Research
Draw This First
Dazhi Zhong, Rowan Bradbury, Grant Davis
Workshop Submission · (Aug 2026)
Collaboration with Wand Technologies
Draw This First reframes sketch generation by separating what a drawing looks like from the order in which it is constructed. Rather than autoregressively predicting strokes one after another, the method adapts a pretrained latent flow-matching image model to predict an image-native intermediate representation in which drawing progress is encoded spatially as a continuous 2D order field. A modified VAE decoder then recovers this field together with foreground and stroke-instance segmentation, after which deterministic vectorization reconstructs ordered polylines. The model supports both text-to-sketch generation and image-to-vector derendering, while allowing drawing order itself to be specified in natural language. Central to the work is a new corpus of ~50k drawings commissioned from 50 artists, captured from Apple Pencil trajectories and averaging 77.5 strokes and 6,864 points per drawing. The dataset is hierarchically annotated by regions and subjects, and its recorded stroke sequences are programmatically permuted—with natural within-region progression preserved—to create matched language descriptions of alternative valid drawing orders.
Draw This First
Dazhi Zhong, Rowan Bradbury, Grant Davis
Workshop Submission · (Aug 2026)
Collaboration with Wand Technologies
Draw This First reframes sketch generation by separating what a drawing looks like from the order in which it is constructed. Rather than autoregressively predicting strokes one after another, the method adapts a pretrained latent flow-matching image model to predict an image-native intermediate representation in which drawing progress is encoded spatially as a continuous 2D order field. A modified VAE decoder then recovers this field together with foreground and stroke-instance segmentation, after which deterministic vectorization reconstructs ordered polylines. The model supports both text-to-sketch generation and image-to-vector derendering, while allowing drawing order itself to be specified in natural language. Central to the work is a new corpus of ~50k drawings commissioned from 50 artists, captured from Apple Pencil trajectories and averaging 77.5 strokes and 6,864 points per drawing. The dataset is hierarchically annotated by regions and subjects, and its recorded stroke sequences are programmatically permuted—with natural within-region progression preserved—to create matched language descriptions of alternative valid drawing orders.
Your Latent Mask is Wrong: Pixel-Equivalent Latent Compositing for Diffusion Models
Rowan Bradbury, Elea Zhong
Accepted to CVPR 2026 Main Track · (Dec 2025)
Funded by Modal
Latent inpainting in diffusion models still relies almost universally on linearly interpolating VAE latents under a downsampled mask. This 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
Accepted to CVPR 2026 Main Track · (Dec 2025)
Funded by Modal
Latent inpainting in diffusion models still relies almost universally on linearly interpolating VAE latents under a downsampled mask. This 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.