[User Survey] Selection of Future Model Formats

Update: Thanks to the advises of you. I tend to try lycoris (without pivotal), if it works well, i will use it. But if lycoris + pivotal performs much better than lycoris, i will use lycoris + pivotal instead.

First, here are some current situations:

• The newly released v1.5 version of the model has compatibility issues on some onsite generators, which have been reported by many users.

• On the other hand, although the A1111's webui already supports the pivotal tuning training method we use (which produces both a LoRA file and a pt file that need to be used together), other mainstream platforms or toolchains have yet to offer support (even though we've already contacted the development teams).

The above situations significantly hinder many users who like our models from using them properly and also prevent our models from being used on many toolchains. This seriously affects the dissemination of the model, contrary to the original intention of all the technological research and development efforts we have put in so far.

Therefore, we are facing an important decision—what kind of model format to use for training the character model in the future. We hereby solicit opinions from users, including but not limited to:

• LoRA (Pros: almost all tools and platforms can support it perfectly; Cons: the overall capability of the model may be average)

• LoRA + pivotal tuning (the current training method used; Pros: no need for technical changes, relatively less prone to overfitting; Cons: the training output contains two files, which many platforms and toolchains cannot support)

• Lycoris (formerly Locon) (Pros: stronger ability to learn details; Cons: better learning ability also means it's easier to overfit, and a small number of platforms and toolchains cannot support it)

• Lycoris + pivotal tuning (Pros: relatively less prone to overfitting compared to Lycoris alone; Cons: the training output contains two files, which many platforms and toolchains cannot support)

• Other training approaches (welcome to propose, but it is recommended to explain the pros and cons)

(Please note that all the abovementioned approaches are currently ensured to run properly on the A1111's webui.)

We look forward to real feedback from users, which is very important to us.

We will extensively understand user opinions and decide which training method to use based on a combination of subsequent technical experiments (the actual training effects of some training methods still need further confirmation).