Versions: LORA SD1.5 | LORA SD 2.1-768 | LYCORIS SD2.1-768
Zenkai System Explained: https://civarchive.com/articles/199
Beta1 sub models, models can be conditioned to provide additional flavours.
Zenkai Ultra Hermit released 30 April 2023
variation weights using my oldest successful recipeZenkai Ultra Shift TBA
variation using a new recipeZenkai Ultra Instinct TBA
variation using newest recipe
JDZ Beta1 - 22 April 2023
Beta1 is released for testing and evaluation. While convergence was reached, there are 230 styles/concepts contained within. This was trained on my original datasets used to create many CKPT / LORA already released on Civit. The Training Data contains combined instructions on making the most from this new versatile model format.
UPDATE - added new "combined-instructions.zip" now includes Instructions.txt and wildcards.zip. Wildcards are now included for use with Dynamic prompts.
JDZ-CaptionList.txt - ~5000 Example prompts used in captioning the 230 sets
JDZ-StyleTriggers.txt - ~230 "Trigger words" used for each style
JDZenkai-GenPrompts2200.txt - ~4200 "GPT2 Generated Prompts" trained on the Caption List
(Above pictured) Wildcard files now included
Future releases will be trained to Epoch 40 (~200,000 steps)
The Beta1 model is currently trained to Epoch 19 (~97000 steps.
There are many ways to use this model as it contained all my CKPT styles in one model.
You do not need to use Triggers and can prompt normally.
You can prompt using words listed in DJZ-Styletriggers.txt to anchor to a specific style, or use multiple triggers with attention weighting to control the output.
You can use one or more prompts as shown in the DJZ-CaptionList.txt.
You can use the DJZ-CaptionList.txt and/or JDZenkai-GenPrompts2200.txt as a wildcard with Dynamic Prompts, simply place it into the /wildcards/ folder and rename it to your choosing.
I trained a GPT2 Generator which created many variations on the Example Prompts, an example of these are the JDZenkai-GenPrompts2200.txt
This is an ongoing project so expect a LORA version next. It will take some time to create the LORA version due to a difference in the captioning. This is trained using a Parallel Dreambooth Method which prevents concept/style bleeding, but you can still mix by interpolating trained terms/triggers in the prompt.
TLDR there is a missing comma, can you find it?
Description
FAQ
Comments (2)
There are too many version of the model, which one should I choose? Do you have a suggestion?
yes this is an ongoing research project. " Beta1 " version is cleanest currently, UltraHermit and Shift are variants/derivative models from Beta1.
Even Pegasus (mutation shift) and Chimera Beta2 (Revision Learning) are derivatives that trained in 75% more data on top of Beta1, however using different methodology for each.
This makes the Beta1 model the cleanest representation of my work, however only contains 230 of my models (all in one)
That being said, The Lycoris Versions are the most advanced project right now. Johnson-Desu-Zenkai for 1.5 is using the Beta1 dataset, 230 models. The newest Lycoris is only available for 2.1-768 at this time, it is a full retrain with 350 model datasets. That is Johnson-Desu-Zenkai-Beta2 for 2.1
Lycoris requires an extension for Automatic which is built in with some webUI. However if the tool you are using does not support the Lycoris models, you can also use classic the provided LORA versions that were trained on 1.5 and 2.1-768.
TLDR, if you don't use Lora/Lycoris, the Beta1 Checkpoint version is the best. Chimera and Pegasus are research models built on Beta1