.0
I'd like to thank everyone who helped me on the AlbedoBase XL Pre side. Without you guys, the release date would probably have been much later. Thank you so much!
I have written a custom script to converge the existing AlbedoBase XL models into one. Intricately aligning the row and column weights of all U-NET and CLIP blocks according to a unique formula of mine.
If you encounter a bug in image generation (if nothing is generated), please switch to CLIP SKIP 2 or modify the prompt slightly! There may be combinations of prompts that CLIP does not recognize. In that case, you can change the order of words, use different words, or, most simply, change the CLIP SKIP. I will gradually work on resolving these issues in the future like v1.3.
The spec grid(403.5 MB): download
v1.3
In order to illustrate the quality associated with the model's randomness, I standardized the seed value at '9' for all showcase images intended for sampling and proceeded with their immediate generation.
Especially with this version, due to the significant impact of negative prompts, leaving the negative prompt field empty is likely to produce the nice quality.
The spec grid(438.7 MB): download
As you can see, as the number of Steps increases, it becomes available for all samplers, and the quality also improves.
Due to the effect of the LoRA I developed and merged, as described below, using sentence-form prompts rather than tag (a list of words) prompts is directly related to improving quality.
I merged 45 checkpoints and 7 LoRAs. After that, I merged AlbedoBase v0.4 and v0.3 in order, less than 0~5%, to reawaken the diluted merged models that had become outdated.
Among the 7 LoRAs, one is created by me. It involves analyzing and annotating captions for a total of 174 high-quality pictorial photos using GPT4-V. Merging this LoRA resulted in astonishingly clear images and an impressively excellent understanding of prompts.
My self-created LoRAs are exclusively available for purchase to my Ko-fi supporters at the Creative level or higher. I plan to release more and more updates in the future. The prices range from $10 to $50.