Trigger word

@Lora1111

Thank you for your comment.

We have conducted tests on the generation of LoRA outputs with more than 20 different types of models.

After achieving a satisfactory LoRA with all the models, we have adjusted the LoRA Block Weight. The Block Weight applied is 1,0,0,1,1,0,1,0,1,1,1,1,0.5,1,0,1,1.

From the LoRA posting page, both the image file used for training along with LoRA and the configuration file can be downloaded.

We hope the downloaded files will be of reference to you.

Does this answer your question?

Thanks.

@Lora1111 

Based upon my accumulated experience and insights, it is widely acknowledged that the optimal settings for dim/alpha generally range between dim:16-128 and alpha:1-64. Importantly, it has been recommended that the value of alpha should always be set at or below half of the corresponding dim to achieve the most favorable outcomes. This principle is consistent across not only the Concept LoRA but also various other types of LoRA. Consequently, configurations such as dim:200 or alpha:200 are considered to be beyond the recommended spectrum.

From my own analytical observations, I interpret the dim parameter as a metric that gauges the granularity of how images within a dataset are learned. When this parameter is configured to 128, there is a palpable sense that the images from the dataset are profoundly integrated into the learning process. On the contrary, a setting of 32 gives an impression of insufficient learning. The alpha parameter, in my perspective, acts as a pivotal threshold to modulate the intensity of learning initiated by dim. In the majority of scenarios, a value that lies between 8 and 16 for alpha is deemed apt.

In the process of crafting LoRA, I occasionally venture into using larger configurations, such as dim:196 coupled with alpha:128. This is particularly true when there is an underlying intention to accentuate a distinct concept. Admittedly, these adventurous configurations do not always lead to success, yet they offer invaluable insights into the intricate interplay between dim and alpha.

The selection of the optimizer_type holds paramount importance. The quintessential values for dim and alpha can exhibit variability contingent upon the optimizer_type in use. Presently, my preference gravitates towards DAdaptAdam, an optimizer characterized by its extended training duration but commendably mitigates the risk of overfitting. Furthermore, the ideal learning rate is intrinsically linked to the configurations of optimizer_type, dim, alpha, and epoch. In my recent endeavors, I have embarked on a comprehensive examination of learning rates, spanning from 5e-5 to 1.0.

Navigating the challenge of dominantly projecting dataset image characteristics onto diverse model styles remains an ongoing endeavor for me. One of the strategies I've adopted entails meticulous tagging adjustments during the training phase. For instance, when the dataset encompasses an anime-themed depiction of a "girl consuming a banana", I diligently incorporate tags such as anime and anime_coloring to the image caption, aiming to mitigate the dominance of that specific style. For images resonating with a real-life photographic essence, I bestow them with the real_photo tag, facilitating a clearer demarcation between disparate styles. In summation, if the objective is to extricate specific elements from LoRA, a prudent approach would be to intentionally blend datasets comprising those elements and judiciously annotate their captions with pertinent tags.

(I genuinely appreciate your patience in perusing this text. I would kindly draw your attention to the fact that this text has been meticulously translated for a comprehensive understanding in English, and there might be subtle nuances in interpretation.)