I would argue the opposite — image generation is the clear loser. If you've ever tried to do it yourself, grabbing a bunch of LoRAs from Civitai to try to convince a model to draw something it doesn't initially know how to draw — it becomes clear that there's far too much unavoidable correlation between "form" and "representation" / "style" going on in even a SOTA diffusion model's hidden layers.

Unlike LLMs, that really seem to translate the text into "concepts" at a certain embedding layer, the (current, 2D) diffusion models will store (and thus require to be trained on) a completely different idea of a thing, if it's viewed from a slightly different angle, or is a different size. Diffusion models can interpolate but not extrapolate — they can't see a prompt that says "lion goat dragon monster" and come up with the ancient-greek Chimera, unless they've actually been trained on a Chimera. You can tell them "asian man, blond hair" — and if their training dataset contains asian men and men with blonde hair but never at the same time, then they won't be able to "hallucinate" a blond asian man for you, because that won't be an established point in the model's latent space.

---

On a tangent: IMHO the true breakthrough would be a model for "text to textured-3D-mesh" — where it builds the model out of parts that it shapes individually and assembles in 3D space not out of tris, but by writing/manipulating tokens representing shader code (i.e. it creates "procedural art"); and then it consistency-checks itself at each step not just against a textual embedding, but also against an arbitrary (i.e. controlled for each layer at runtime by data) set of 2D projections that can be decoded out to textual embeddings.

(I imagine that such a model would need some internal "blackboard" of representational memory that it can set up arbitrarily-complex "lenses" for between each layer — i.e. a camera with an arbitrary projection matrix, through which is read/written a memory matrix. This would allow the model to arbitrarily re-project its internal working visual "conception" of the model between each step, in a way controllable by the output of each step. Just like a human would rotate and zoom a 3D model while working on it[1]. But (presumably) with all the edits needing a particular perspective done in parallel on the first layer where that perspective is locked in.)

Until we have something like that, though, all we're really getting from current {text,image}-to-{image,video} models is the parallel layered inpainting of a decently, but not remarkably exhaustive pre-styled patch library, with each patch of each layer being applied with an arbitrary Photoshop-like "layer effect" (convolution kernel.) Which is the big reason that artists get mad at AI for "stealing their work" — but also why the results just aren't very flexible. Don't have a patch of a person's ear with a big earlobe seen in profile? No big-earlobe ear in profile for you. It either becomes a small-earlobe ear or the whole image becomes not-in-profile. (Which is an improvement from earlier models, where just the ear became not-in-profile.)

[1] Or just like our minds are known to rotate and zoom objects in our "spatial memory" to snap them into our mental visual schemas!

Stable Fast 3D: Rapid 3D Asset Generation from Single Images