The Happiness Throw, or Sophitia's grab in Soul Calibur.
Made this awhile back but I'm just going to upload now.
I made some other LoRA's but they were almost immediately taken down by CivitAI :(. Hopefully this one stays up.
Description
FAQ
Comments (5)
800megs is pretty big? could the size be reduced?
as if I know what I'm doing...
What could I do to reduce it without sacrificing quality?
I haven't made any Lora's myself, but I know the settings that are people used for training lora are baked into them and you can see them by hitting the info tab on them. Most for illust end up being about 200-300 megs (225 is the average).
My goggling resulted in this so I would compare this with other lora's on the site that are similar to see the settings that are used to train them:
To reduce LoRA file size in 2026, primarily lower the network_dim (rank) to 8–32, use network_alpha at half the rank, and train with --fp16 or --bf16 precision. Smaller dimensions significantly reduce parameter counts and file size, while techniques like 8-bit Adam optimizer reduce VRAM usage during training.
Key Methods to Reduce LoRA Size
Reduce Network Rank (network_dim): This is the most critical factor. Lower the rank (dimension) from default settings (often 64 or 128) to 16 or 32 for smaller files.
Adjust Network Alpha (network_alpha): Set the alpha to half the rank (e.g., Rank 32, Alpha 16) to maintain training stability while keeping files smaller.
Use fp16 or bf16 Precision: Train using 16-bit floating-point (half-precision) instead of 32-bit to immediately cut the file size in half.
Apply 8-bit Optimizers: Use --optimizer AdamW8bit or similar to reduce VRAM requirements during training, allowing for more efficient training runs that can still result in smaller, better-optimized models.
Target Specific Layers: Instead of training all layers, focus the training on specific Transformer layers (e.g., MLP/MoE).
Recommended Settings for Smaller LoRAs (2026)
Rank (Dim): 16–32 (Use 8-16 for simple character faces, 32-64 for complex subjects).
Alpha: 16 (or half of your rank).
Optimizer: AdamW8bit.
Data Prep: Crop images to uniform sizes (e.g.,
) to avoid inefficient bucket training.
Medium +3
Using these techniques in training tools like Kohya_ss or Unsloth will generate smaller, more efficient LoRA models.
@Arturito Also took me a while to re-find this but this is actually directly about file size to quality
https://civitai.com/articles/26599
in b4 Videl/Gohan