Sign In

GhostXL

1.2k
8.7k
332
Verified:
SafeTensor
Type
Checkpoint Trained
Stats
8,668
Reviews
Published
Feb 19, 2024
Base Model
SDXL 1.0
Trigger Words
realistic anime art style
realistic art style
Hash
AutoV2
EE1FDC86F7

GhostXL is an anime model. If you like GhostXL, please give it a 5 stars review. It is very inspiring for me to improve GhostXL.(中文版本在liblib和tusi都有)

Any questions or corporations? Find me at Discord or by Wechat: ghostinshell10. Thanks

Hello everyone, I am GhostInShell, the author of GhostMix. First and foremost, I would like to express my sincerest thanks to each and every user of GhostMix. With so many models out there, it is my honor that you guys choose and love GhostMix. Once again, thank you all. Secondly, I sincerely apologize for the long delay of GhostXL. Because I have high standards for the model and strive for perfection, a lot of time has been spent training and adjusting the model. Although GhostXL-V1.0 is not as good as I had imagined, it has a unique style and meets my standards in various aspects, therefore I decided to release it.Before using GhostXL, I recommend reading the user instructions below. Thank you.

User Instructions for GhostXL-V1.0

  1. After extensive testing, we found that GhostXL-V1.0 can directly generate images with high precision in the resolution range of 1024-1536, and there's no need for upsampling or refining operations.

  2. xx art style: The "xx art style" option allows for fine adjustments of the style. Here are some examples illustrating the effects of different art styles with the same prompts and seeds. (PS: If you want effects similar to GhostMix, the "realistic" or "realistic anime" art styles are personally recommended.)

  3. Personal observation: The compatibility of GhostXL seems slightly less than GhostMix, and its image quality might be a bit inferior, although it tends to produce a higher percentage of satisfactory results. (These comparisons are made against an upscaled version of GhostMix.) The primary reason for releasing GhostXL is the current market's lack of models with this particular style. On the bright side, GhostXL has an edge in producing immediate outputs, making upsampling unnecessary.

GhostXL Model Details

Analysis of the Pros and Cons of GhostXL-V1.0

As I mentioned in my previous articles, the main factors I consider when creating models are: 1) Compatibility; 2) Image quality; 3) Rate of generating satisfactory results.

  1. Compatibility: As a model that generates images in a 2D-2.5D style, GhostXL-V1.0 shows lesser compatibility with different styles when compared to GhostMix, especially for 3D styles. GhostXL is a finetune of Animagine XL V3, combined with features from other models (70% base from the finetuned model and 30% from the blend of others). As of the 1.0 version, GhostXL is still primarily a 2D-2.5D model, but I plan to keep developing it to see whether the style compatibility can be improved. For prompts compatibility, thanks to Animagine XL V3, GhostXL gives a good response to prompts such as 'masterpiece', 'best quality', as well as anime characters. However, it doesn't perform as well with more stylized prompts.

  2. Image Quality: Personally, I feel that the image quality is generally satisfactory, considering that GhostXL generates images in the range of 1024-1536 directly (one-stage generation), unlike GhostMix which generates in 512-768 and then undergoes upsampling (a two-stage generation process). I am quite satisfied with the quality it achieves using the one-stage approach.

  3. The Rate of Generating Satisfactory Results: In my experience, I believe GhostXL yields a higher rate of satisfactory results than GhostMix. However, for the same prompts, the variation in results is less broad in GhostXL relative to GhostMix. To increase diversity within the same theme, prompts may need to be adjusted.