Training a LoRA on a finetuned base model overfits fast — epoch 1 beats epoch 10
Junko_Enoshima_18 · reddit · 2026-10-26
A user trained 3 anime character LoRAs on Civitai. The first two used the default Illustrious base and worked at epoch 10. The third, trained on Wai-Illustrious, overfit badly: epoch 1 matched expectations while epoch 3+ simply reproduced the training images (only 48 images, including 3D-model screenshots for a rare character). Epoch 1 still struggles with fingers. Open questions: base-model choice, dataset size, or epoch count — and whether training on a finetuned base breaks compatibility with other Illustrious-family models.
More from Multimodal
- Dense instance segmentation via Muse Wonder + SAM 3.1 on Meta's Model API — nikhilaravi · 2026-10-26
- LichtFeld Studio v0.5.4.3 ships community-built SteadySplats, POPSpa and Voxelization — janusch_patas · 2026-10-26
- One prompt to a 2-3 min narrated short: hands-on test of Antigravity video generation — AI_Andrew · 2026-10-26
- Creator makes 10,564-frame animated short entirely with Gemini toolchain — AI_Andrew · 2026-10-26
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26