On the Diffusibility of High-Dimensional Latents

A paper finds that x0-prediction (predicting clean data) beats velocity prediction for diffusion in the high-dimensional, low-effective-dimensionality latents of reconstruction-finetuned visual encoders, improving text-to-image generation.

A paper titled 'On the Diffusibility of High-Dimensional Latents,' tracked on Hugging Face daily papers on 2026-09-22 and listed on arXiv cs.AI, cs.LG, and cs.CL on 2026-09-23, studies Representation Autoencoders, which let diffusion models generate in the feature spaces of pretrained visual encoders. Many off-the-shelf encoders are not optimized for faithful reconstruction and discard fine-grained visual detail; finetuning them for reconstruction recovers that detail but reduces the representation's effective dimensionality and changes its geometry. In these spaces, standard flow-matching velocity prediction spends capacity fitting noise directions orthogonal to the low-dimensional signal manifold, which is inefficient. Predicting clean data with x0-prediction instead focuses learning on the signal manifold and consistently improves text-to-image performance across several strong-reconstruction encoders. The two reports are consistent and give no conflicting figures, versions, or dates.

• Paper title: 'On the Diffusibility of High-Dimensional Latents'
• Tracked on Hugging Face daily papers on 2026-09-22 and on arXiv cs.AI, cs.LG, and cs.CL on 2026-09-23
• Representation Autoencoders let diffusion models operate in the feature spaces of pretrained visual encoders
• Many off-the-shelf visual encoders discard fine-grained visual detail
• Reconstruction finetuning recovers fine detail but lowers the latents' effective dimensionality and changes their geometry (geometry change noted in the Hugging Face report)
• Standard flow-matching velocity prediction wastes capacity fitting noise directions orthogonal to the low-dimensional signal manifold
• x0-prediction (predicting clean data) focuses learning on the signal manifold
• x0-prediction consistently improves text-to-image generation across multiple strong-reconstruction encoders