That’s definitely a valid concern, and I believe we need to rethink how LLMs and other generative models are deployed and used in practice. We do explicitly discuss some of these considerations.

As a conlanger myself, I was mainly curious to explore whether LLMs could be used as a creative assistant for humans, as well as procedural generation in games with unbounded worlds. I hope this gets more people interested in conlanging and experimenting themselves.

Thanks for the heads-up, fixing this.

Try out some of the newest languages on our project page:
conlangcrafter.github.io

ConlangCrafter could potentially be used in pedagogy, typological and NLP work, and many entertainment applications. Imagine a video game where aliens can speak countless new procedurally-generated languages.

To enhance consistency and diversity, our pipeline incorporates randomness injection and self-refinement mechanisms. This is measured by our novel evaluation framework, providing rigorous evaluation for the new task of computational conlanging.

The ConlangCrafter pipeline harnesses an LLM to generate a description of a constructed language and self refines it in the process. We decompose language creation into phonology, grammar, and lexicon, and then translate sentences while constructing new needed grammar points.

Conlangs (Constructed Languages), from Tolkien’s Elvish to Esperanto, have long been created for artistic, philosophical, or practical purposes.
As generative AI proves its creative power, we ask:
Can it also take on the laborious art of conlang creation?

Check out our project page and paper for more info:
Project page: wildcat3d.github.io
Paper: arxiv.org/abs/2506.13030
(5/5)

At inference time, we inject the appearance of the observed view to get consistent novel views. This also enables cool applications like appearance-conditioned NVS! (4/5)

To learn from this data, we use a novel multi-view diffusion architecture adapted from CAT3D, modeling appearance variations with a bottleneck encoder applied to VAE latents and disambiguating scene scale via warping. (3/5)

Photos like the ones below differ in global appearance (day vs. night, lighting), aspect ratio, and even weather. But they give clues to how scenes are build in 3D. (2/5)

See our paper, project page, and GitHub for more details and a full implementation!
ArXiv: arxiv.org/abs/2502.00129
Project page: tau-vailab.github.io/ProtoSnap/
GitHub: github.com/TAU-VAILab/P...

Finally we show that ProtoSnap-aligned skeletons can be used as conditions for a ControlNet model to generate synthetic OCR training data. By controlling the shapes of signs in training, we can achieve SOTA on cuneiform sign recognition. (Bottom: synthetic generated sign images)

Our results show that ProtoSnap effectively aligns wedge-based skeletons to scans of real cuneiform signs, with global and local refinement steps. We provide a new expert-annotated test set to quantify these results.

ProtoSnap uses features from a fine-tuned diffusion model to optimize for the correct alignment between a skeleton matched with a prototype font image and a scanned sign. Perhaps surprising that image generation models can be applied to this sort of discriminative task!

We tackle this by directly measuring the internal configuration of characters. Our approach ProtoSnap "snaps" a prototype (font)-based skeleton onto a scanned cuneiform sign using a multi-stage pipeline with SOTA methods from computer vision and generative AI.

Some prior work has tried to classify scans of signs categorically, but signs' shapes differ drastically in different time periods and regions making this less effective. E.g. both signs below are AN, from different eras. (Top: font prototype; bottom: scan of sign real tablet)