SelfLift | Self-Recovering Resolution Transition

01 / Abstract

Abstract

Bottleneck

Recent few-step diffusion models have shifted the inference bottleneck from sampling steps to spatial token computation. Resolution transition offers an effective way to exploit the coarse-to-fine generation process, but existing transition strategies can introduce artifacts and semantic drift that are difficult to recover under few-step sampling.

Self-recovering transition

We propose SelfLift, a self-recovering resolution transition framework for efficient few-step diffusion models. SelfLift-zero introduces a training-free artifact-aware consistency lift to correct transition errors without additional denoising steps or external super-resolution models. Building upon this stable transition, SelfLift-rich further introduces an on-policy self-recovering training strategy for the same low-to-high resolution transition trajectory used at inference, leveraging high-resolution generation priors to recover richer semantics and fine-grained details while avoiding off-policy mismatch.

Quality preserved

SelfLift is lightweight and orthogonal to existing acceleration techniques. On FLUX.2-Klein-9b and Z-Image-Turbo, SelfLift-rich reduces native few-step latency by 41.5% and 44.1%, reaching 29.61× and 19.21× overall speedups while improving preference, aesthetics, CLIP alignment, and GenEval over SelfLift-zero.

Native speedNo external SR model Richer recoveryStudent-trajectory supervision

Few-Step Diffusion Models Efficient Inference Resolution Transition Artifact-Aware Consistency Lift On-Policy Self-Recovering

02 / Effect Comparison

Effect Comparison

SelfLift targets aggressive progressive-resolution inference, where direct transitions often trigger ghosting, text drift, spatial drift, and oversmoothing.

SelfLift qualitative comparison against baseline methods — Vanilla, RaLu, and Speed exhibit visible transition artifacts under aggressive acceleration. SelfLift-zero stabilizes model-native transitions, and SelfLift-rich recovers richer text and high-frequency details.

Stage-wise runtime breakdown for SelfLift — Low-resolution rollout shrinks early spatial-token computation, while SelfLift keeps the remaining high-resolution path short and model-native.

Quality versus latency comparison plot — SelfLift-zero forms a training-free native transition, while SelfLift-rich moves the frontier further with trajectory-aligned high-resolution recovery.

03 / Analysis

Analysis

The paper figures isolate the core transition issue: speed depends on moving resolution later, but reliability depends on recovering a target-resolution latent state.

Paper Figure 1 overview of SelfLift progressive-resolution inference — Reliable transition unlocks later low-to-high switches: prior transitions trade fidelity for speed, while SelfLift preserves native high-resolution quality.

Paper Figure 2 analysis of latent lifting and consistency — Direct latent lifting is sharp but inconsistent; pixel-VAE anchoring is stable but smooth. SelfLift corrects only high-risk regions toward the stable anchor.

Appendix Figure 1 target-VAE round-trip consistency analysis — Round-trip consistency shows the same pattern: directly lifted latents deviate most from target-VAE reachable states.

04 / Method Diagram

Method Overview

The overview figure summarizes SelfLift-zero and SelfLift-rich: a training-free Artifact-Aware Consistency Lift, followed by optional On-Policy Self-Recovering training.

05 / Experimental Table

Experimental Table

Quantitative results are shown in the original paper-style table format, without reformatting into a separate web table.

Main quantitative results table from the SelfLift paper — Full paper-style table from the latest source values.