Optimal Transport for Rectified Flow Image Editing: Unifying Inversion-Based and Direct Methods
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Optimal Transport for Rectified Flow Image Editing".
Title and authors: Tom: The paper's main contribution seems to be twofold, right? They introduce a zero-shot transport-guided inversion framework that works during the reverse diffusion process, and they also extend optimal transport principles to improve those direct editing methods by correcting the velocity fields <ref:2508.02363#pg1>.
Lu: That’s a big deal. They are proposing that for inversion-based editing, you can make sure the inversion trajectories follow what's called geodesics in transport space, which means they stay closer to the actual path between images <ref:2508.02363#pg3>.
Meng: If they can enforce those paths mathematically, that should mean better structural coherence when you're trying to invert an image and then edit it. It tackles that trajectory deviation problem directly <ref:2508.02363#pg1>.
Jane: But they also didn't just focus on inversion; they took their optimal transport ideas and applied them to FlowEdit, which is a direct editing method that bypasses the inversion step entirely <ref:2508.02363#pg4>.
Tom: Yeah, that’s interesting because FlowEdit is already better than some other direct methods, but it's missing that principled guidance mechanism to make its pathways more optimized <ref:2508.02363#pg1>.
Lalam: The paper explicitly lists four main contributions: they give you a theoretical framework linking optimal transport distances to the quality of editing across different approaches, and then they show transport-guided inversion with adaptive scheduling for those trade-offs <ref:2508.02363#pg4>.
Lu: That adaptive scheduling part, using cosine annealing for the temporal scope of intervention, sounds like they are being smart about when to apply the transport correction during the denoising process <ref:2508.02363#pg1>.
Meng: I wonder how much it actually helps in practice. If this is just a theoretical thing that doesn't save us time or money, it won't ship at scale <ref:2508.02363#pg4>.
The paper's summary: Tom: So, to summarize what they did: they established that optimal transport theory gives us a mathematical framework for finding the best pathways between distributions, which is key because it aligns well with how we perceive image similarity <ref:2508.02363#pg4>.
Jane: They then apply this theory in two ways: first, for inversion-based editing, they make sure the inversion trajectories look like the shortest paths between images in transport space, which improves structural coherence <ref:2508.02363#pg4>.
Lu: And second, for direct editing methods like FlowEdit, they enhance the velocity field by adding a term related to that optimal transport direction to improve semantic consistency and reduce trajectory deviation <ref:2508.02363#pg1>.
Meng: So essentially, they are trying to make the editing path cleaner whether you're going through inversion or not by using these transport-based corrections instead of just relying on standard flow paths <ref:2508.02363#pg4>.
Tom: And they validated this across FLUX and Stable Diffusion three showing that they see consistent improvements in both paradigms without having to retrain the models or do any test time optimization <ref:2508.02363#pg1>.
Jane: That’s a significant point for us. If it works training-free and is computationally efficient, it means we don't have to introduce massive overhead when we want to apply these edits <ref:2508.02363#pg4>.
Lalam: The paper shows they can achieve results ranging from seventy-eight percent up to ninety-nine point three percent improvement across different metrics, which is a wide range of gains <ref:2508.02363#pg1>.
The paper's improvements: Tom: Let’s look at the actual performance claims. For inversion-based editing, RF-Inversion plus Optimal Transport Inversion, they got LPIPS scores of zero point zero zero one versus a score of zero point one three five, and an SSIM of zero point nine nine two against a baseline of zero point eight three three on face editing benchmarks <ref:2508.02363#pg1>.
Jane: That’s a massive jump in quality for the inversion side, moving from that baseline to those numbers, which shows how much better the structural preservation is when you use transport guidance <ref:2508.02363#pg1>.
Lu: And for the direct editing part, Transport-Enhanced FlowEdit shows consistent improvements in semantic consistency and structure preservation across various editing scenarios, like an eleven point two percent improvement in identity preservation <ref:2508.02363#pg4>.
Meng: Eleven point two percent is a concrete number. That’s what we need to see when we're trying to deploy something that actually works reliably on real traffic or user photos <ref:2508.02363#pg1>.
Tom: And they showed they can control the editing precisely by varying three parameters: controller guidance eta, phase parameter phi, and transport strength beta zero <ref:2508.02363#pg1>.
Jane: That means we have fine-grained control over the trade-off between how much structure you preserve versus how much semantic flexibility you get in the edit <ref:2508.02363#pg1>.
Lalam: They also showed robustness across different datasets, like LSUN datasets, and that it works on both FLUX and Stable Diffusion three architectures <ref:2508.02363#pg4>.
Conclusion: Tom: So wrapping this up, the paper titled "Optimal Transport for Rectified Flow Image Editing: Unifying Inversion-Based and Direct Methods" shows a unified way to handle the trajectory problems in rectified flow editing using optimal transport theory <ref:2508.02363#pg1>.
Jane: The main implication is that we can now improve both the methods that invert images and the direct editing methods by enforcing geometrically principled paths through optimal transport, leading to better structural preservation overall <ref:2508.02363#pg4>.
Lu: What this means for the bigger picture is that we have a method that handles complex image modifications like gender transformation or sequential object insertion while maintaining coherence across multiple steps <ref:2508.02363#pg1>.
Meng: If this framework can handle those complex operations reliably, it moves us closer to having editing tools that are robust enough for real-world applications beyond just simple attribute changes <ref:2508.02363#pg4>.
Lalam: The paper’s limitation is that they didn't explicitly state what the absolute limits are, but they did mention the three parameters—eta, phi, and beta zero—which suggests there are still trade-offs you have to manage when you choose your specific setting <ref:2508.02363#pg1>.
Tom: Exactly. It’s not a magic solution that fixes everything; it’s a structured way to tune the balance between fidelity and flexibility using those transport parameters <ref:2508.02363#pg1>.
Jane: So we're leaving with a framework that brings together theory and practice for image editing in rectified flow models, which is pretty interesting stuff to think about as we keep looking at these new generation models <ref:2508.02363#pg4>.
Marian Lupasćcu, Mihai Sorin Stupariu
Department of Computer Science, University of Bucharest
cs.CV
Submitted: 2025-08-04
Updated: 2025-09-20
Comments: 27 pages, 26 figures, WACV conference
Journal ref: 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 6764-6774
DOI: 10.1109/WACV61042.2026.00654
Code: https://github.com/marianlupascu/OT-RF
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Image editing in rectified flow models remains challenging due to the fundamental trade-off between reconstruction fidelity and editing flexibility While inversion-based methods suffer from…
Key concepts
- Optimal Transport Theory
- This mathematical framework provides a principled way to find the most efficient pathways between different probability distributions. It uses the Wasserstein distance to measure perceptual similarity, ensuring that transformations between images follow geometrically sound paths that preserve structure during editing.
- Inversion-Based Editing
- This method tries to edit an image by first mapping it back into a latent representation using an inversion process. The paper addresses its weakness—trajectory deviation—by using optimal transport to ensure the inversion path stays close to the ideal, semantically meaningful route.
- Transport-Enhanced FlowEdit
- This approach improves direct editing methods (like FlowEdit) by adding guidance from optimal transport during the denoising process. This guidance modifies the image's velocity field, leading to better pathway optimization and superior structural preservation in the final edited image.
Terminology
Summary
Image editing in rectified flow models remains challenging due to the fundamental trade-off between reconstruction fidelity and editing flexibility While inversion-based methods suffer from trajectory deviation, recent inversion-free approaches like FlowEdit offer direct editing pathways but can benefit from additional guidance to improve structure preservation In this work, we demonstrate that optimal transport theory provides a unified framework for improving both paradigms in rectified flow editing. p. 0 Abstract
Image editing in rectified flow models remains challenging
due to the fundamental trade-off between reconstruction fidelity and editing flexibility. While inversion-based methods suffer from trajectory deviation, recent inversion-free
approaches like FlowEdit offer direct editing pathways but
can benefit from additional guidance to improve structure
preservation.Where the paper says it
Unified Framework and Contributions
The authors introduce a zero-shot transport-guided inversion framework that leverages optimal transport during the reverse diffusion process, and extend optimal transport principles to enhance inversion-free methods through transport-optimized velocity field corrections. The main contributions include:
-
A theoretical framework linking optimal transport distances to RF editing quality across different editing approaches. p. 1
Where the paper says it
For inversion-free methods, we enhance FlowEdit with optimal transport guidance for improved semantic consistency. Our main contributions are: (i) A theoretical framework linking optimal transport distances to RF editing quality across different editing approaches; (ii) Transport-guided inversion with adaptive scheduling for optimal reconstruction-editability trade-offs; (iii) Transportenhanced FlowEdit with improved pathway optimization and structural preservation; (iv) Comprehensive evaluation across FLUX [24] and Stable Diffusion 3 [12], demonstrating consistent improvements in both paradigms; (v) Training-free methods preserving rectified flow computational efficiency [26, 27].
-
Transport-guided inversion with adaptive scheduling for optimal reconstruction-editability trade-offs. p. 1
Where the paper says it
For inversion-free methods, we enhance FlowEdit with optimal transport guidance for improved semantic consistency. Our main contributions are: (i) A theoretical framework linking optimal transport distances to RF editing quality across different editing approaches; (ii) Transport-guided inversion with adaptive scheduling for optimal reconstruction-editability trade-offs; (iii) Transportenhanced FlowEdit with improved pathway optimization and structural preservation; (iv) Comprehensive evaluation across FLUX [24] and Stable Diffusion 3 [12], demonstrating consistent improvements in both paradigms; (v) Training-free methods preserving rectified flow computational efficiency [26, 27].
-
Transport-enhanced FlowEdit with improved pathway optimization and structural preservation. p. 1
Where the paper says it
For inversion-free methods, we enhance FlowEdit with optimal transport guidance for improved semantic consistency. Our main contributions are: (i) A theoretical framework linking optimal transport distances to RF editing quality across different editing approaches; (ii) Transport-guided inversion with adaptive scheduling for optimal reconstruction-editability trade-offs; (iii) Transportenhanced FlowEdit with improved pathway optimization and structural preservation; (iv) Comprehensive evaluation across FLUX [24] and Stable Diffusion 3 [12], demonstrating consistent improvements in both paradigms; (v) Training-free methods preserving rectified flow computational efficiency [26, 27].
-
Comprehensive evaluation across FLUX and Stable Diffusion 3, demonstrating consistent improvements in both paradigms. p. 1
Where the paper says it
For inversion-free methods, we enhance FlowEdit with optimal transport guidance for improved semantic consistency. Our main contributions are: (i) A theoretical framework linking optimal transport distances to RF editing quality across different editing approaches; (ii) Transport-guided inversion with adaptive scheduling for optimal reconstruction-editability trade-offs; (iii) Transportenhanced FlowEdit with improved pathway optimization and structural preservation; (iv) Comprehensive evaluation across FLUX [24] and Stable Diffusion 3 [12], demonstrating consistent improvements in both paradigms; (v) Training-free methods preserving rectified flow computational efficiency [26, 27].
-
Training-free methods preserving rectified flow computational efficiency. p. 1
Where the paper says it
For inversion-free methods, we enhance FlowEdit with optimal transport guidance for improved semantic consistency. Our main contributions are: (i) A theoretical framework linking optimal transport distances to RF editing quality across different editing approaches; (ii) Transport-guided inversion with adaptive scheduling for optimal reconstruction-editability trade-offs; (iii) Transportenhanced FlowEdit with improved pathway optimization and structural preservation; (iv) Comprehensive evaluation across FLUX [24] and Stable Diffusion 3 [12], demonstrating consistent improvements in both paradigms; (v) Training-free methods preserving rectified flow computational efficiency [26, 27].
Theoretical Foundation of Optimal Transport
Optimal transport theory provides a mathematical framework for establishing principled pathways between probability distributions. The Wasserstein distance provides natural alignment with perceptual similarity through geometric structure preservation. p. 1 Integration with generative models
has enhanced training stability and sample quality through
Wasserstein GANs [2, 14] and neural optimal transport [21]. The Wasserstein distance provides natural alignment with
perceptual similarity through geometric structure preservation [35], making it well-suited for semantic image editing
applications across different editing paradigms. Approach
Our key insight is that optimal transport theory provides a
unified mathematical framework for both inversion-based
and inversion-free rectified flow editing without requiring
model retraining. We propose a comprehensive approach
that leverages transport-optimal trajectories: the Optimal
Transport Inversion Pipeline for inversion-based editing,
and Transport-Enhanced FlowEdit for direct editing approaches. For
inversion-based editing, our method ensures that inversion
trajectories approximate geodesics in transport space, preserving structural coherence and semantic relationships. For direct editing, we enhance FlowEdit’s pathway construction with transport-theoretic optimization to improve
semantic consistency and reduce trajectory deviation.Where the paper says it
Where the paper says it
Where the paper says it
Implementation Details of Transport Guidance
The authors define the optimal transport direction dOT as a closed-form solution using the Brenier map property for quadratic costs, which reduces computational complexity from O(n 3) to O(n). The temporal scaling 1/(T − t) intensifies transport guidance as the denoising endpoint is approached, and an adaptive weighting schedule α(t) using cosine annealing controls the temporal scope of transport intervention. p. 3 Early stages (t ≈ T) prioritize global
structure preservation during noise removal, while later
stages (t ≈ 0) require precise semantic alignment with
prompt conditioning. We model this temporal evolution through an adaptive weighting schedule: α(t) =
β0 · S T −t
T
, where S implements cosine annealing:
S(s) = 1
2
1 + cos
min
s
ϕ
, 1
· π
. We enhance this approach by
incorporating optimal transport guidance to improve pathway optimization and semantic consistency. The enhanced FlowEdit process modifies the original
FlowEdit velocity computation according to the formula
venh(zt, t) = vFE(zt, t) + γ(t) · dOT(zt, ztarget, t), where
vFE represents the original FlowEdit velocity difference
(vtar − vsrc), and γ(t) is the transport guidance weightWhere the paper says it
Where the paper says it
Experimental Validation and Performance Gains
The unified framework achieves significant improvements across both paradigms. For inversion-based editing, RF-Inversion+OTC shows LPIPS scores of 0. p. 5 Reconstruction Performance
Table 1 presents reconstruction quality across both editing paradigms. For inversion-based methods, our transport guidance achieves improvements: RF-Inversion+OTC
shows LPIPS score of 0. For inversion-based methods, our transport guidance achieves improvements: RF-Inversion+OTC
shows LPIPS score of 0. 135 for baseline RFInversion, SSIM of 0. 001 and SSIM of 0. 992 on face editing benchmarks, observing 7. 9% improvements over RF-Inversion on
LSUN datasets. For inversion-free editing with FlowEdit
on FLUX and Stable Diffusion 3, we demonstrate consistent improvements in semantic consistency and structure
preservation across diverse editing scenarios. Our semantic face editing experiments show an 11. 2% improvement
in identity preservation and enhanced perceptual quality. Optimal performance occurs at β0 = 0. These ablations show that our three-parameter framework enables precise control of editing, with each parameter addressing distinct aspects of the faithfulnesseditability trade-off.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Robustness and Versatility Across Tasks
The framework is validated across various tasks on LSUN datasets, showing improvements in reconstruction quality and semantic editing performance. Qualitative comparisons using the FLUX backbone demonstrate superior edit realism through improved semantic consistency and enhanced structural preservation compared to baseline methods. p. 1 For inversion-free methods, we enhance FlowEdit with optimal transport guidance for improved semantic consistency. Our main contributions are: (i) A theoretical framework linking optimal transport distances to RF editing quality across different editing approaches; (ii)
Transport-guided inversion with adaptive scheduling for optimal reconstruction-editability trade-offs; (iii) Transportenhanced FlowEdit with improved pathway optimization
and structural preservation; (iv) Comprehensive evaluation
across FLUX [24] and Stable Diffusion 3 [12], demonstrating consistent improvements in both paradigms; (v)
Training-free methods preserving rectified flow computational efficiency [26, 27]. For inversion-free methods, we enhance FlowEdit with optimal transport guidance for improved semantic consistency. Our main contributions are: (i) A theoretical framework linking optimal transport distances to RF editing quality across different editing approaches; (ii)
Transport-guided inversion with adaptive scheduling for optimal reconstruction-editability trade-offs; (iii) Transportenhanced FlowEdit with improved pathway optimization
and structural preservation; (iv) Comprehensive evaluation
across FLUX [24] and Stable Diffusion 3 [12], demonstrating consistent improvements in both paradigms; (v)
Training-free methods preserving rectified flow computational efficiency [26, 27]. Figure 7 showcases editing versatility across both
paradigms, demonstrating consistent performance improvements across expression modification, age progression, gender transformation, and object insertion. The generalizability across diverse tasks validates our unified transport guidance framework for rectified flow editing approaches.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
This work establishes a framework for rectified flow editing that addresses trajectory optimization challenges by incorporating transport-theoretic trajectory corrections for both inversion-based and inversion-free methods, operating without model retraining or test-time optimization. The approach achieves consistent improvements—ranging from 7. p. 5 Our transport enhancements provide consistent
improvements regardless of the base method architecture. For inversion-based methods, RF-Inversion+OTC
achieves face recognition distance of 0. 65 dB)
with minimal runtime overhead (<0. While the approach has limitations—Gaussian approximations may not
hold universally, trajectory stability emphasis may constrain extreme transformations, and optimal parameters
vary between paradigms—the framework establishes optimal transport as a viable theoretical foundation for rectified
flow editing with immediate applicability to existing architectures and potential for extension to other generative models facing similar trajectory deviation challenges. Experimental Evaluation
We evaluate our unified optimal transport framework across
both inversion-based and inversion-free rectified flow editing paradigms. Our experiments demonstrate significant
improvements over existing methods through three objectives: (i) pure reconstruction quality, (ii) reconstruction
from corrupted inputs, and (iii) controlled semantic modifications with structure preservation.Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
Optimal Transport for Rectified Flow Image Editing: Unifying Inversion-Based and Direct Methods Marian Lupas¸cu1 Mihai Sorin Stupariu1 marianlupascu15@gmail.com stupariu@fmi. p. 0 Optimal Transport for Rectified Flow Image Editing: Unifying Inversion-Based
and Direct Methods
Marian Lupas¸cu1 Mihai Sorin Stupariu1
marianlupascu15@gmail. com stupariu@fmi. com stupariu@fmi. ro
1Department of Computer Science, University of Bucharest, Romania
(a) Ref. ro
1Department of Computer Science, University of Bucharest, Romania
(a) Ref. content
(b) Ref. Transport-guided rectified flow editing enables image modifications across both paradigms. Our framework enhances
inversion-based and inversion-free methods, applying prompt-based transformations without fine-tuning. Abstract
Image editing in rectified flow models remains challenging
due to the fundamental trade-off between reconstruction fidelity and editing flexibility. While inversion-based methods suffer from trajectory deviation, recent inversion-free
approaches like FlowEdit offer direct editing pathways but
can benefit from additional guidance to improve structure
preservation.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 2 ---
- Introduction Rectified flow (RF) models [13, 26] achieve high-quality text-to-image generation with computational advantages over diffusion models [19, 41], as demonstrated by systems like FLUX [24] or Stable Diffusion 3 [12]. p. 1However, effective image editing in rectified flows remains challenging, requiring either mapping real images back to editable latent representations through inversion, or developing principled direct editing approaches that bypass inversion altogether Image inversion in rectified flows presents unique challenges compared to diffusion models. p. 1
Where the paper says it
Introduction Rectified flow (RF) models [13, 26] achieve high-quality text-to-image generation with computational advantages over diffusion models [19, 41], as demonstrated by systems like FLUX [24] or Stable Diffusion 3 [12]. However, effective image editing in rectified flows remains challenging, requiring either mapping real images back to editable latent representations through inversion, or developing principled direct editing approaches that bypass inversion altogether.
The deterministic nature of RF trajectories eliminates stochastic regularization that aids diffusion model inversion [19], leading to trajectory deviation during the inversion process. p. 1Where the paper says it
However, effective image editing in rectified flows remains challenging, requiring either mapping real images back to editable latent representations through inversion, or developing principled direct editing approaches that bypass inversion altogether. Image inversion in rectified flows presents unique challenges compared to diffusion models.
Small numerical errors compound across the deterministic flow path, causing inverted latents to drift from semantically meaningful representations [42]. p. 1Where the paper says it
Image inversion in rectified flows presents unique challenges compared to diffusion models. The deterministic nature of RF trajectories eliminates stochastic regularization that aids diffusion model inversion [19], leading to trajectory deviation during the inversion process.
This creates a trade-off: methods preserving reconstruction fidelity produce latents resistant to editing, while approaches enabling modification introduce visual artifacts. p. 1Where the paper says it
The deterministic nature of RF trajectories eliminates stochastic regularization that aids diffusion model inversion [19], leading to trajectory deviation during the inversion process. Small numerical errors compound across the deterministic flow path, causing inverted latents to drift from semantically meaningful representations [42].
Existing RF inversion approaches like DDIM inversion [41] adapted to rectified flows accumulate discretization errors, while optimization-based methods like Null-Text Inversion [32] require expensive test-time computation that contradicts RF efficiency advantages. p. 1Where the paper says it
Small numerical errors compound across the deterministic flow path, causing inverted latents to drift from semantically meaningful representations [42]. This creates a trade-off: methods preserving reconstruction fidelity produce latents resistant to editing, while approaches enabling modification introduce visual artifacts.
Recent work has proposed inversion-free alternatives to address these limitations. p. 1Where the paper says it
This creates a trade-off: methods preserving reconstruction fidelity produce latents resistant to editing, while approaches enabling modification introduce visual artifacts. Existing RF inversion approaches like DDIM inversion [41] adapted to rectified flows accumulate discretization errors, while optimization-based methods like Null-Text Inversion [32] require expensive test-time computation that contradicts RF efficiency advantages.
FlowEdit [22] constructs direct pathways between source and target image distributions, bypassing the problematic inversion step entirely While this approach achieves improved structural preservation compared to inversion-based methods, it lacks principled guidance mechanisms for trajectory optimization, potentially leading to suboptimal editing paths and semantic inconsistencies in complex transformations. p. 1Where the paper says it
Existing RF inversion approaches like DDIM inversion [41] adapted to rectified flows accumulate discretization errors, while optimization-based methods like Null-Text Inversion [32] require expensive test-time computation that contradicts RF efficiency advantages. Recent work has proposed inversion-free alternatives to address these limitations.
Optimal transport theory provides a mathematical framework for establishing principled pathways between probability distributions The Wasserstein distance provides natural alignment with perceptual similarity through geometric structure preservation [35]. p. 1Where the paper says it
FlowEdit [22] constructs direct pathways between source and target image distributions, bypassing the problematic inversion step entirely. While this approach achieves improved structural preservation compared to inversion-based methods, it lacks principled guidance mechanisms for trajectory optimization, potentially leading to suboptimal editing paths and semantic inconsistencies in complex transformations.
Where the paper says it
The Wasserstein distance provides natural alignment with perceptual similarity through geometric structure preservation [35], making it well-suited for semantic image editing applications across different editing paradigms. We introduce a unified framework that leverages optimal transport theory to improve rectified flow editing across both paradigms.
--- Page 4 ---
We address these limitations by incorporating optimal transport (OT) theory to compute minimalcost transformations between probability distributions For inversion-based editing, our method ensures that inversion trajectories approximate geodesics in transport space, preserving structural coherence and semantic relationships. For direct editing, we enhance FlowEdit with optimal transport guidance for improved semantic consistency. p. 1 For inversion-free methods, we enhance FlowEdit with optimal transport guidance for improved semantic consistency. Our main contributions are: (i) A theoretical framework linking optimal transport distances to RF editing quality across different editing approaches; (ii)
Transport-guided inversion with adaptive scheduling for optimal reconstruction-editability trade-offs; (iii) Transportenhanced FlowEdit with improved pathway optimization
and structural preservation; (iv) Comprehensive evaluation
across FLUX [24] and Stable Diffusion 3 [12], demonstrating consistent improvements in both paradigms; (v)
Training-free methods preserving rectified flow computational efficiency [26, 27]. For inversion-free methods, we enhance FlowEdit with optimal transport guidance for improved semantic consistency. Our main contributions are: (i) A theoretical framework linking optimal transport distances to RF editing quality across different editing approaches; (ii)
Transport-guided inversion with adaptive scheduling for optimal reconstruction-editability trade-offs; (iii) Transportenhanced FlowEdit with improved pathway optimization
and structural preservation; (iv) Comprehensive evaluation
across FLUX [24] and Stable Diffusion 3 [12], demonstrating consistent improvements in both paradigms; (v)
Training-free methods preserving rectified flow computational efficiency [26, 27].Where the paper says it
Where the paper says it
--- Page 5 ---
Algorithm 1 Transport-Guided RF Inversion Algorithm 1 Input: Source image x0, target image xtarget, target prompt p, RF model vθ, guidance strength η, phase parameter ϕ, transport strength β0 Output: Edited image xˆ0 z0 ← Encode(x0) ▷ Encode to latent space ztarget ← Encode(xtarget) ▷ Encode to latent space zT ← RFInvert(z0, vθ) ▷ Standard inversion for t = T to 0 step −∆t do vtar ← vθ(zt, t, p) ▷ Velocity with target prompt vref ← vθ(zt, t, ∅, z0) ▷ Reference-conditioned velocity vRF ← vtar + η · (vref − vtar) ▷ Controller guidance dOT ← ztarget−zt max(T −t,δ) ▷ Transport direction α ← β0 · S T −t / T do venh ← vRF + α · clip(dOT, τ) ▷ Combined velocity zt−∆t ← zt + ∆t · venh ▷ Euler integration end for xˆ0 ← Decode(z0) ▷ Decode result Algorithm 2 Transport-Enhanced FlowEdit Algorithm 2 Input: Source image x0, source prompt psrc, target prompt ptar, RF model vθ, transport strength βOT, steps T Output: Edited image xˆ0 zsrc ← Encode(x0) zT ← zsrc = Initialize editing trajectory for t = T to 0 step −∆t do ε ∼ N (0, I) zsrc t ← (1 − t) · zsrc + t · ϵ ▷ Forward diffusion ztar t ← zt + zsrc t - zsrc ▷ FlowEdit coupling vsrc ← vθ(zsrc t, t, psrc) vtar ← vθ(ztar, t, ptar) vFE ← vtar − vsrc ▷ Original FlowEdit velocity dOT ← ztar t −z src t max(T −t,δ) γ ← βOT · S t / T venh ← vFE + γ · clip(dOT, τ) zt−∆t ← zt + ∆t · venh end for xˆ0 ← Decode(z0) ▷ Decode result
--- Page 13 ---
Table 4 Optimal hyperparameter configurations for different editing tasks Parameter Symbol Reconstruction Semantic Editing Stroke-to-Image Controller guidance η Transport strength β0 Phase parameter ϕ Starting time s Stopping time τ Clipping threshold τclip Integration steps T Guidance scale w Stability parameter δ Portrait editing requires lower transport strength to preserve facial identity
--- Page 14 ---
Table 5 Optimal hyperparameter configurations for Transport-Enhanced FlowEdit across FLUX and SD3 architectures Parameter Symbol FLUX SD3 Recon. Semantic Stroke-to-Image Recon. p. 4 For stroke-to-image reconstruction, we use
L2 distance, PSNR, and KID [4]. For semantic editing,
we use face recognition distance for identity preservation,
DINO score [9] for structural coherence, CLIP-T [36] for
text–image alignment, and CLIP-I for visual quality. FlowEdit [22] constructs direct pathways between source and target image distributions, bypassing the problematic inversion step entirely. While this approach achieves improved structural preservation compared to inversion-based methods, it lacks principled guidance mechanisms for trajectory optimization, potentially leading to suboptimal editing paths and semantic
inconsistencies in complex transformations. FlowEdit [22] constructs direct pathways between source and target image distributions, bypassing the problematic inversion step entirely. While this approach achieves improved structural preservation compared to inversion-based methods, it lacks principled guidance mechanisms for trajectory optimization, potentially leading to suboptimal editing paths and semantic
inconsistencies in complex transformations.Where the paper says it
Where the paper says it
Where the paper says it
--- Page 15 ---
Figure 8 Controller guidance parameter ablation (η ∈ [0, 0.8]) for age editing using prompt “face of a young man” Parameter controls progression from original image (η = 1) toward promptaligned result (η = 0) Top: RF-Inversion; bottom: our approach Figure 9 Phase parameter ablation (ϕ ∈ [0, 1]) with fixed β0 = 0. p. 7 8])
for age editing using prompt “face of a young man”. Parameter
controls progression from original image (η = 1) toward promptaligned result (η = 0). 6, prompt “face of a young man”. p. 7 Controller guidance parameter ablation (η ∈ [0, 0. 8])
for age editing using prompt “face of a young man”. Controller guidance parameter ablation (η ∈ [0, 0. 8])
for age editing using prompt “face of a young man”.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 17 ---
Figure 14 Parameter interaction: η vs. ϕ. Results for prompt ”face of a young man” with β0 = 0. p. 7 Phase parameter ablation (ϕ ∈ [0, 1]) with fixed β0 =
0. 6, prompt “face of a young man”. Phase parameter ablation (ϕ ∈ [0, 1]) with fixed β0 =
0. 6, prompt “face of a young man”. 8 1
Figure 9. Phase parameter ablation (ϕ ∈ [0, 1]) with fixed β0 =
0. 8 1
Figure 9. Phase parameter ablation (ϕ ∈ [0, 1]) with fixed β0 =
0. Controller guidance parameter ablation (η ∈ [0, 0. 8])
for age editing using prompt “face of a young man”. Controller guidance parameter ablation (η ∈ [0, 0. 8])
for age editing using prompt “face of a young man”.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 18 ---
Figure 14 Parameter interaction: η vs. ϕ. Results for prompt ”face of a young man” with β0 = 0.1 fixed Figure 15 Parameter interaction: η vs. β0. Results for prompt ”face of a young man” with ϕ = 0.3 fixed
--- Page 19 ---
Figure 16 Comparison using Flux backbone Figure 17 Comparison on LSUN-Bedroom dataset Figure 18 Comparison on LSUN-Church dataset.
--- Page 27 ---
Table 1 Reconstruction quality comparison on SFHQ dataset using LPIPS, SSIM, face recognition distance, and CLIP-I similarity metrics under null prompt conditions Table 1 presents reconstruction quality across both editing paradigms.
--- Page 7 ---
Figure 5 Editing method comparison Figure 7 Semantic editing comparison.
--- Page 11 ---
Figure 16 Qualitative comparison using the FLUX backbone.
--- Page 19 ---
Figure 7 Comparison across editing paradigms Figure 7 Semantic editing comparison.
--- Page 14 ---
Table 1 Reconstruction quality comparison on LSUN datasets using LPIPS, SSIM, face recognition distance, and CLIP-I similarity metrics under null prompt conditions.
--- Page 15 ---
Table 2 Stroke-to-image reconstruction comparison on LSUN datasets using L2 distance Table 3 Semantic editing evaluation on facial modification tasks. p. 7 952
Table 3. Semantic editing evaluation on facial modification
tasks.Where the paper says it
--- Page 16 ---
Table 4 Optimal hyperparameter configurations for different editing tasks.
--- Page 17 ---
Figure 8 Controller guidance parameter ablation (η ∈ [0, 0.8]) for age editing using prompt “face of a young man” Figure 9 Phase parameter ablation (ϕ ∈ [0, 1]) with fixed β0 = 0. p. 7 Phase parameter ablation (ϕ ∈ [0, 1]) with fixed β0 =
0. 6, prompt “face of a young man”. Phase parameter ablation (ϕ ∈ [0, 1]) with fixed β0 =
0. 6, prompt “face of a young man”. Transport strength parameter ablation (β0 ∈ [0, 1])
with fixed ϕ = 0. 6, prompt “face of a young man”. Transport strength parameter ablation (β0 ∈ [0, 1])
with fixed ϕ = 0. 6, prompt “face of a young man”.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 27 ---
Table 6 Comparison of FlowEdit-SD3 and FlowEdit-Flux across different transport strengths (TS) Table 7 Comparison of FlowEdit results (SD3 and Flux) across different transport strengths (TS).
--- Page 14 ---
Table 1 Reconstruction quality comparison on LSUN datasets using LPIPS, SSIM, face recognition distance, and CLIP-I similarity metrics under null prompt conditions Table 2 Stroke-to-image reconstruction comparison on LSUN datasets using L2 distance. Table 3 Semantic editing evaluation on facial modification tasks.
--- Page 17 ---
Figure 8 Controller guidance parameter ablation (η ∈ [0, 0.8]) for age editing using prompt “face of a young man” Figure 9 Phase parameter ablation (ϕ ∈ [0, 1]) with fixed β0 = 0.3, η = 0.6, prompt “face of a young man” Figure 10 Transport strength parameter ablation (β0 ∈ [0, 1]) with fixed ϕ = 0.1, η = 0.6, prompt “face of a young man”.
--- Page 14 ---
Table 1 Reconstruction quality comparison on LSUN datasets using LPIPS, SSIM, face recognition distance, and CLIP-I similarity metrics under null prompt conditions Table 2 Stroke-to-image reconstruction comparison on LSUN datasets using L2 distance Table 3 Semantic editing evaluation on facial modification tasks.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Optimal Transport Inversion Pipeline for inversion-based editing allows methods to
approximate geodesics in transport space
for inversion trajectories, ensuring thatreconstruction fidelity and editing flexibility
are balanced by minimizing trajectory deviation during the reverse diffusion process. -
Transport-enhanced FlowEdit improves direct editing by modifying the velocity field with optimal transport guidance, resulting in enhanced velocities that allow the method to achieve
improved pathway optimization and structural preservation.
-
The framework enables inversion-based methods to achieve high fidelity reconstruction, such as
LPIPS scores of 0.001 and SSIM of 0.992 on face editing benchmarks,
showing significant improvements over baselines like RF-Inversion on LSUN datasets. -
Inversion-free editing methods benefit from transport enhancement by achieving consistent improvements in semantic consistency and structure preservation across diverse editing scenarios, such as an
11.2% improvement in identity preservation and enhanced perceptual quality
in semantic face editing experiments. -
The system can perform controlled attribute modifications, including
gender transformation
andage manipulation,
while maintaining facial identity throughout the process via the controller guidance parameter eta, which is shown to result insmoother interpolation and better preservation of facial identity across the full parameter range.
-
The framework supports sequential editing tests where it maintains coherence through multiple modifications, demonstrating a
compositional property emerging from our transport formulation maintaining consistent reference to the original image.
-
The system can handle complex multi-modal transformations by simultaneously performing
Disney cartoon stylization and facial expression generation,
balancing competing objectives through itsprincipled geometric framework.
-
The model can generate high-quality text-to-image outputs across diverse prompts, such as generating images of a
dragon soaring through the sky
or agenius scientist,
maintaining visual coherence and faithful adherence to text descriptions even at lower discretization steps.
Abstract
Image editing in rectified flow models remains challenging due to the fundamental trade-off between reconstruction fidelity and editing flexibility. While inversion-based methods suffer from trajectory deviation, recent inversion-free approaches like FlowEdit offer direct editing pathways but can benefit from additional guidance to improve structure preservation. In this work, we demonstrate that optimal transport theory provides a unified framework for improving both paradigms in rectified flow editing. We introduce a zero-shot transport-guided inversion framework that leverages optimal transport during the reverse diffusion process, and extend optimal transport principles to enhance inversion-free methods through transport-optimized velocity field corrections. Incorporating transport-based guidance can effectively balance reconstruction accuracy and editing controllability across different rectified flow editing approaches. For inversion-based editing, our method achieves high-fidelity reconstruction with LPIPS scores of 0.001 and SSIM of 0.992 on face editing benchmarks, observing 7.8% to 12.9% improvements over RF-Inversion on LSUN datasets. For inversion-free editing with FlowEdit on FLUX and Stable Diffusion 3, we demonstrate consistent improvements in semantic consistency and structure preservation across diverse editing scenarios. Our semantic face editing experiments show an 11.2% improvement in identity preservation and enhanced perceptual quality. The unified optimal transport framework produces visually compelling edits with superior detail preservation across both inversion-based and direct editing paradigms. Code is available for RF-Inversion and FlowEdit at: https://github.com/marianlupascu/OT-RF
Sources
- FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing
- On the Relation between Rectified Flows and Optimal Transport
- Direct Inversion: Boosting Diffusion-based Editing with 3 Lines of Code
- FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models
- FLUX-Text: A Simple and Advanced Diffusion Transformer Baseline for Scene Text Editing
- Steering Rectified Flow Models in the Vector Field for Controlled Image Generation
- Taming Rectified Flow for Inversion and Editing
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs