TrOCR Users Probe Multilingual Decoder Swaps

The core instinct is right, but the easy answer is not plug-and-play - the decoder/tokenizer contract is doing a lot of work here.

• –TrOCR's decoder is autoregressive and cross-attentive, so the architecture can support sequence generation from image features.
• –mT5 is the closer candidate conceptually, but you would still need to rebuild the text side around its tokenizer and generation setup.
• –MuRIL is not a causal decoder, so it does not satisfy the swap-in-decoder requirement the way a seq2seq model would.
• –For Hindi OCR, the bigger bottleneck is usually script coverage and vocabulary, so a multilingual tokenizer plus fine-tuning often matters more than the exact pretrained decoder.