Your LoRA might be fine — it may just be the wrong base model, says SDXL-era trainer
Inevitable_Pen9043 · reddit · 2026-11-07
A Reddit user shared a commonly overlooked lesson: if your LoRA can't learn a style even after 100-300 epochs, the problem may not be the LoRA, hyperparameters, or dataset — it may be the base model.
Drawing on experience since the Stable Diffusion XL days, he found that a LoRA that seemed undertrained on the base model (or early fine-tunes) could perform perfectly on a different fine-tuned checkpoint. So when judging a LoRA, test it across base models rather than just adding epochs.
More from Multimodal
- Kling AI Heads to Adobe MAX 2026, Teases 'Surprise' Upcoming Product — Kyrannio · 2026-11-07
- OpenAI opens GPT-Image-2.6 GIF creation to everyone — OpenAIDevs · 2026-11-07
- Higgsfield launches Miracle: layer-based pixel-level AI video editing in real time — azed_ai · 2026-11-07
- One Midjourney sref code, three subjects: consistent blue ink-wash lion, antelope and Pegasus — sergeantsref · 2026-11-07
- GPT Image 2.6 sweeps image leaderboards: Flare ~2x faster, Sunburst takes #1 — ArtificialAnlys · 2026-11-07
- Psychedelic AI MV 'Chemical Romance' built with Suno and Dreamina — AIandDesign · 2026-11-07