prof. adithya vadapalli's lab at iitk was looking at discrete diffusion for text, mostly sedd and llada style setups. the pitch is simple: instead of generating left to right like every autoregressive model, you denoise the whole sequence at once. sounds great until you point it at hindi.
first problem is the tokenizer. every checkpoint we pulled was trained on an english-heavy corpus, so devanagari gets chopped into garbage subword pieces. the model spends half its capacity just learning to spell before it can learn anything about grammar.
second problem is noise scheduling. the diffusion process corrupts tokens randomly and learns to reverse it, but hindi has way more morphological variation per word than english. the same schedule that works fine on english destroys hindi sentences faster, so early denoising steps give you almost nothing useful.
we ended up retraining a hindi-specific tokenizer and tuning the noise schedule by hand, checkpoint by checkpoint, watching loss curves that didn't behave like the paper said they would. no shortcut. run it, read the output, adjust.
still not done. hindi nlp is underfunded compared to english, and diffusion text models are new enough that half the tooling assumes things that don't hold outside english. that gap is basically the whole research problem.