Research map / Image generation / Fast diffusion sampling

Continuous-Multiple Image Outpainting in One-Step via Positional Query and A Diffusion-based Approach

ICLR 2024 · Image generation › Fast diffusion sampling · approach: Distillation for few-step generation · 3 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 PQDiff, a diffusion-based image outpainting method that uses relative positional embeddings between randomly cropped views as queries, enabling continuous-multiple, one-step outpainting without a pretrained backbone.

Problem

Image outpainting currently requires multiple passes for large expansion multiples, is restricted to discrete multiples, and relies on GAN/MAE pipelines with pretrained encoders.

Approach

Train a diffusion model to reconstruct a target crop from a clean anchor crop plus relative positional embeddings (sin-cos of the relative offset/scale between the two random crops), with position-aware cross-attention where positional embeddings query the encoded anchor tokens; at inference, specifying a target position yields outpainting at any continuous multiple in one step.

Key results

• PQDiff + Copy achieves FID 20.100 at 2.25x on Scenery vs 20.366 for QueryOTR+Copy, and 39.465 at 11.7x vs 60.977.
• At 11.7x on Scenery, PQDiff reports FID 21.512 (abstract) and sampling time at 10.2% of QueryOTR's time.
• PQDiff with 20 timesteps takes 9.638 s to generate 64 images at all multiples versus 23.728 s (2.25x) and 71.184 s (11.7x) for QueryOTR.

Measured effects reported in the paper

Surprising findings

• Without the copy operation, PQDiff still reproduces input pixels in the center region, sometimes exactly (infinite PSNR), and on WikiArt PQDiff without copy beats PQDiff + Copy at 5x and 11.7x.
• Predicting x0 instead of noise performs much worse across all multiples.

Limitations

• Relies on the frozen VQVAE and fixed resize to 192x192 for inputs, limiting resolution and scaling evaluation.
• Discrete-trained variant generalizes poorly to outer-distribution multiples (e.g., 4x, 9x, 16x); scaling to other domains is untested.
• Abstract and text report inconsistent FID values (21.512 vs 20.100 for Scenery, 36.212 vs 36.216 for WikiArt), and PSNR results exclude infinite values.

Experimental setup

Datasets: Scenery Building Facades WikiArt ImageNet

Benchmarks: Scenery Building Facades WikiArt

Metrics: FID Inception Score center PSNR sampling time

Baselines: NSIPO SRN IOH Uformer QueryOTR Vanilla diffusion

Builds on: DDPM DDIM QueryOTR MAE VQVAE

Closest papers by meaning

• Unboxed: Geometrically and Temporally Consistent Video Outpainting (CVPR 2025 · similarity 0.88)
• Infinite-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation (AAAI 2025 · similarity 0.87)
• Autoregressive Omni-Aware Outpainting for Open-Vocabulary 360-Degree Image Generation (AAAI 2024 · similarity 0.86)
• Novel View Synthesis with Pixel-Space Diffusion Models (CVPR 2025 · similarity 0.86)
• Image Inpainting via Tractable Steering of Diffusion Models (ICLR 2024 · similarity 0.86)
• Understanding the Generalization of Pretrained Diffusion Models on Out-of-Distribution Data (AAAI 2024 · similarity 0.86)