LoRA Training Complete Guide 2025 - How Many Headshots and Body Shots Do You Really Need?

You're ready to train your first character LoRA, but the internet gives you wildly conflicting advice. Some tutorials say 5-10 images is enough, others demand 200+. Nobody agrees on how many should be headshots versus full body shots. And what if you want to train a LoRA that handles both SFW and NSFW content?

After testing dozens of training runs with datasets ranging from 20 to 200+ images, clear patterns emerge about what actually works. The truth? Dataset size and composition matter enormously, but the optimal configuration depends entirely on what you want your LoRA to do.

This guide cuts through the confusion with tested, real-world strategies for building LoRA training datasets that produce consistent, high-quality results. For using your trained LoRAs in ComfyUI workflows, see our ComfyUI basics guide and essential custom nodes.

Understanding LoRA Training Fundamentals - What Actually Matters

Before diving into dataset specifics, understanding what LoRAs are learning helps you make informed decisions about training data composition.

What LoRAs Actually Learn: LoRAs (Low-Rank Adaptations) learn to modify base model outputs by capturing patterns specific to your training data. They're learning facial features, body proportions, clothing styles, lighting preferences, and artistic characteristics present in your dataset.

The more consistently these patterns appear, the better the LoRA captures and reproduces them.

Why Dataset Composition Matters:

The Overfitting Problem: Too many similar images cause overfitting - the LoRA memorizes specific photos rather than learning general character features. This creates problems when you try to generate scenes different from your training data.

Diversity in angles, lighting, and context prevents overfitting while maintaining character consistency.

Quality vs Quantity: Ten high-quality, well-composed, varied images outperform fifty nearly identical selfies. Quality, variety, and consistency matter more than raw image count.

This doesn't mean more images can't help - it means throwing random images at training won't produce better results.

Training Time and Resources:

Understanding these resource requirements helps you plan training runs effectively. If you're working with limited VRAM, see our complete low-VRAM survival guide for optimization strategies.

For users who want excellent LoRAs without managing training infrastructure, platforms like Apatero.com provide streamlined training interfaces with automatic optimization.

The Tested Formula - Dataset Sizes That Actually Work

Based on extensive testing across dozens of training runs, here are the dataset configurations that consistently produce high-quality results for different LoRA types.

Face-Only LoRA (Headshots/Portraits Only): If your goal is generating headshots and half-length portraits, you don't need full-body images. Focus entirely on facial consistency.

Optimal Configuration: 100+ face-focused images

• 70-80 close-up headshots (shoulders and above)
• 20-30 half-length portraits (waist and above)
• Variety of expressions, angles, and lighting
• Consistent subject across all images

Real-World Results: In testing, 100+ face images produced excellent facial consistency with strong identity preservation across different prompts, styles, and contexts. The LoRA reliably generates recognizable character faces in varied scenes. For visual novel character creation requiring extreme consistency, also see our VNCCS guide.

Smaller datasets (20-30 face images) worked but showed weaker consistency and occasional facial feature drift.

Full-Body LoRA (Complete Character): For generating full-body images with consistent character appearance from head to toe, you need body proportion training data.

Optimal Configuration: 100+ total images split 50/50

• 50+ headshots and close-up portraits
• 50+ full-body shots (head-to-toe visible)
• Mix of poses, clothing, and contexts
• Consistent character across all images

Why The 50/50 Split Works: This balanced approach ensures the LoRA learns facial details from close-ups while understanding body proportions from full-body shots. Skewing too heavily toward either type creates weaknesses.

Too many headshots and the LoRA struggles with body generation. Too many full-body shots and facial consistency suffers.

Multi-Purpose LoRA (SFW + NSFW): For LoRAs handling both safe-for-work and adult content with consistent character representation, dataset separation and volume matter significantly.

Optimal Configuration: 200+ total images split by content type

• 100+ SFW images (50+ headshots, 50+ body shots)
• 100+ NSFW images (50+ headshots, 50+ body shots)
• Maintain angle and variety balance within each category
• Same character across all images

Why NSFW Training Needs More Images: The model has less pre-existing knowledge about NSFW compositions, requiring more training data to learn these patterns while maintaining character consistency.

The 100/100 split ensures adequate representation of both content types without the LoRA overfitting to either category.

Testing Results:

The Minimum Viable Dataset: While 100+ images is optimal, you can train usable LoRAs with 20-30 high-quality, diverse images. Expect weaker consistency and less versatility, but the LoRA will capture basic character features.

This minimal approach works for personal projects and experimentation but isn't recommended for professional or commercial work.

Dataset Preparation - Building Your Training Set

Quality dataset preparation matters as much as quantity. Here's how to build training sets that produce excellent LoRAs.

Image Selection Criteria:

Aspect Ratio Distribution: Modern LoRA training handles multiple aspect ratios. Vary your training data to match how you'll use the LoRA.

Recommended Distribution:

• 40% square (1:1) - headshots, close-ups
• 30% portrait (3:4 or 2:3) - full-body standing
• 20% landscape (4:3 or 3:2) - full-body action
• 10% ultra-wide or ultra-tall - creative compositions

Image Quality Requirements:

What to Avoid in Training Data: Don't include heavily filtered or edited images - Instagram filters confuse training. Avoid images with multiple people unless you can crop to single subject. Skip images where the subject is partially obscured or cut off. Exclude low-resolution or heavily compressed images.

Captioning Your Dataset:

Directory Structure: Organize your dataset logically for training tools. Create a training_dataset folder with subfolders for headshots, body_shots, sfw, and nsfw categories. Each image file should have a corresponding .txt caption file with the same name.

Most training tools expect images and corresponding .txt caption files in the same directory.

Training Parameters That Actually Matter

Beyond dataset composition, training parameters significantly affect LoRA quality. Here are tested configurations that consistently produce good results.

Core Training Parameters:

Learning Rate Impact: Learning rate controls how aggressively the LoRA learns from training data. Too high causes overfitting and instability. Too low means insufficient learning even with many steps.

Start with conservative learning rates (1e-4) and decrease for larger datasets to prevent overfitting.

Step Count Determination: Calculate total steps as: (number_of_images × epochs) / batch_size

For 100 images with 30 epochs and batch size 2: (100 × 30) / 2 = 1500 steps

Most training tools calculate this automatically based on your epoch setting.

Network Rank Explained:

Higher rank allows the LoRA to learn more complex features but requires more training data to avoid overfitting.

Training Platform Comparison:

Monitoring Training Progress: Watch for overfitting signs - training loss approaching zero while validation loss increases indicates overfitting. Sample generation every few hundred steps to visualize learning progress.

Stop training when sample quality plateaus - additional steps won't improve results.

Common Training Mistakes and How to Avoid Them

Even experienced creators make training mistakes that degrade LoRA quality. Here are the most common issues and their solutions.

Mistake 1 - Insufficient Dataset Variety:

Mistake 2 - Overfitting from Too Many Similar Images: Training on 100 nearly identical selfies produces a LoRA that only works for that specific pose and lighting. The model memorizes photos rather than learning character features.

Solution: Curate datasets for maximum diversity within consistent character representation.

Mistake 3 - Inconsistent Subject: Using multiple different people or characters in a single dataset confuses training. The LoRA tries to learn all subjects simultaneously, producing inconsistent results.

Solution: One LoRA = one subject. Train separate LoRAs for different characters.

Mistake 4 - Wrong Learning Rate:

Mistake 5 - Ignoring Training Metrics: Blindly running training without monitoring loss curves leads to suboptimal results. Training might overfit long before completion or might need more steps than initially planned.

Solution: Check sample outputs every 200-500 steps and watch loss curves.

Mistake 6 - Low-Quality Source Images:

Mistake 7 - Dataset Too Small for Complexity: Trying to train a multi-style, multi-outfit, multi-context LoRA with 20 images doesn't provide enough data for the model to learn all those variations.

Solution: Match dataset size to complexity goals. Simple character LoRA = 20-30 images. Complex versatile LoRA = 100+ images. For more common pitfalls to avoid, see our guide on 10 common ComfyUI beginner mistakes.

Advanced Training Strategies and Optimization

Beyond basic training, advanced techniques optimize LoRA quality and versatility.

Multi-Concept Training: Training a single LoRA on multiple related concepts (same character in different styles) requires careful dataset separation and increased image counts.

Approach: 50+ images per concept/style you want to capture. Use distinct caption keywords for each concept to help the LoRA differentiate.

Progressive Training: Start training with low learning rate and small network rank, then increase both gradually. This builds stable foundation before learning complex details.

Implementation:

• Phase 1: Rank 8, LR 5e-5, 500 steps
• Phase 2: Rank 16, LR 1e-4, 1000 steps
• Phase 3: Rank 32, LR 5e-5, 1500 steps

Dataset Augmentation:

Regularization Images: Include generic images of similar subjects (not your specific character) to prevent overfitting and maintain model capabilities.

Ratio: 1 regularization image per 2-3 training images. Example: 100 character images + 40 regularization images.

Tag Weighting: Use weighted caption tags to emphasize important features.

Example caption: (masterpiece:1.3), (character_name:1.5), blue eyes, blonde hair, red dress

The weights tell training to emphasize those tagged features more strongly.

Checkpoint Selection:

Multi-Resolution Training: Train on varied resolutions to improve LoRA flexibility. Include images at 512x512, 768x768, 1024x1024, and non-square ratios.

This produces LoRAs that work well across different generation resolutions.

Testing and Iterating Your LoRA

Training doesn't end when the process completes. Systematic testing reveals strengths, weaknesses, and iteration opportunities.

Initial Testing Protocol:

Quality Assessment Criteria:

Identifying Overfitting: Test with prompts significantly different from training data. If the LoRA struggles to generate anything outside training contexts, overfitting occurred.

Example: If all training images showed indoor scenes and the LoRA fails generating outdoor scenes, the model overfit to indoor contexts.

Iteration Strategy:

Version Management: Save training checkpoints at different step counts. This provides multiple LoRA versions to test and choose from.

Many creators find their best LoRA is from 70-80% through training rather than the final checkpoint.

Community Feedback: Share test generations in LoRA training communities for feedback. Experienced trainers quickly identify issues and suggest improvements.

Real-World Training Examples and Results

Here are specific training runs with exact configurations and results to demonstrate these principles in practice.

Example 1 - Portrait LoRA:

• Dataset: 120 face-focused images (90 headshots, 30 half-length)
• Parameters: Rank 32, LR 1e-4, 3000 steps, SDXL base
• Results: Excellent facial consistency across varied prompts and styles. LoRA weight 0.7-0.9 produced best results. Struggled with full-body generation as expected.
• Best Use: Headshot generation, avatar creation, portrait art. For face swapping workflows, see our ComfyUI face swap guide

Example 2 - Full Character LoRA:

• Dataset: 100 images (50 headshots, 50 full-body)
• Parameters: Rank 32, LR 5e-5, 2500 steps, SD 1.5 base
• Results: Good balance of facial and body consistency. Versatile across scenes and contexts. Slight facial drift at very high resolutions.
• Best Use: General character generation, varied scenes

Example 3 - Multi-Purpose LoRA (SFW/NSFW):

• Dataset: 220 images (110 SFW split 55/55, 110 NSFW split 55/55)
• Parameters: Rank 64, LR 1e-5, 5000 steps, SDXL base
• Results: Excellent consistency across both content types. Character recognizable in all contexts. Slightly longer training time justified by versatility.
• Best Use: Commercial character work, comprehensive character representation

Example 4 - Minimal Dataset:

• Dataset: 25 images (15 headshots, 10 body shots)
• Parameters: Rank 16, LR 1e-4, 1500 steps, SD 1.5 base
• Results: Recognizable character but inconsistent details. Worked well at specific LoRA weights (0.8-0.9) but weak outside that range. Prone to generating training outfit.
• Best Use: Personal projects, quick character concepts

Training Cost Comparison:

Lessons from Testing: The jump from 25 to 100 images dramatically improves consistency and versatility. Beyond 100 images, improvements become incremental rather than transformative.

The 50/50 split for full-body LoRAs consistently outperforms other ratios. Training on SDXL produces higher quality but requires more VRAM and time compared to SD 1.5.

Conclusion - Building Training Datasets That Work

LoRA training dataset composition makes the difference between mediocre results and excellent character consistency. The tested formulas in this guide provide starting points for your specific needs.

Key Takeaways: Face-only LoRAs work excellently with 100+ face-focused images. Full-body LoRAs need 100+ images split 50/50 between headshots and body shots. Multi-purpose LoRAs handling SFW and NSFW content benefit from 200+ images split 100/100. Quality and variety matter more than raw image count.

Your Training Strategy: Start with clear goals - what will this LoRA generate? Match dataset size and composition to those goals. Curate for quality and diversity over quantity. Test systematically and iterate based on actual results.

Platform Considerations: Local training provides complete control but requires technical setup and GPU resources. Cloud platforms like Apatero.com streamline the process with optimized training pipelines. CivitAI training offers beginner-friendly interfaces with guided workflows. For deploying your trained LoRAs in production workflows, see our workflow to production API guide.

What's Next: Build your first training dataset following these guidelines. Start with a modest 50-image dataset to learn the process, then scale up based on results. Join LoRA training communities to share results and learn from experienced trainers.

The Bottom Line: Great LoRAs come from thoughtful dataset preparation, appropriate training parameters, and systematic iteration. Follow these tested strategies, and you'll create consistent, versatile LoRAs that bring your characters to life across any context.

Your training data defines your LoRA's capabilities. Invest time in dataset preparation, and the results will reflect that quality.