Research map / Image generation / Autoregressive and latent generation

RaPD: Resolution-Agnostic Pixel Diffusion via Semantics-Enriched Implicit Representations

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis · 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 RaPD, which runs flow-matching diffusion directly in a continuous Neural Image Field latent and renders it at arbitrary resolutions with a coordinate-queried attention renderer, using semantic distillation from DINOv2 to make the latent generation-aware.

Problem

Most image generators operate on discrete grids, so generating at arbitrary unseen resolutions needs more tokens and compute; existing continuous neural-field representations are reconstruction-oriented and not suitable as generative latents.

Approach

Diffuse over a dense, continuous NIF latent made generation-aware via DINOv2 semantic distillation (margin cosine and distance-matrix losses) plus a resolution-dependent timestep shift, and render with a coordinate-conditioned Swin-style windowed attention renderer over output-coordinate tokens so diffusion cost is independent of output size.

Key results

• On GenEval at 512x512, RaPD reaches overall 0.85, above the re-run PixelGen (0.79) and DeCo (0.83) and matching the reported DeCo 0.86 only below it.
• On DPG-Bench, RaPD averages 81.6, above re-run PixNerd (81.5), DeCo (81.2) and PixelGen (80.4).
• At 4096x4096 RaPD takes 2.13 s per image on a B200 versus 61.46 s for PixelGen and 62.72 s for DeCo, with the denoising token count fixed at 1,024.

Surprising findings

• Reported DeCo GenEval overall (0.86) from its paper is higher than RaPD (0.85), though the authors' re-run of DeCo gives 0.83.

Limitations

• Trained on about 36M BLIP3o images, far less than commercial-scale models, limiting absolute quality.
• Frozen Qwen3-1.7B text encoder may limit complex prompt understanding.
• Inference stack lacks FlashAttention, fused window-attention kernels, FP8 and compiler fusion, so latency figures are conservative.

Experimental setup

Datasets: BLIP3o DPG-Bench prompts GenEval prompts

Benchmarks: GenEval DPG-Bench rFID

Metrics: GenEval overall score DPG-Bench average rFID inference time

Baselines: LDM SD3 FLUX.1-dev PixArt-α DALL-E 3 PixelFlow PixNerd DeCo PixelGen SD v2

Builds on: LIIF CLIF PixelGen REPA flow matching

Closest papers by meaning

• DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.89)
• DemoFusion: Democratising High-Resolution Image Generation With No $$$ (CVPR 2024 · similarity 0.88)
• PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.88)
• PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · similarity 0.88)
• FAM Diffusion: Frequency and Attention Modulation for High-Resolution Image Generation with Stable Diffusion (CVPR 2025 · similarity 0.88)
• ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.88)