Accuracy plateaus around 59% — could more granular style labels or a larger dataset help?
Thanks for sharing this — interior style classification is a genuinely underexplored task on the Hub compared to text-to-image generation.
Looking at the training curve, accuracy plateaus around 58-59% after epoch 5-6, and the training set looks fairly small (only ~13 steps/epoch at batch size 128). A couple of things that might help push this further:
- More granular/fewer overlapping style labels — a lot of interior style datasets suffer from label ambiguity (e.g. "modern" vs "contemporary" vs "minimalist" are visually very close), which caps accuracy regardless of model size.
- Data augmentation or a larger pretrained backbone (e.g. ViT-Large or a CLIP-based classifier) might squeeze out more headroom without needing much more labeled data.
On a related note — we've been tackling the adjacent problem of going from a classified room style straight to a rendered, style-consistent visualization (rather than just labeling), which might be a useful reference if you're thinking about downstream applications for this classifier: AI Interior Design Generator
Curious what dataset this was trained on — is it public? Would be interesting to see if the label taxonomy is the bottleneck here.