I renamed the model because it involves combining elements rather than just a simple union.
Krea2
This model was created by extracting and recombining specific elements from several of Krea2's anime models. As this is an experimental prototype, it is provided only in FP8 (Non-Scaled) format.
This model fuses anime-style elements with the characteristics of Kroma. Please use it with approximately 8 steps.
Anima
This model was created as an experiment by simply aggregating the learned content of various models (this test using "Anima" models, chosen for their lightweight nature). The process involves iteratively combining components: given definitions like Model A (AA + AB + Base) and Model B (BB + AB + Base), the system generates a composite of AA + AB + BB + Base (essentially Model A plus BB). During addition, duplicate elements are excluded so that only unique components are merged.
The diagram below illustrates the merging process. Although there was some backtracking along the way, the final result reflects this general workflow.
While mechanically combining learned content does not guarantee high-quality images, it does enable the model to handle a broader range of concepts. My findings suggest that, although certain models were better left out to avoid undesirable interference, the resulting images were less prone to structural collapse than anticipated, making this a highly viable technique. Although the test excluded photorealistic models, the specific models used were selected somewhat arbitrarily, leading to unexpected stylistic outcomes; this implies that the art style could be controlled through careful model selection. It should also be possible to achieve even greater flexibility and ease of use by gathering and combining components equivalent to the "BB" module.
Description
FAQ
Comments (7)
It's a lovely visual style and solid composition. Very visually pleasing.
Thanks!
I know you said it was experimental... but it is awesome, int8 convrot please!
This model was created using a technique that decomposes multiple models into subsets and then reconstructs them. Since all intermediate data was processed in FP8, this represents the highest level of precision currently available to me. While higher precision could be achieved by performing the same process in BF16, doing so would require far more time and resources than most people realize. Furthermore, since some of the source models were originally provided in 8-bit precision, simply increasing the processing precision would not yield the best possible results.
Above all, I do not intend for Krea2 to be anything more than an experimental project utilizing existing models (as I am not fond of the licensing terms). However, since I plan to use Krea2's output as training material for other types of models, I am releasing these models as a way to make a contribution to the community.
Is it true that this "Krea2" model can only generate images in the style of anime? Can't the realistic style of reality be utilized?
Since the base model is capable of handling both anime-style art and photorealistic images, it is certainly possible to generate images in other styles. However, because the model was reconstructed by aggregating anime-oriented elements, prompts requesting a photographic style will likely still result in an anime-like image (though not necessarily a "cel-shaded" look; attempting that with the current model tends to make it look excessively like Studio Ghibli art).
While a similar approach could be applied to photographic styles, I do not believe it is particularly necessary, given that Krea2 already offers a wide range of models designed for photorealism.
Regarding the model format: while the model is essentially converted to FP8, the txtfusion, tproj, and txtmlp components retain BF16 precision.