Lora that mimic the style of Regdic or Regudeku
Trained locally with my own tool, ComfyUI-AnimaTrainer (https://github.com/FR-Mister-T/ComfyUI-AnimaTrainer)
About V2.1 [Same as previous model in term of dataset, with about 25% more training steps, and a different trigger word that won’t trigger the pretrained style from base model]
Trigger word is r3gd1c for V2.1
The V2.1 is kind of experimental for me and the goal is to test model behavior and interaction between lora and the base training. It is also to learn how prompting weight and placement inside the prompt does affect the output.
Long story short and random information about prompting:
you can call upon base model knowledge with “@regdic”
you can call upon lora V2.1 knowledge with “r3gd1c”
You can stack both but result can be messy and need weight tuning
The Lora is better at giving the recent style from the artist than the base model
Moderate weighting on both Lora and trigger word is surprisingly effective and can help a lot, example
(r3gd1c:1.2) with lora at 1.1
Base model output @regdic
With lora at 1.0 and r3gd1c
with lora at 1.2 and (r3gd1c:1.3)
About V2 [Anima base V1, new training parameters, new images, new caption and tags, trained with regularization]
Regularization lengthen the training but allow great flexibility and improve promptability in case of lora stack or mixed style.
Weight from 0.4 to 1.8, you must use @regdic trigger word for effect, you can start low like 0.5 or 0.6 as the base model already know the artist, the lora act as a reinforcement
Example @regdic on base model
@regdic + lora @ 0.5
@regdic + lora @1.6 !
About V1B
Weight from 0.8 to 1.0 , mostly character focus , no trigger
I had hard time to decide which Lora stepping/epoch I should publish, I run into lot of small issues that force me to retrain it 3 times and to redo dataset. And some inconsistency in Inference were annoying as well. The lora supplied here is the most predictable for the moment imho.
I will revisit it later once I'll have built a better datasets and better training parameters. (check V2 is out)
For the time being, enjoy !
o7
Description
trigger word required @regdic.
New training strategy (1 set captioned in hybrid mode + 1 set captioned in tags),
400+ images datasets including landscape and some 2025-2026 content,
Prodigy with lower start training rate,
Training done with regularization.
trigger word required @regdic.
Base model already know about regdic with that trigger word, Lora act as reinforcement you can get improved styling while maintains great prompt ability down to 0.5 weight, you can crazy and push up to 1.8 weight, above that prompt is sensitive , quality collapse at 2.2.
trained locally on AMD GPU using https://github.com/FR-Mister-T/ComfyUI-AnimaTrainer