Research map / Image generation / Editing and style transfer
SplitFlow: Flow Decomposition for Inversion-Free Text-to-Image 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 SplitFlow, an inversion-free text-guided image editing method for rectified flow models that decomposes the target prompt into LLM-generated sub-prompts, runs independent editing flows, and aggregates them via latent trajectory projection and cosine-similarity-weighted velocity aggregation.
Problem
Text-guided image editing with rectified flow models suffers from inaccurate inversion, gradient entanglement across multiple target attributes, and suboptimal fidelity/editability trade-offs in existing inversion-free ODE-based methods.
Approach
Decompose the target prompt into semantically separate sub-prompts with an LLM, compute an independent inversion-free (FlowEdit-style) flow per sub-prompt, then aggregate them at one timestep using projection onto the full-target latent (LTP) and softmax weights from pairwise cosine similarity of velocity fields (VFA), inspired by gradient conflict resolution in multi-task learning.
Key results
- On PIE-Bench with SD3.5, SplitFlow reports Structure Distance 11.68 and PSNR 27.12, versus FlowEdit (SD3.5) at 11.80 and 26.97.
- With SD3, SplitFlow achieves the best Whole CLIP similarity (26.96) and Edited CLIP similarity (23.83), above FlowEdit SD3 at 26.83 and 23.67.
- Ablation: naive averaging of sub-flows (AVG) raises PSNR from 22.13 to 23.36 over the FlowEdit baseline while keeping CLIP similarity near baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decompose the target flow and aggregate sub-target flows with LTP and VFA | FlowEdit with SD3 · PIE-Bench | Structure Distance ×10^3: 27.24 -> 25.96 | Helped |
| decompose the target flow and aggregate sub-target flows with LTP and VFA | FlowEdit with SD3 · PIE-Bench | PSNR: 22.13 -> 22.45 | Helped |
| decompose the target flow and aggregate sub-target flows with LTP and VFA | FlowEdit with SD3 · PIE-Bench | CLIP Similarity, Edited: 23.67 -> 23.83 | Helped |
| decompose the target flow and aggregate sub-target flows with LTP and VFA | FlowEdit with SD3.5 · PIE-Bench | PSNR: 26.97 -> 27.12 | Helped |
| use simple averaging of decomposed sub-flows (ablation) | FlowEdit with SD3 · PIE-Bench | PSNR: 22.13 -> 23.36 | Helped |
| use simple averaging of decomposed sub-flows (ablation) | FlowEdit with SD3 · PIE-Bench | CLIP Similarity, Edited: 23.67 -> 23.67 | No clear effect |
| add LTP to simple averaging (ablation) | decomposed FlowEdit with SD3 · PIE-Bench | CLIP Similarity, Edited: 23.67 -> 23.82 | Helped |
| add LTP to simple averaging (ablation) | decomposed FlowEdit with SD3 · PIE-Bench | PSNR: 23.36 -> 22.37 | Hurt |
| add VFA to LTP (ablation) | decomposed FlowEdit with SD3 · PIE-Bench | PSNR: 22.37 -> 22.45 | Helped |
| add VFA to LTP (ablation) | decomposed FlowEdit with SD3 · PIE-Bench | CLIP Similarity, Edited: 23.82 -> 23.83 | Helped |
| extend decomposition by setting η_dec to 27 instead of 28 (ablation) | SplitFlow · PIE-Bench | CLIP Similarity, Edited: 23.83 -> 23.85 | Helped |
| remove the maximum of three sub-target prompts (ablation) | SplitFlow · PIE-Bench | Structure Distance ×10^3: 25.96 -> 26.55 | Hurt |
Surprising findings
- Naive averaging of sub-flows alone improves background preservation over the baseline while keeping CLIP similarity on par.
Limitations
- Performance depends on the LLM used for prompt decomposition, and results vary across LLMs.
- Higher inference cost than the FlowEdit baseline (about 15 extra steps plus LLM decomposition time).
- Over-segmentation with unconstrained sub-prompt count (average 4.2) slightly degrades performance.
Experimental setup
Datasets: PIE-Bench
Benchmarks: PIE-Bench
Metrics: PSNR LPIPS MSE SSIM Structure Distance CLIP similarity
Baselines: DDIM Null-text inversion PnP-Inv RF-Inversion RF-Solver FireFlow iRFDS FTEdit FlowEdit
Builds on: FlowEdit Rectified Flow PCGrad
Closest papers by meaning
- Unveil Inversion and Invariance in Flow Transformer for Versatile Image Editing (CVPR 2025 · similarity 0.90)
- EditInfinity: Image Editing with Binary-Quantized Generative Models (NeurIPS 2025 · similarity 0.89)
- FluxSpace: Disentangled Semantic Editing in Rectified Flow Models (CVPR 2025 · similarity 0.89)
- Object-Aware Inversion and Reassembly for Image Editing (ICLR 2024 · similarity 0.89)
- Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations (ICLR 2025 · similarity 0.89)
- PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code (ICLR 2024 · similarity 0.88)
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.