Follow me on X/Twitter for exclusive content: https://x.com/Citron_Legacy š
Check out all the Art styles Iāve made: https://civarchive.com/user/CitronLegacy/models?tag=style
Please support the artist by checking out their pageš: https://twitter.com/diivesart
Please check out the Illustrious version: https://civarchive.com/models/185388
Description
, <lora:DiivesStyle_ANIMA:1.0>
FAQ
Comments (6)
good to see you training stuff for Anima, i rly believe this is the best candidate to be illust successor so far.
Thanks š Agreed! I'm excited to see what Anima will bring.
I just reviewed your Anima lora it looks great.
Its nice to find a friend in the Anima party. I was hoping to see you making Anima loras and I'm very curious to see what you will think of it as you continue to create with it.
@CitronLegacyĀ having an llm for an encoder is still need time getting used to when training, other then that the model is a treat to prompt with xD
@navimixuĀ I remember reading something about that when I was setting up Kohya-ss for training.
How is the encoder impacting your training? Is there a benefit to training it?
I wanted to have this setting cache_text_encoder_outputs=True so that I use less VRAM, but if the encoder outputs are cashed then the encoder can't be trained.
I'm not limited by my VRAM so I guess I could train the encoder but I don't know if there is a benefit in training it.
I still have a lot to learn/adjust about my training process so I could be missing something.
--
I was working with Claude Sonnet 4.6 and I had it research options for me.
The following is copied exactly from the notes Claude generated.
Benefits:
- Significantly reduced VRAM usage - Text encoder is freed from GPU memory
- Faster training - Less computation per training step (no text encoder backward pass)
- Effective for most LoRA use cases - DiT-only LoRA provides excellent results
- Memory efficiency - Allows training with lower VRAM GPUs
- Recommended by documentation - Standard configuration for memory-constrained training
Disadvantages of training the encoder:
- Much higher VRAM usage - Keeps Qwen3-0.6B (600M params) loaded on GPU
- Slower training - Additional backward pass through text encoder each step
- Higher memory requirements - May not work on lower VRAM GPUs
- Rarely needed - Most LoRA applications don't benefit from text encoder modifications
@CitronLegacyĀ
@CitronLegacyĀ oh yes training the text encoder was vary beneficial on models using clip epically if you wanna change meaning of work or even make the lora more creative with less prompting , though i dont think its doing anything on ones with llm as a text encoder (looking at two quick experiments i made) , though im yet to try training a lora with a high step count/low learning rate , alot of unknowns here means even more tests xD
@navimixuĀ Oh I see what you mean! That aligns with some other things I've seen where some models allow for lazy prompting and then "magically" make something good.
I'll definitely try training with the encoder. High step count/low learning rate is a good idea!