These are LoRA versions of the anime-style models I have previously released. The "Ckpt_to_LoRA" tool can also be used with MOSS-Image. The LoRAs are divided into two separate models based on their layer ranges: layers 0–23 and layers 24–41. They are identified by the suffixes "L" and "U," respectively.
MOSS-FET01L2_LoRA.safetensors
MOSS-FET01U2_LoRA.safetensors
MOSS-FET02L2_LoRA.safetensors
MOSS-FET02U2_LoRA.safetensors
MOSS-FET03L2_LoRA.safetensors
MOSS-FET03U2_LoRA.safetensors
To be more specific, the model is split into two halves as follows:
L: 0–23 = 1.0, 24–39 = 0, 40–41 = 0.3
H: 0–23 = 0, 24–39 = 1.0, 40–41 = 0.7
Since the high-tier group appears to have a stronger influence, I have incorporated some overlap.
These were also utilized in the creation of MOSS-FET ver. 0.4.
The image above shows the result of applying the model to a base model with a strength setting of 1.5, in order to examine its specific characteristics.
Upon seeing "01U," I initially felt that this single model might be sufficient on its own; however, it is also possible to combine three different versions—each featuring distinct character styles—to achieve a blended effect.
As the example above illustrates, it appears that character-specific information and fine-grained rendering styles reside in the "upper" layers, while elements like composition are found in the "lower" layers.
I had somewhat anticipated this—having observed in the heatmaps that features like facial details were concentrated in the upper layers—but the roles of the upper and lower layers turned out to be even more inverted than I had expected. I verified this multiple times, eventually rebuilding everything from scratch to confirm. With the exception of version 0.3, the pattern holds: character details are in the upper layers, and composition is in the lower layers. Relying solely on the lower layers results in a somewhat blurry image. While it might actually make more sense to describe these as "inverted" (with the roles swapped), the processing order remains unchanged; it seems that the way these layers are utilized differs significantly from previous models.
Description
This is a LoRA version of a model from the same series as the anime-style models I have previously released. The Ckpt_to_LoRA tool can also be used with MOSS-Image. While all my previously published models were trained for only 1 epoch, this LoRA was created using a model trained for 3 epochs. The LoRA is split into two separate models based on layer ranges: layers 0–23 and layers 24–41, identified by the suffixes "L" and "U," respectively.
These were also used in the creation of MOSS-FET ver. 0.4.
The following image shows the result of applying these settings to a base model with a strength of 1.5, in order to observe their individual characteristics.
Upon seeing "01U," I initially felt that this might be all that’s needed; however, it is possible to combine three different types—each with distinct styles, such as those for specific characters—and use them together.
As the example above illustrates, it appears that character-specific information and fine-grained rendering styles reside in the "upper" layers, while elements like composition are found in the "lower" layers.
Based on the heatmaps—which showed that features like faces are located in the upper layers—I had some inkling of this, but the roles of the upper and lower layers turned out to be even more inverted than I had anticipated. I verified this multiple times, eventually rebuilding everything from scratch to confirm. With the exception of version 0.3, the pattern holds: characters are in the upper layers, and composition is in the lower layers. Using only the lower layers results in a somewhat blurry image. While it might feel more intuitive to describe these as "reversed" (top-to-bottom), the actual processing order remains unchanged; it seems that the way these layers are utilized differs significantly from previous models.