Provide more accessible/free GPU compute credits

We're working on a project training normalizing flow models for image generation, and it's amazing how much GPU compute these things demand. We're doing this at university scale, and even getting something that looks decent (not even SOTA) takes a lot of computeresources. It's interesting because, compared to LLMs, you can't just use open-source models that are close to SOTA, and it's even impossible to train on a specific task and get reasonable results. (I know that for general-purpose LLMs you need tons of compute...) I'm wondering why? Besides the good open-source models, images probably contain more information, so they're harder to generate. But it also makes me think we're doing something wrong here. One idea (maybe counterintuitive): image generation models are actually smaller compared to LLMs. So why do these "smaller" image models still need so much compute to train well? Maybe it's the data (images are high-dimensional and contain a lot of "noise" that's harder to learn), maybe it's the training dynamics (diffusion/flow matching is expensive per step), or perhaps we just haven't figured out the right architectures yet. Could it be that for generation tasks, larger models are actually easier to train, and you need less effort to fit the data? There's theoretical and empirical work suggesting larger models are easier to optimize and more sample-efficient (the scaling laws work, overparameterized networks finding global minima, infinite overparameterization results...). But I haven't seen anything that formalizes this well for image generation specifically. Images lie in a low-dimensional manifold that we need to learn, but you have tons of pixels that don't really matter. In some sense, they're noise (or highly correlated with other pixels). How does model size interact with learning this structure efficiently? I don't have a good answer. Anyway, till then, if someone has available GPUs, let me know 😁