Research map / Image generation / Autoregressive and latent generation
Your Latent Mask is Wrong: Pixel-Equivalent Latent Compositing for Diffusion Models
Rowan Bradbury, Dazhi Zhong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace heuristic latent blending with DecFormer (soft masks, 1024px) | FLUX VAE compositing · Halo L1 | Halo L1: 0.050 -> 0.018 | Improved |
| replace heuristic blending with DecFormer (binary masks) | FLUX VAE compositing · PSNR | dB: 28.4 -> 35.7 | Improved |
| replace heuristic blending with DecFormer inside FLUX.1-Dev inpainting | FLUX.1-Dev · COCO-2017 val inpainting | FID: 23.514 -> 20.556 | Improved |
| add inpainting LoRA on top of DecFormer | FLUX.1-Dev + DecFormer · COCO-2017 val inpainting | FID: 20.556 -> 19.280 | Improved |
| DecFormer+LoRA vs fully finetuned Flux Fill | FLUX.1-Dev · COCO-2017 val inpainting | PSNR: 14.231 vs 16.750 | Hurt |
| add halo-weighted L1 loss (ablation) | DecFormer · network ablation | Halo L1: 0.0973 -> 0.0829 | Improved |
| add residual shift head alongside alpha (ablation) | DecFormer · network ablation | LPIPS: 0.0514 -> 0.0303 | Improved |
| add halo loss (ablation) | DecFormer · network ablation | LPIPS/MSE: LPIPS 0.0299 vs 0.0303 | No clear effect |
Closest papers by meaning
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.87)
- PixPerfect: Seamless Latent Diffusion Local Editing with Discriminative Pixel-Space Refinement (NeurIPS 2025 · similarity 0.86)
- PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining (NeurIPS 2026 · similarity 0.86)
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.86)
- Towards Enhanced Image Inpainting: Mitigating Unwanted Object Insertion and Preserving Color Consistency (CVPR 2025 · similarity 0.85)
- Color Alignment in Diffusion (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
