Are you about to train your first (or next) LoRA on TensorArt and you're unsure whether to use Pony Diffusion v6 or Illustrious XL 2.0 BASE?
This article is based on real experience and can help you avoid wasting credits and understand the practical differences between both models.
⚠️ Personal experience
I trained a LoRA using Pony Diffusion v6 with:
30 images
20 repeats per image
5 epochs
Result: the style was clearly visible by the 3rd epoch. The LoRA worked well and was usable.
Then I tried the same setup with Illustrious XL 2.0 BASE:
30 images
20 repeats per image
10 epochs
Result: after 5 epochs, there was no visible change. Even after 10 epochs, only minimal improvement. In short: the LoRA didn’t work as expected, and it felt like a waste of credits.
🔍 Side-by-side comparison: Pony vs Illustrious
Feature | Pony Diffusion v6 (SD 1.5) | Illustrious XL 2.0 BASE (SDXL)
--------------------------|---------------------------------|--------------------------------
Min. recommended images | 20–40 | 60–100
Ideal repeats per image | 10–20 | 5–10
Epochs to see results | 3–5 | 8–12
Early style visibility | Yes | No
Training speed | Fast | Slow
Quality with low data | Good | Poor
Credit efficiency | High | Low
Best suited for... | Anime, cartoon, simple styles | Realism, portraits, fine detail
✅ Practical recommendations
Use Pony Diffusion if you...
Are new to LoRA training
Have limited resources or time
Work with anime/cartoon styles
Want fast results with smalldatasets
Use Illustrious XL 2.0 BASE if you...
Have previous experience
Can prepare datasets with 60+ diverse images
Aim for realism, portraits, or high detail
Are ready for longer testing and refinement (10+ epochs)
💡 Tips for training LoRAs on TensorArt
Avoid using more than 15 repeats per image on SDXL if your dataset is under 50 images.
Check your results every 2–3 epochs using the same prompt and seed.
Use simple test prompts, like: <lora:your_lora_name:0.8>, portrait of <your_tag>
Don’t overload your prompt with terms like “masterpiece” or “anime style” when testing an early LoRA.
If you're experimenting with a small dataset, start with Pony. Once it works well, you can consider scaling it to SDXL.
You can boost your dataset with variations (poses, backgrounds, lighting) to improve generalization.
📌 Conclusion
Illustrious XL is not just a better version of Pony Diffusion — it’s a different beast.
If you try to apply the same training rules from SD 1.5 directly to SDXL, you'll likely run into frustration or wasted resources.
Pony Diffusion remains the most reliable option for fast, small-scale, anime-style LoRAs.
Illustrious XL shines in more complex use cases — but only if you're well-prepared.
Not every failed experiment is a mistake... if you learn something from it.
I hope this article helps you train smarter!
well in the end all this is good from my experience and anyone can have a different point of view...
I don't have the absolute truth, I'm just someone on the Internet. Don't pay attention to my posts, I do them more for experiments than anything else :p
