75-image ablation: how Reddit criticism made us rethink the FLUX-LoRA pipeline
A systematic study: 5 pipeline variants, training data leakage via Cyrillic text, and why the "sandwich" approach is a workaround, not a fix.
We trained a FLUX-LoRA for generating illustrations in the Soviet matchbox art style. The initial results looked good — until Reddit tore them apart. The criticism forced us to run an honest ablation study.
We tested 5 pipeline variants across multiple random seeds: pure LoRA, two-pass "sandwich," different LoRA scale values, with and without Cyrillic prompts. 75 images, blind comparison.
Key finding: the two-pass "sandwich" was masking problems, not solving them. Training data leakage showed up as Cyrillic text on images. LoRA scale was suboptimal — at the right value, single-pass generation matched the "sandwich" results.
The real fix is expanding the dataset and retraining the LoRA, not relying on pipeline tricks. The Reddit criticism was uncomfortable but led us to the right answer.
Takeaway for anyone training LoRAs: if your results depend on pipeline tricks, that's a signal the problem is in training, not inference. An ablation study is a mandatory step before claiming results.