TruBass版本v1.1 (text) (text) 综合资源合集综合资源合集

A sleeping giant rises from the depths...

Welcome all. This is a joint collaborative project between myself and https://civitai.com/user/speach1sdef178. I've been developing the TruBass model series, while she has been developing the Project 0 model series.

Periodically, as the development progressed on our individual models we would share the results with each other and each of us would take turns creating our own merge.

Working towards similar goals with differing perspectives, our respective versions of the model began to become both similar and unique.

All of the training that we have done, has been done using paramaters achievable on 24GB of VRAM, and much of it was trained online using Tensor Art's online training, which allows users to train on their own custom models. What wasn't trained on Tensor art was trained locally, using AI Toolkit.

Furthermore, until recently, the majority of our work has been focused on training LORA and using them, in both positive and negative weights to substantiate impactful changes which steer the model towards our intended goals.

Generally speaking we have succeeded in a number of areas, but we have also identified several key areas in which the currently available Flux models are significantly lacking. And frankly both of us have been struggling with how to proceed with implementing our vision without losing the general functionality of the model and reducing overall prompt adherence.

As this is a FLUX DEV model, it is necessary to state all of the changes made to it as part of the model license agreement for using and customizing it.

In this case, it can be summarized as follows:

We trained ~1000 styles into the model using our own captioning style for the mile high styler.
We then tested over 1000 styles and identified which styles were lacking.
We created synthetic datasets using the models own output, which reflected the worst mistakes of a particular style on a case by case basis.
We trained individual bad LORA using those bad datasets, one by one.
Then we follow this up by collecting and curating a small dataset of real-world data for the individual style. Which is then used to train a good lora.
The "bad" lora is then used in negative weight to remove unwanted elements, while the "good" lora is used both to return some of the lost weight to the model and to shape it more accurately towards your intended style output.
The resulting combination of positive and negative weighted lora is then merged with the model and saved as a checkpoint.
We repeated this process multiple times a day, for approximately 3 months. Periodically merging our works together and occasionally incorporating additional community models like ShuttleDiffusion, Crystal Clear Super, Jibmix and ArtsyDream for added context.
When we merged with a community model, it was necessary to re-apply the negative weighted lora to avoid it getting rid of my existing progress during the merge process.
In some cases it was also necessary to re-apply the positive weighted lora also.

The intention of this model is to create a replacement for the FLUX DEV model which is so much improved that for most cases additional LORA aren't needed. We would have liked to have added characters to the model, including celebrities. But fundementally it seems to be above our paygrade as simple model merging lora trainers.

This has been an extremely difficult endeavour despite overall being a simple process. Much of the difficulty has stemmed from the research elements and developing the system by which the model can be consistently trained on different datasets with the same paramaters. We've had to make several compromises because the model architecture is either unfixable, or requiring a total overhaul from scratch.

The latest version(s) of the model can be tested online using Tensor Art. And I will, as I develop and test, release them here on Civitai for public download.

https://tensor.art/models/816904519431515667

The reason for this being that the cost to train on Tensor is really much lower than Civitai. Where a model release can take several days, to a month or more to generate enough Buzz to train a follow-up model onsite, Tensors prices make it possible to train a new lora every single day, even if you aren't succesful at all. And the more success you have on the platform the more LORA you can train each day without having to spend a thing.

So in order to make sure that all of us get the benefit of free access, I'll keep the test releases exclusively online only on Tensor to fund the continued ongoing training process.

We fall ever deeper into the abyss as we seek the light.