SDXL LoRA Learns 5 of 6 Features, the Tail Never Trains: Conv Layers or Captions?
Relevant_Meat_9418 · reddit · 2026-08-23
Training an SDXL LoRA of a Bf 109 G with kohya sd-scripts (dim 32/alpha 16, 2700 steps, RealVisXL V5.0 base), the author gets the canopy right in 20/20 stills while the vertical fin always reverts to the base model's generic lobe and the rudder hinge never appears.
Three levers already failed: LoRA strength sweep 0.4–0.6, dataset 80→97 with tail close-ups, and tail-crop augmentation to 24% per epoch.
Two hypotheses: A. Conv layers — default LierLa only adapts Linear and 1x1 convs, leaving ResNet 3x3 convs untouched (confirmed in source); if part silhouettes live in spatial filters, the tail was untrainable while attention-carried features trained fine. Fix: networkargs convdim=16 convalpha=8 (beware: convalpha silently defaults to 1.0 if omitted). B. Captions — current captions are 15 words with zero component names; a Civitai guide reports the same symptom on a MiG-29 (stabilizers becoming smoke trails) fixed by 90+ word captions naming every part, contradicting the usual tag-what-varies rule.
The author asks for LoCon vs LoRA A/B results on hard-surface part reversion, whether exhaustive part naming helps, and whether masked loss beats cropping.
More from Multimodal
- NeurIPS 2026 to host BabyVLM workshop on learning like babies — LChoshen · 2026-08-23
- LTX v2v Analysis: Fine Geometry Collapse and Flat Surface Hallucinations — Relevant_Meat_9418 · 2026-08-23
- Achieving Studio Quality AI Music with a 70/30 Udio and Suno Split — Ok_pettech · 2026-08-23
- Short film made with Minimax, heavy post-processing applied — Disastrous-Agency675 · 2026-08-23
- AI Demo: Visualizing Dreams and Afterlife in 8-bit Pixel Style — pickover · 2026-08-23
- Boson AI targets voice market with Higgs RealTime model supporting 100+ languages — smolix · 2026-08-23