LoRA Training Parameters - Subject vs Style Training Differences (2025)

I wasted three weeks training LoRAs with wrong parameters before understanding that subject and style LoRAs need completely different approaches. Used style training settings for a character LoRA and got a model that barely recognized the character but applied weird artistic filters. Used subject settings for a style LoRA and it barely affected the output style at all.

The parameter differences aren't documented clearly anywhere. You either stumble onto working settings through trial and error or copy settings from someone who already figured it out. Here's what actually matters for each type.

Quick Answer: Subject LoRAs (characters, objects, specific people) require lower learning rates (0.0001-0.0003), more training epochs (15-30), higher network dimension (32-128), and diverse pose/angle training data to capture identity while allowing flexibility. Style LoRAs (artistic styles, aesthetics, techniques) need higher learning rates (0.0003-0.0008), fewer epochs (5-15), lower network dimension (16-64), and consistent style examples to learn artistic patterns without overfitting to content. Subject training prioritizes feature consistency, style training prioritizes pattern abstraction. Using wrong parameter sets for your LoRA type produces weak, overfitted, or ineffective results.

Understanding the Fundamental Difference

Before diving into parameters, understand what you're actually trying to achieve with each LoRA type.

Subject LoRAs teach the model to recognize and reproduce specific identities. A character's face, a particular person, a unique object design, a specific location. The goal is "this exact thing" regardless of pose, angle, or context. You want the model to identify the subject correctly while maintaining flexibility in how it's depicted.

Style LoRAs teach the model artistic patterns and techniques. A painter's style, an animation studio's aesthetic, a photography technique, a rendering approach. The goal is "this type of visual treatment" applied to any subject. You want the model to learn the style essence while applying it to different content.

The training tension differs completely. Subject LoRAs must memorize specific features while generalizing everything else. Style LoRAs must abstract patterns while avoiding memorizing specific subjects. These opposing requirements demand opposite parameter strategies.

Overfitting manifests differently. Overtrained subject LoRAs reproduce exact training poses and can't depict the subject differently. Overtrained style LoRAs reproduce training image content instead of just the style. Undertrained subject LoRAs don't recognize the subject reliably. Undertrained style LoRAs barely affect output style.

Success metrics differ too. Subject LoRA succeeds when you recognize the subject across varied depictions. Style LoRA succeeds when varied subjects all share the learned aesthetic. You're measuring different outcomes.

Understanding this philosophical difference explains why parameters diverge. You're solving different problems with the same training framework.

Learning Rate - The Make or Break Parameter

Learning rate is where most training failures happen. The right range depends entirely on LoRA type.

Subject LoRA learning rates sit in 0.0001 to 0.0003 range typically. Lower rates let the model slowly learn identifying features without overfitting to specific poses or compositions. The gradual learning captures "what makes this character recognizable" rather than "this exact image."

Start at 0.0002 for subject training. If the LoRA is too weak after full training, increase to 0.00025 for next attempt. If it's overfitting (only works in training poses), decrease to 0.00015. The adjustments are small because the range is narrow.

Style LoRA learning rates run 0.0003 to 0.0008, often higher than subject rates. Styles are patterns that need stronger influence to override the base model's defaults. Too low and the style barely applies. The higher rate lets style patterns imprint without needing excessive epochs.

Start at 0.0005 for style training. If style is too subtle, try 0.0006 or 0.0007. If you're getting content memorization instead of style abstraction, drop to 0.0004. The range is wider because style training is less sensitive to exact values.

Constant vs scheduled learning rates also split by type. Subject LoRAs often benefit from constant learning rate maintaining steady learning throughout training. Style LoRAs sometimes improve with cosine schedule that starts higher and decays, learning broad patterns early then refining.

Optimizer choice interacts with learning rate. AdamW8bit works well for both types with different rates. Prodigy optimizer auto-adjusts learning rate but tends toward subject-appropriate values, making it better for subject than style training.

Testing learning rates requires patience. Train a LoRA to completion at one rate, evaluate thoroughly, adjust for next training run. Don't try to adjust mid-training. Each training run to completion gives you data for optimizing the next attempt.

Signs of wrong learning rate appear in results. Too low for subjects means weak character identification. Too high means rigid overfitting. Too low for styles means barely visible effect. Too high means content memorization. The symptoms guide correction.

The learning rate is THE parameter that must match your LoRA type. Get this wrong and no amount of tweaking other parameters salvages the training.

Network Dimension and Alpha Settings

Network dimension controls LoRA capacity and behavior differently for each type.

Subject LoRA dimensions typically run 32, 64, or 128. Higher dimensions (64-128) capture complex subjects with lots of distinctive features better. Lower dimensions (32) work for simpler subjects or when you want the LoRA to apply more subtly.

A complex character with distinctive outfit, hairstyle, facial features, and accessories benefits from 64 or 128 dimension. Simple character or object can work at 32 dimension. The complexity of identifying features guides dimension choice.

Style LoRA dimensions usually sit lower at 16, 32, or 64. Style patterns are often learnable at lower dimensions. Going too high can cause content memorization. The abstraction goal doesn't need as much capacity as subject identification.

Painterly styles, photography techniques, or animation aesthetics work well at 32 dimension. Very distinctive or complex styles might justify 64. Rarely need 128 for pure style LoRAs.

Alpha settings conventionally equal dimension (32/32, 64/64) for both types. This isn't universal requirement but works reliably. The alpha controls how LoRA weights scale during application. Matching dimension simplifies tuning.

Higher dimension tradeoffs include larger file sizes, longer training times, and increased overfitting risk. Subject LoRAs justify these costs for complex subjects. Style LoRAs rarely do.

Testing dimension effects by training same dataset at different dimensions shows where diminishing returns hit. Often 64 captures subject or style adequately without needing 128's overhead.

Hardware constraints affect viable dimensions. 8GB VRAM struggles with 128 dimension training. 12GB handles it comfortably. 16GB+ doesn't care. Match dimension ambition to hardware reality.

The dimension choice is less critical than learning rate but still significantly impacts results. Start conservative (32 for styles, 64 for subjects), increase only if results justify it.

Training Epochs and When to Stop

How long to train varies dramatically between subject and style LoRAs.

Subject LoRA epochs typically run 15-30 depending on dataset size and complexity. Small datasets (10-15 images) need more epochs (25-30). Larger datasets (30-50 images) can work with fewer epochs (15-20). The model needs enough exposure to learn identifying features across the diversity.

Monitor preview images every few epochs. The character should become recognizable around epoch 8-12. Continue until features stabilize. If still improving at epoch 20, continue to 25 or 30. If degrading (overfitting), you went too far.

Style LoRA epochs run shorter at 5-15 typically. Styles imprint faster than subject identification. Too many epochs cause content memorization. Small style datasets (5-10 images) work at 8-12 epochs. Larger style collections (20-30 images) might only need 5-8 epochs.

Watch for style application without content bleeding. If preview images show the style but not the training image content, you're on track. If you're recreating training images instead of applying style to new content, you've overtrained.

Preview-based stopping is more reliable than arbitrary epoch counts. Generate preview images from prompts not in your training set every 2-3 epochs. Stop when quality peaks before degradation starts.

Diminishing returns point varies by LoRA type. Subject LoRAs improve gradually then plateau. Style LoRAs improve quickly then risk degrading. Recognizing the plateau/peak prevents wasted training time.

Multiple checkpoint saving lets you review training progression. Save every 5 epochs, test all checkpoints, pick the best. Often the optimal isn't the final epoch.

Dataset size interaction affects ideal epochs. This isn't linear - doubling dataset doesn't mean half the epochs. Experimentation with your specific subject or style finds the sweet spot.

The epoch count is something you dial in through experience with your datasets. Start with conservative estimates, extend if improving, stop when degrading.

Dataset Curation Differences

The training images you select matter as much as the parameters you set.

Subject dataset diversity is key. You want your subject in different poses, angles, expressions, and contexts. Training on 20 identical frontal portraits produces LoRA that only works for frontal portraits. Training on varied angles and poses produces flexible LoRA.

Include shots from front, sides, three-quarter views, slightly above, slightly below. Different expressions if it's a character. Different lighting if practical. The diversity teaches the model what stays constant (the identifying features) versus what's variable (everything else).

Style dataset consistency is opposite goal. You want consistent style application across the examples. Training a painting style on 20 images from one artist in their consistent technique works well. Training on 20 different art styles mixed together confuses the LoRA about what pattern to learn.

The subjects in style dataset should vary but the treatment should be consistent. Different characters/scenes painted in the same style. Different photographs processed with the same technique. The model learns "this treatment" abstracted from content.

Image quality standards matter more than quantity. 15 high-quality, well-chosen images beat 50 mediocre images for both types. Curate ruthlessly. Include only images that exemplify what you want the LoRA to learn.

Resolution considerations affect training. Higher resolution training images (768x768+) capture more detail but train slower and need more VRAM. Standard resolution (512x512) works fine for most LoRAs and trains faster.

Captioning strategy differs between types. Subject LoRAs benefit from detailed captions describing everything except the subject's identity. Style LoRAs benefit from minimal captions focusing on content variation while the style remains constant.

Problematic training data includes low quality images, heavily compressed images with artifacts, images with watermarks or text overlays, and images with elements you don't want learned. These corrupt training and should be excluded ruthlessly.

Augmentation considerations like flipping images can increase effective dataset size. Works fine for subjects if mirroring makes sense. Can confuse style learning if the style has directional aspects.

The dataset is where training success or failure is determined. Parameters can optimize but can't salvage bad data curation.

Batch Size and Training Efficiency

Batch size affects training speed and quality differently for each LoRA type.

Batch size fundamentals control how many images process before weight updates. Larger batches provide more stable gradient estimates but need more VRAM. Smaller batches update weights more frequently but with noisier gradients.

Subject LoRA batch sizes typically run 2-4. The diversity in subject datasets means smaller batches capture variation adequately. Batch size 2 works well for most subject training on 12GB VRAM cards. Batch size 4 is comfortable on 16GB+.

Style LoRA batch sizes can go slightly higher at 4-8 if VRAM allows. The consistency in style datasets means larger batches don't lose important variation. Batch size 4 on 12GB VRAM, 6-8 on 16GB+ works well.

VRAM optimization through gradient checkpointing or gradient accumulation lets you train effectively with less VRAM. Gradient accumulation simulates larger batch sizes by accumulating gradients over multiple forward passes before updating.

Training speed scales with batch size but hits diminishing returns. Doubling batch size doesn't double training speed. The optimal is largest batch size your VRAM comfortably handles without hitting OOM errors.

Quality effects of batch size are subtle for both types. Slightly smaller batches sometimes improve subject LoRA generalization. Slightly larger batches sometimes improve style LoRA consistency. The differences are minor versus other parameters.

Experimentation value is limited for batch size. Pick a value your hardware handles, stick with it. Optimizing batch size produces minimal improvements versus optimizing learning rate or dataset.

The batch size is more about hardware utilization than training optimization for most LoRA training scenarios.

Advanced Techniques for Each Type

Beyond basic parameters, advanced approaches improve results for experienced trainers.

Subject LoRA techniques include training multiple resolutions to capture detail at different scales, using regularization images to prevent overfitting, and training with varied CFG scales to improve flexibility.

Multiple subject aspects can train as separate concept within one LoRA. Character's face, outfit, and props as distinct tagged concepts. Allows mixing and matching aspects during inference.

Style LoRA techniques include training on preprocessed images emphasizing style elements, using style-focused captioning that describes artistic technique, and limiting epochs aggressively to prevent content bleeding.

Style strength calibration by training multiple checkpoints and selecting version with ideal strength. Too-weak style LoRA is unusable, too-strong overwhelms everything. Finding the sweet spot requires testing multiple training depths.

Hybrid LoRAs attempting to capture both subject and style rarely work well. The parameter conflicts usually produce mediocre results for both goals. Better to train separate LoRAs and combine during inference.

LoRA merging for combining multiple trained LoRAs happens outside training but affects strategy. If planning to merge, train individual LoRAs conservatively (lower strength) anticipating combination will amplify.

Platform-specific optimizations in tools like Kohya, OneTrainer, or cloud training services provide advanced options. Understanding your training platform's capabilities enables optimization beyond basic parameters.

The advanced techniques matter more for professional use or problematic cases. Most LoRAs succeed with proper basic parameter selection and good dataset curation.

Troubleshooting Common Training Failures

Recognizing failure modes helps diagnosis and correction.

Weak subject LoRA that barely affects output indicates learning rate too low, too few epochs, or dataset lacking consistency in identifying features. Solution is incrementally increasing learning rate or extending epochs, and reviewing dataset for consistency.

Overfitted subject LoRA that only reproduces training poses means learning rate too high or too many epochs. Features are learned but with too much context memorization. Solution is reducing learning rate and stopping training earlier in future runs.

Invisible style LoRA barely affecting output suggests learning rate too low, insufficient epochs, or dataset lacking consistent style. Solution is higher learning rate (try 0.0006-0.0008) and verifying dataset consistency.

Content-memorizing style LoRA that reproduces training images instead of applying style indicates too many epochs or insufficiently varied training subjects. Solution is stopping much earlier (try 8 epochs instead of 15) and improving dataset subject diversity.

Inconsistent results where LoRA works sometimes but not reliably suggests training instability. Try lower learning rate with more epochs for more stable learning. Or dataset has too much variation confusing the training.

File corruption or unusable LoRA files indicates training process errors or hardware issues. Verify sufficient disk space, stable power, no system crashes during training. Retry training and if problem persists, check for hardware issues.

Compatibility problems where LoRA doesn't work with intended base model means training on wrong base model. Verify you're training on the model you plan to use the LoRA with. LoRAs are model-specific.

Most training problems trace to learning rate misalignment with LoRA type or dataset quality issues. Fixing these two factors solves majority of failures.

Frequently Asked Questions

Can you use subject training parameters for style LoRAs and vice versa?

Technically yes but results will be poor. Subject parameters on style LoRAs produce barely visible style application. Style parameters on subject LoRAs cause overfitting and rigidity. Always match parameters to LoRA type for effective training.

How many training images do you actually need?

Subject LoRAs work with 15-30 quality images showing necessary variation. Style LoRAs can work with as few as 5-10 highly consistent examples. More isn't always better - 50 mediocre images often underperform 20 carefully curated ones. Quality and appropriateness matter more than quantity.

Can you train one LoRA that does both subject and style?

Not effectively in most cases. The parameter conflicts mean compromising both objectives. Better to train separate LoRAs and combine them during inference using both simultaneously. The combined approach produces better results than trying to train hybrid LoRA.

Does training on SDXL versus SD1.5 require different parameters?

Base model affects optimal parameters slightly but the subject/style distinction matters more than model architecture. SDXL generally trains similarly to SD1.5 for LoRAs with minor adjustments. The LoRA type remains the primary parameter determinant.

How do you know if your learning rate is correct?

Preview images during training show if learning rate works. Steady improvement without overfitting indicates good learning rate. Too-fast improvement then degradation suggests too high. Minimal improvement suggests too low. Empirical observation beats theoretical calculation.

Can you recover from wrong parameter choices mid-training?

Not really. Stop training, adjust parameters, start new training run. Trying to correct mid-training rarely works well. Each training run is test iteration for optimizing next attempt. Build knowledge through complete training cycles.

Do these parameter guidelines work for all training platforms?

The principles apply across platforms (Kohya, OneTrainer, Dreambooth) but specific values might need slight adjustment for different training implementations. Use these as starting points and calibrate for your specific platform and hardware.

What's more important - perfect parameters or perfect dataset?

Dataset quality matters more. Good dataset with okay parameters beats poor dataset with perfect parameters. Spend more time curating training images than obsessing over parameter optimization. Parameters can't fix fundamentally inadequate training data.

Parameter Cheat Sheet

Quick reference for proven parameter combinations.

Basic Subject LoRA (Character/Person):

• Learning rate: 0.0002
• Network dimension: 64
• Network alpha: 64
• Epochs: 20 (adjust based on dataset size)
• Batch size: 2-4
• Optimizer: AdamW8bit

Basic Style LoRA (Art Style/Aesthetic):

• Learning rate: 0.0005
• Network dimension: 32
• Network alpha: 32
• Epochs: 10 (adjust based on dataset size)
• Batch size: 4-6
• Optimizer: AdamW8bit

Complex Subject (Detailed Character/Object):

• Learning rate: 0.00015
• Network dimension: 128
• Network alpha: 128
• Epochs: 25-30
• Batch size: 2
• Optimizer: AdamW8bit

Subtle Style (Photography Technique/Light Style):

• Learning rate: 0.0004
• Network dimension: 16-32
• Epochs: 8-12
• Batch size: 4
• Optimizer: AdamW8bit

These are starting points, not absolute rules. Adjust based on your specific needs and observed results. The most important lesson is recognizing that subject and style LoRAs are different training tasks requiring different approaches. Stop using one-size-fits-all parameters and start matching parameters to your training objective.

Success in LoRA training comes from understanding what you're trying to achieve, curating appropriate training data, setting parameters matching your goal, and iterating based on results. The subject versus style distinction is the foundational split that determines everything else. Get this right and the rest follows naturally.