Research map / Image generation / Fast diffusion sampling
FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing
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 FireFlow, a training-free second-order-accuracy ODE solver for rectified flow (FLUX) that reuses the previous midpoint velocity so each step costs one NFE, enabling fast inversion and prompt-based image editing in 8 steps.
Problem
Fast and accurate inversion of rectified-flow (ReFlow) models such as FLUX back to structured noise, which is needed for reconstruction and semantic image editing, while existing high-order solvers cost extra function evaluations.
Approach
Reuse the midpoint velocity computed at the previous step as the velocity estimate for the current step, so a midpoint-style update costs only one model evaluation per step; the authors argue this keeps second-order accuracy because ReFlow velocities are nearly constant.
Key results
- On PIE-Bench, FireFlow with 8 steps (18 NFE) reaches structure distance 0.0271 and PSNR 23.03 versus RF-Solver's 0.0311 and 22.90 at 60 NFE.
- Per-image runtime on RTX 3090 at 512x512 is 7.70s versus 23.76s for vanilla ReFlow inversion, a 3.09x speedup.
- On MSCOCO T2I with 11 NFE, FireFlow gets FID 25.16, better than FLUX-dev (26.77, 20 NFE) and RF-Solver (25.93, 20 NFE).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the reused-velocity second-order solver instead of vanilla FLUX-dev sampling | FLUX-dev · MSCOCO Caption 2014 val, 10K | FID: 26.77 -> 25.16 | Helped |
| use the reused-velocity second-order solver instead of RF-Solver | FLUX-dev · MSCOCO Caption 2014 val, 10K | CLIP Score: 31.35 -> 31.42 | No clear effect |
| use FireFlow rather than RF-Solver at 30 steps | FLUX-dev inversion and reconstruction · DCI, first 1K images (30 steps) | LPIPS: 0.2926 -> 0.1579 | Helped |
| use FireFlow rather than RF-Solver at 30 steps | FLUX-dev inversion and reconstruction · DCI, first 1K images (30 steps) | SSIM: 0.7078 -> 0.8160 | Helped |
| use FireFlow rather than RF-Solver at 8 steps | FLUX-dev inversion and reconstruction · DCI, first 1K images (8 steps) | LPIPS: 0.5010 -> 0.4111 | Helped |
| use FireFlow rather than RF-Solver at 15 steps | FLUX editing · PIE-Bench | Background PSNR: 22.90 -> 23.28 | Helped |
| use FireFlow rather than RF-Solver at 15 steps | FLUX editing · PIE-Bench | CLIP Similarity, Edited: 22.88 -> 22.94 | Helped |
| use FireFlow rather than RF-Solver | ReFlow inversion-based editing · 512×512 | Speedup: 0.94× -> 3.09× | Helped |
| increase editing steps from 2 to 8 (ablation) | FireFlow · Prompt-guided image editing | editing performance | Helped |
| increase editing steps from 8 to 10 or 12 (ablation) | FireFlow · Prompt-guided image editing | editing performance | No clear effect |
| add K to Add Q (ablation) | FireFlow editing strategy · PIE-Bench, 8 steps, guidance [1,2,2,...] | Structure Distance: 0.0590 -> 0.0537 | Helped |
| add K and V to Add Q (ablation) | FireFlow editing strategy · PIE-Bench, 8 steps, guidance [1,2,2,...] | Structure Distance: 0.0590 -> 0.0416 | Helped |
Surprising findings
- Reusing the previous step's midpoint velocity keeps second-order-like accuracy at first-order (one NFE per step) cost, despite the paper's own analysis suggesting the velocity-approximation error bound is only O(Δt).
- Replacing only the value features in self-attention (rather than also adding query or key features) gives the best structure preservation, while adding key features improves prompt fidelity at the cost of preservation.
Limitations
- Struggles with editing that changes object colors or uncommon scenes, and with unusual prompts such as 'stormtrooper with blue hair'.
- Adding key features to self-attention to fix these failures reduces preservation of original structure and background.
- The theoretical error analysis relies on smoothness and bounded-derivative assumptions and the O(Δt) velocity-error claim is empirical on FLUX-dev.
Experimental setup
Datasets: MSCOCO Caption 2014 validation (10K subset) Densely Captioned Images (DCI, first 1K) PIE-Bench (700 images)
Benchmarks: MSCOCO Caption 2014 DCI PIE-Bench
Metrics: FID CLIP score LPIPS SSIM PSNR structure distance NFE runtime
Baselines: Prompt-to-Prompt MasaCtrl Pix2Pix-zero Plug-and-Play DiffEdit DirectInversion RF-Inversion RF-Solver Add-it PnP-Inv DDIM inversion
Builds on: Rectified Flow FLUX RF-Solver RF-Inversion Plug-and-Play
Closest papers by meaning
- Taming Rectified Flow for Inversion and Editing (ICML 2025 · similarity 0.90)
- Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations (ICLR 2025 · similarity 0.90)
- SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image Editing (NeurIPS 2025 · similarity 0.88)
- Unveil Inversion and Invariance in Flow Transformer for Versatile Image Editing (CVPR 2025 · similarity 0.87)
- Root-Selecting Fixed-Point Inversion for Rectified Flows via Trajectory Straightness (NeurIPS 2026 · similarity 0.86)
- Lightning-Fast Image Inversion and Editing for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.86)
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.