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.

Original post →

More from Multimodal

Multimodal channel →