Research
Your Latent Mask is Wrong: Pixel-Equivalent Latent Compositing for Diffusion Models
Rowan Bradbury, Elea Zhong
(In Review) · (Dec 2025)
Funded by Modal
Latent mixing by mask interpolation is fundamentally wrong because modern VAEs are nonlinear and globally entangled - so blending latents doesn’t blend pixels. We introduce Pixel-Equivalent Latent Compositing (PELC) and DecFormer, a 7.7M-parameter compositor that predicts per-channel blend weights and a residual to enforce pixel-consistent latent fusion. It eliminates seams, halos, and color drift, improves edge fidelity by up to ~50% with only ~3.5% compute overhead, and slots into diffusion pipelines as a drop-in replacement for heuristic masking, approaching dedicated inpainting models when paired with a small LoRA.
Your Latent Mask is Wrong: Pixel-Equivalent Latent Compositing for Diffusion Models
Rowan Bradbury, Elea Zhong
(In Review) · (Dec 2025)
Funded by Modal
Latent mixing by mask interpolation is fundamentally wrong because modern VAEs are nonlinear and globally entangled - so blending latents doesn’t blend pixels. We introduce Pixel-Equivalent Latent Compositing (PELC) and DecFormer, a 7.7M-parameter compositor that predicts per-channel blend weights and a residual to enforce pixel-consistent latent fusion. It eliminates seams, halos, and color drift, improves edge fidelity by up to ~50% with only ~3.5% compute overhead, and slots into diffusion pipelines as a drop-in replacement for heuristic masking, approaching dedicated inpainting models when paired with a small LoRA.
Deterministic Continuous Replacement: Fast and Stable Module Replacement in Pretrained Transformers
Rowan Bradbury, Aniket Srinivasan Ashok, Sai Ram Kasanagottu, Gunmay Jhingran, Shuai Meng
Accepted to NeurIPS 2025 Workshop ScaleOpt · (Dec 2025)
A deterministic module-replacement method for pretrained transformers that smoothly transitions from teacher to student without stochastic gates. By removing gate-induced gradient variance and adding near-free feature alignment, DCR converges faster and more stably than Theseus-style or distillation baselines in controlled attention-replacement experiments, establishing a clean foundation for swapping in efficient attention operators.
Deterministic Continuous Replacement: Fast and Stable Module Replacement in Pretrained Transformers
Rowan Bradbury, Aniket Srinivasan Ashok, Sai Ram Kasanagottu, Gunmay Jhingran, Shuai Meng
Accepted to NeurIPS 2025 Workshop ScaleOpt · (Dec 2025)
A deterministic module-replacement method for pretrained transformers that smoothly transitions from teacher to student without stochastic gates. By removing gate-induced gradient variance and adding near-free feature alignment, DCR converges faster and more stably than Theseus-style or distillation baselines in controlled attention-replacement experiments, establishing a clean foundation for swapping in efficient attention operators.