Research map / Image generation / Editing and style transfer

PixPerfect: Seamless Latent Diffusion Local Editing with Discriminative Pixel-Space Refinement

NeurIPS 2025 · Image generation › Editing and style transfer · approach: Inversion-based diffusion editing · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PixPerfect, a GAN-based pixel-space refiner that corrects seams, color shifts and texture mismatches left by latent diffusion inpainting and local editing models, using a learned discriminative tone-mapped loss and synthetic artifact simulation.

Problem

Latent diffusion inpainting and local editing produce pixel-level inconsistencies (chromatic shifts, texture mismatches, seams) at mask boundaries that existing latent-decoding and harmonization fixes do not fully eliminate or generalize across latent spaces and tasks.

Approach

A pixel-space GAN refiner trained with a per-sample polynomial tone mapping that amplifies color/texture discrepancies (discriminative pixel space), plus a synthetic artifact simulation pipeline and inference-time jitter pooling.

Key results

• On MISATO, FLUX-Fill-PixPerfect reduces FID from 14.66 to 10.87 and LPIPS from 0.195 to 0.141 versus FLUX-Fill.
• On RORDS object removal, OmniPaint+PixPerfect improves FID from 23.05 to 18.87 and PSNR from 24.67 to 27.96.
• Refiner adds about 2.7 s to a 9.7 s FLUX-Fill sampling run at 512x512 on a single A100, roughly 21.8% of total inference time.

Measured effects reported in the paper

Surprising findings

• Poisson blending, a classical baseline, was evaluated only with ground-truth-derived gradients, which the authors note is a leakage issue and not a deployable baseline.
• Replacing only the masked region in the FLUX VAE latent space causes global background drift in the decoded image, indicating spatial entanglement in the latent space.

Limitations

• Cannot correct major semantic errors from the generative model.
• Depends on reasonably accurate initial predictions and predefined edited regions.
• Training requires a very large curated dataset (about 300M images) and 32 A100 GPUs for a week.

Experimental setup

Datasets: Places2 MISATO RORDS CoModGAN masks

Benchmarks: MISATO Places2 RORDS object insertion triplet dataset (300 triplets)

Metrics: FID LPIPS L1 PSNR U-IDS P-IDS MUSIQ MANIQA

Baselines: Asymmetric VQ-GAN DiffHarmony++ ASUKA Poisson blending BrushNet CLIPAway PowerPaint OmniPaint Pbe ObjectStitch AnyDoor

Builds on: CMGAN LPIPS CoModGAN

Closest papers by meaning

• DesignEdit: Unify Spatial-Aware Image Editing via Training-free Inpainting with a Multi-Layered Latent Diffusion Framework (AAAI 2025 · similarity 0.89)
• PIXELS: Progressive Image Xemplar-based Editing with Latent Surgery (AAAI 2025 · similarity 0.89)
• PixelMan: Consistent Object Editing with Diffusion Models via Pixel Manipulation and Generation (AAAI 2025 · similarity 0.89)
• CTRL-D: Controllable Dynamic 3D Scene Editing with Personalized 2D Diffusion (CVPR 2025 · similarity 0.88)
• PS-Diffusion: Photorealistic Subject-Driven Image Editing with Disentangled Control and Attention (CVPR 2025 · similarity 0.87)
• PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.87)