LoRA Trainer Guide

Overview

Help users pick the right training configuration and understand what the parameters actually do. Ships with 38 real-world presets covering SD1.5, SDXL, SD3, and Flux1 — standard LoRA, LyCORIS variants (LoHA, LoKR, LoCon, iA3), finetuning, and dreambooth.

The presets live in the presets/ folder as JSON files. They're directly usable with Kohya SS, sd-scripts, and the Ktiseos-Nyx-Trainer.

When to Use

• User wants to train a LoRA and needs help picking settings
• User has a preset and wants to understand what the params mean
• User is troubleshooting training results ("my outputs look fried, what do I change?")
• User wants to compare LoRA types (standard vs LoHA vs LoKR etc.)

Opening Behavior

Always start by asking: "How familiar are you with LoRA training?"

Adjust your responses based on their answer:

• New to it — explain concepts, define terms, walk them through choices step by step
• Done it a few times — skip basics, focus on "which preset fits your use case" and key params to tweak
• Experienced — be concise, jump straight to recommendations, reference specific params

Preset Selection Flow

Walk the user through these questions to find the right preset:

1. What model architecture?

Architecture	Resolution	Key differences
SD 1.5	512x512	Oldest, most presets available, clip_skip=2 common
SDXL	1024x1024	Dual text encoder, needs sdxl: true, more VRAM
SD3	Varies	Newer architecture, limited preset support
Flux1	512x512+	Requires fp8, T5 text encoder, different LoRA type

2. What training type?

Type	Use when	File size	Flexibility
Standard LoRA	General purpose, good balance of quality and size	Medium	High
LoHA	Easier concepts, multi-concept, want better generalization	Medium	Very high
LoKR (small, factor=-1)	Want tiny file size (<2.5MB)	Very small	Lower (model-specific)
LoKR (large, factor~8)	Want LoRA-like quality with Kronecker math	Medium-large	High
LoCon	Need convolution layer training specifically	Medium	Medium
iA3	Style learning, want ultra-tiny files (<1MB)	Tiny	Low (hard to transfer)
Finetune	Full model training, best possible quality	Very large	Highest
Dreambooth	Subject-specific training	Very large	High

3. Recommend a preset

Filter the presets/ folder by model_type and LoRA_type from the JSON files. Each preset has this structure:

{
  "name": "Human-readable name",
  "description": "What this preset is optimized for",
  "model_type": "SDXL",
  "config": {
    "LoRA_type": "Standard",
    "optimizer": "AdamW8bit",
    "network_dim": 32,
    "network_alpha": 32,
    "learning_rate": 0.0001,
    ...
  }
}

Present matching presets as numbered options with a brief explanation of what makes each one different (usually the optimizer and dim/alpha choices).

What to Tweak

After selecting a preset, highlight the params the user is most likely to want to adjust. Don't dump every parameter — focus on what actually matters.

Always worth checking

Parameter	What it does	Typical range	Watch out for
network_dim	LoRA rank — higher = more expressive, bigger file	8–128	Higher isn't always better; 32 is a solid default
network_alpha	Scales the learning rate relative to dim	1 to same-as-dim	alpha=1 vs alpha=dim behave VERY differently — see pitfalls
learning_rate	How fast the model learns	1e-5 to 1e-3	Depends heavily on optimizer choice
train_batch_size	Images per step	1–8	Higher = more VRAM, may need to adjust LR
max_train_epochs / max_train_steps	When to stop	Varies	More isn't better — watch for overtraining
optimizer	Learning algorithm	See reference.md	Each optimizer wants different LR ranges
save_every_n_epochs	Checkpoint frequency	1–5	Save often so you can pick the best epoch

Per-architecture specifics

SD 1.5:

• clip_skip: usually 2 for anime models, 1 for realistic
• max_resolution: 512,512
• mixed_precision: fp16 or bf16

SDXL:

• sdxl: must be true
• max_resolution: 1024,1024
• sdxl_no_half_vae: usually true (prevents NaN)
• clip_skip: usually 1
• Consider min_snr_gamma: 5 helps stabilize training

Flux1:

• fp8_base: must be true (Flux needs fp8)
• flux1_checkbox: true
• LoRA_type: must be "Flux1"
• t5xxl_max_token_length: 512
• timestep_sampling: "sigmoid"
• discrete_flow_shift: 3
• model_prediction_type: "raw"
• Cache text encoder outputs to disk (saves VRAM)

Common Pitfalls

Problem	Likely cause	Fix
Loss goes NaN	LoHA + high dim, or LR too high	Lower LR, reduce dim, or switch to LoRA
Output looks nothing like training data	Undertrained or LR too low	More epochs, higher LR, check captions
Output is copy-paste of training images	Overtrained	Fewer epochs, lower dim, use an earlier checkpoint
Style bleeds into everything	Model learned style too hard	Use LoHA (dampens style), lower dim, add regularization images
Colors are washed out / wrong	VAE issues on SDXL	Set sdxl_no_half_vae: true
Flux training OOMs immediately	Not using fp8	Set fp8_base: true, cache text encoder outputs
alpha=1 trains differently than expected	alpha scales effective LR	alpha=1 with dim=32 means effective LR is multiplied by 1/32. Set alpha=dim for "normal" behavior, or adjust LR accordingly
Increasing batch size makes results worse	LR wasn't scaled	When you increase batch size, increase LR proportionally

See Also

• reference.md in this skill folder — full parameter definitions, algorithm comparison tables, optimizer reference
• Presets are in presets/ — each is a self-contained JSON file ready to load