acceptodds
Under review as a conference paper at ICLR 2027

Free Lunch in Diffusion Inversion: Tensorial Classifier-Free Guidance at Its Origin

Abstract

Inversion errors can compromise source preservation before semantic editing begins. We propose POLARIS, a lightweight, training-free guidance plugin for improving the source trajectory in diffusion-based editing. Building on reconstruction-residual control in diffusion inversion, POLARIS specializes the local inverse-consistency objective to the classifier-free guidance (CFG) coefficients. With adjacent network predictions held fixed, the deterministic DDIM cycle residual is a timestep-dependent scalar multiple of the adjacent guided-noise discrepancy. This yields a coordinate-wise least-squares update using existing conditional and unconditional predictions, without gradient-based optimization or an auxiliary network. Structured-guidance ablations and measured cycle residuals support the local analysis, while comparisons with both high-CFG and null-branch fixed-scale controls demonstrate improved reconstruction fidelity. Paired evaluations further show better background preservation when POLARIS is integrated with existing image, audio, and video editors. A separate transport-based extension improves background fidelity on Wan2.1 video editing without additional model evaluations. Across the measured image backbones, peak memory overhead is 2.0–3.2%. These results support adaptive guidance as a practical component for source preservation; improved inversion alone does not guarantee target-semantic editing.

Primary area:
generative models
Keywords:
Diffution models, image editing

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fDyzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Diffusion-based image editing typically first inverts a source image into noise space and then regenerates it. Inversion errors can cause the reconstructed image to deviate from the original, compromising preservation outside the edited region. Conventional CFG uses a fixed scalar and cannot readily accommodate differences in error across spatial locations and feature channels. This paper proposes POLARIS, a training-free method for tensor-level adaptive CFG. It establishes an analytical relationship between inversion error and CFG coefficients and derives directly solvable, coordinate-wise optimal guidance coefficients, enabling improved inversion accuracy through low-cost inference-time control.

Critical strengths

The method is lightweight and relatively novel, the derivation is reasonable, and the experiments show good visual results.

Critical weaknesses

The executable algorithm and the evidence connecting configurations across experiments remain insufficiently clear. The closed-form solution involves adjacent predictions, whereas the implementation description only states that predictions are “updated from the cache and replayed.” The paper needs to specify which step produces each prediction, which transition uses the resulting coefficients, whether there is a one-step lag, and all stabilization hyperparameters.

Questions for authors

Which differences in data, scheduling, or initialization account for the discrepancy between the correct replay results in Table 12 and the 50-step COCO results in Table 1? Please provide the exact sample counts, selection rules, backbone/code versions, and a configuration mapping table for all experiments.

3 comments

We thank the reviewer for the constructive feedback and for identifying the need for a clearer description of the executable algorithm and the correspondence between experimental configurations.

Below we address all raised concerns point by point.

1. Clarification of the executable algorithm and stabilization settings

We clarify the actual inversion and reconstruction procedure below.

At inversion step tt, the currently available coefficient ωt\omega_t is used to combine the unconditional and source-conditioned predictions at the current latent xtx_t,

ϵ^t=ϵu,t+ωt⊙(ϵc,t−ϵu,t).\hat{\epsilon}_t = \epsilon_{u,t} + \omega_t \odot \left(\epsilon_{c,t} - \epsilon_{u,t}\right).

The DDIM inversion transition then uses this guided prediction to produce xt+1x_{t+1}. After xt+1x_{t+1} is obtained, the unconditional and conditional predictions at this new latent are evaluated and cached. Together with the cached predictions at xtx_t, they provide the adjacent prediction differences,

Δϵu=ϵu,t+1−ϵu,t,\Delta\epsilon_u = \epsilon_{u,t+1} - \epsilon_{u,t},

Δϵc=ϵc,t+1−ϵc,t.\Delta\epsilon_c = \epsilon_{c,t+1} - \epsilon_{c,t}.

For each latent coordinate ii, we define the direction

di=Δϵc,i−Δϵu,i,d_i = \Delta\epsilon_{c,i} - \Delta\epsilon_{u,i},

and compute the raw coefficient as

ωraw,t+1,i=−Δϵu,idi.\omega_{\mathrm{raw},t+1,i} = -\frac{\Delta\epsilon_{u,i}}{d_i}.

The newly computed coefficient ωt+1\omega_{t+1} is not used to generate xt+1x_{t+1}. It is first applied to the subsequent inversion transition. Thus, the implementation has an explicit one-step lag. The coefficient ωt\omega_t generates xt+1x_{t+1}, while ωt+1\omega_{t+1} is computed from the adjacent predictions at xtx_t and xt+1x_{t+1} and controls the next transition. This ordering avoids circular dependence between the latent state and the coefficient estimated using the prediction at that state.

For numerical stability, the SD1.5 implementation uses the following coordinate-wise safeguards.

  • The direction dd is evaluated in FP32.
  • The denominator threshold is ϵnum=10−12\epsilon_{\mathrm{num}}=10^{-12}.
  • If ∣di∣≤ϵnum\lvert d_i \rvert \leq \epsilon_{\mathrm{num}}, the coordinate retains its previous coefficient rather than performing an unstable division or reverting to ω=1\omega=1.
  • Otherwise, the raw coefficient is clipped to [0,2][0,2] before it is used.

The coefficient sequence is initialized as ω0=s01\omega_0=s_0\mathbf{1}, consistent with the implementation described in the paper.

During reconstruction, the stored inversion-order coefficient sequence is replayed in reverse order. Starting from xTx_T, reconstruction applies ωT,ωT−1,…,ω1\omega_T, \omega_{T-1}, \ldots, \omega_1 to the corresponding reverse transitions. Equivalently, at reconstruction-loop index kk, the applied coefficient is ωT−k+1\omega_{T-k+1}. This ensures that each stored coefficient is paired with its corresponding noise level.

The prediction cache is also important for the computational cost. The predictions evaluated at xt+1x_{t+1} are retained and reused as the predictions at the next inversion step. Therefore, coefficient estimation does not require additional denoiser evaluations. The total number of denoiser evaluations remains the same as that of the corresponding CFG pipeline.

2. Clarification of the discrepancy between the main reconstruction and replay-strategy ablation results

The two experiments use different COCO2017 evaluation subsets:

  • Table 1 (Main reconstruction experiment): COCO2017 Sub-1000, containing 1,000 images, with an MSE of 349 at 50 steps.
  • Table 18 (Replay-strategy ablation): COCO2017 Sub-100, containing 100 images, with an MSE of 498 for reverse-order replay at 50 steps.

Because these values are computed over different evaluation subsets, the absolute MSE values 349 and 498 should not be directly compared as if they were measured on the same image pool. The difference does not indicate an inconsistency in the reconstruction procedure.

The replay-strategy ablation is specifically designed to investigate the effect of aligning the stored guidance coefficients with their corresponding reconstruction timesteps. Within the Sub-100 evaluation setting, reverse-order replay achieves an MSE of 498, outperforming forward-order replay (642), cosine replay (833), and fixed guidance with ω=1\omega=1 during reconstruction/sampling (1171).

We also clarify that the fixed-ω=1\omega=1 row in Table 18 refers to using fixed guidance during reconstruction/sampling after the adaptive inversion trajectory has been obtained. It is not the same protocol as using fixed guidance throughout both inversion and reconstruction, which is evaluated separately in Table 17.

We have corrected the table cross-reference in the main text: the replay-strategy ablation is presented as Table 18 in Appendix B.9, whereas Table 12 reports fallback and clipping statistics in Appendix B.2.

3. Clarification of experimental configurations

To improve reproducibility, we further summarize the experimental configurations and their correspondence to the tables in the manuscript. In particular, we distinguish the main reconstruction evaluation, timestep-wise diagnostics, numerical-stability analysis, guidance-granularity ablation, and replay-strategy ablation.

ExperimentBackboneDataset and sample countScheduler / StepsGuidance configuration
Main reconstruction (Table 1)SD1.5COCO2017 Sub-1000 (1,000 images); Pick-a-Pic (sample count not separately specified in the table)Deterministic DDIM, 10–100 stepsFixed CFG 7.5 vs. tensor-valued POLARIS; reverse-order replay
Timestep-wise reconstruction diagnostic (Table 2)SD1.5Pick-a-Pic (sample count not separately specified in the table)50-step DDIMFixed ω=1\omega=1 vs. POLARIS; predicted clean-sample MSE tracked across reconstruction steps
Local cycle-residual analysis (Table 9)SD1.5COCO2017, 100 images49 adjacent transitionsFixed guidance vs. POLARIS
Numerical-stability analysis (Tables 10–12)SD1.5Same 100-image evaluation; 49 coefficient updates per image49 transitions; latent dimensions 4×64×644 \times 64 \times 64FP32 denominator direction, ϵnum=10−12\epsilon_{\mathrm{num}}=10^{-12}, previous-coefficient fallback, and clipping to [0,2][0,2]
Guidance-granularity ablation (Table 13)DDIM-based inversion settingDataset and sample count not separately restated in the table caption50-step DDIMFixed, Scalar-, Channel-, Spatial-, and Tensor-POLARIS
Replay-strategy ablation (Table 18)DDIM-based reconstructionCOCO2017 Sub-100 (100 images)10–100 inference stepsReverse-order, forward-order, cosine-decay, and fixed-ω=1\omega=1 sampling
FreeInv / ReNoise integration (Table 7)SD1.5Pick-a-Pic, 100 samples50 stepsOriginal host pipeline vs. host pipeline + POLARIS; seed 42

Several additional implementation details are worth emphasizing.

First, the main reconstruction experiment uses the same SD1.5 backbone and deterministic DDIM scheduler for the fixed-scale baseline and POLARIS. The principal difference is whether the guidance coefficient remains fixed at 7.5 or is adaptively updated as a tensor across latent coordinates and timesteps.

Second, the numerical-stability statistics in Tables 10–12 are collected from 100 images over 49 coefficient-update transitions per image, yielding 4,900 image-transition pairs. The latent dimension for this analysis is 4×64×644 \times 64 \times 64. The reported fallback frequency measures coordinate-wise denominator degeneracy rather than solver failures.

Third, the guidance-granularity ablation in Table 13 uses the same 50-step DDIM inversion setting across all parameterizations, ensuring that the comparison isolates the effect of guidance granularity rather than differences in the sampling schedule.

Finally, the FreeInv and ReNoise integration experiments use a matched setting with SD1.5, 100 Pick-a-Pic samples, 50 inference steps, and seed 42. These experiments evaluate the incremental benefit of integrating POLARIS into the respective host pipelines, rather than presenting a standalone ranking of the inversion methods.

We have consolidated these settings in the revised appendix and explicitly distinguish configurations that are shared across experiments from those specific to individual evaluations. Exact checkpoint identifiers, preprocessing details, and implementation versions have been reported from the corresponding experiment configurations rather than inferred from the benchmark settings.

We thank the reviewer for helping us improve the clarity, implementation transparency, and reproducibility of the manuscript.

Reviewer_st4BOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

POLARIS adjusts classifier-free guidance separately at each latent coordinate during inference. The authors hold adjacent conditional and unconditional predictions fixed, then solve a local least-squares problem to reduce the DDIM inversion–reconstruction mismatch. The resulting coefficients are stored and replayed on the source-preservation branch during reconstruction. A separate channel-wise controller extends the idea to Wan2.1/FlowEdit. The experiments cover image reconstruction and editing, as well as audio and video editing. Most of the reported gains are in reconstruction and source preservation; the image experiments also measure target-prompt alignment.

Critical strengths
  1. The controller is a small change with useful ablations behind it. According to the paper, it reuses predictions already computed by the host pipeline. In Table 7, tensor guidance gives the best reconstruction metrics among the scalar, channel-wise, spatial, and tensor variants. I also appreciate the fixed-zero-guidance control in Table 8: it is much stronger than the CFG 7.5 baseline, and POLARIS still improves several metrics. This gives the idea more support than a high-CFG comparison alone would.

  2. The plugin improves preservation when added to several existing editors. Table 2 includes paired results for DI+P2P, DCI+P2P, SAGE, RF-Edit, and DNAEdit. For DCI+P2P, for example, PSNR goes from 29.38 to 29.89 and scaled LPIPS from 33.01 to 28.58, while CLIP is unchanged at the reported precision. Holding the editor fixed makes these gains easier to interpret.

Critical weaknesses
  1. I can follow the closed-form update, but I couldn't work out the exact algorithm used in the experiments. Equation (19) needs predictions at two adjacent states, and the later state depends on the guidance used to reach it. Section 3.2 mentions cached predictions; Appendix A.1 also makes clear that the frozen-prediction calculation is different from the actual rollout. What I still need is the update order: which coefficient creates each state, when the new coefficient takes effect, and which coefficient is replayed on the reverse step. The denominator threshold and any clipping should be given too. These choices affect reconstruction and whether another model call is needed. For the Wan extension, the cached predictor and the range, step-change, and smoothing settings also remain unclear in Section 3.3 and Appendix A.7.

  2. The closest low-cost alternative is missing from the quantitative comparison. Appendix A.5 identifies FreeInv as particularly close, but there is no matched experiment with it. The fixed-zero, fixed-one, high-CFG, and granularity controls are useful. I'd still like to know what POLARIS offers over an existing inexpensive method for reducing trajectory mismatch. The audio and video experiments show preservation gains, but they don't measure whether the requested edits were completed. For my assessment, a matched image inversion/editing comparison would add more than another modality.

  3. There are several typos and presentation issues. Equation references repeatedly appear as "Eq. equation ...". On p. 9, Figure 6 shows the 50-step initialization experiment. The separate random-guidance experiment is also described as using 50 steps (line 467), but its corresponding Figure 7 extends to 100 inversion steps. The step count for the random-guidance experiment should be clarified. On p. 26, line 1378, "drawingsimages" in "Colored pencil drawingsimages" runs two words together, and on line 1384, "A African" should be "An African". In Figure 12 on p. 28, the bottom-left example shows a pedestrian on a street, but the text below it reads "Full glass of water standing on a mountain".

Questions for authors
  1. Could you give pseudocode for the full inversion and reconstruction procedure? Please show which coefficient creates each latent, which one is replayed in reverse, the denominator threshold, any clipping, and the model-call count.
  2. How exactly is the Transport-POLARIS predictor constructed? What values do you use for the ridge term, coefficient limits, inter-step limits, and EMA? Are they fixed across FiVE-100?
  3. Do you have a comparison with FreeInv using the same backbone, prompts, step count, and downstream editor? That would help me judge whether the method improves the fidelity–cost trade-off.
5 comments

We thank the reviewer for the careful assessment and constructive suggestions. We appreciate the recognition of POLARIS as a lightweight controller that reuses existing network predictions and improves source preservation when integrated into existing editing pipelines.

We have further clarified the executable algorithm, provided the complete numerical configuration of Transport-POLARIS, expanded the matched comparisons with existing inversion methods, and corrected the presentation issues identified by the reviewer. We address each concern below.

1. Executable inversion/reconstruction algorithm and numerical safeguards

We clarify the complete execution procedure below.

At inversion step tt, the currently available coefficient ωt\omega_t is used to combine the unconditional and source-conditioned predictions at the current latent x~t\tilde{\mathbf{x}}_t:

ϵ^t=ϵφ,t+ωt⊙(ϵc,t−ϵφ,t).\hat{\epsilon}_t = \epsilon_{\varphi,t} + \omega_t \odot (\epsilon_{c,t} - \epsilon_{\varphi,t}).

The deterministic DDIM inversion transition then uses this guided prediction to produce x~t+1\tilde{\mathbf{x}}_{t+1}. After obtaining the new latent, we evaluate and cache its unconditional and conditional predictions. Together with the cached predictions at x~t\tilde{\mathbf{x}}_t, they provide the adjacent prediction differences:

Δϵφ,t+1=ϵφ,t+1−ϵφ,t,\Delta\epsilon_{\varphi,t+1} = \epsilon_{\varphi,t+1} - \epsilon_{\varphi,t},

Δϵc,t+1=ϵc,t+1−ϵc,t.\Delta\epsilon_{c,t+1} = \epsilon_{c,t+1} - \epsilon_{c,t}.

For each latent coordinate ii, define

δt+1,i=Δϵφ,t+1,i−Δϵc,t+1,i.\delta_{t+1,i} = \Delta\epsilon_{\varphi,t+1,i} - \Delta\epsilon_{c,t+1,i}.

The unconstrained coordinate-wise minimizer of Eq. (19) is

ωraw,t+1,i=Δϵφ,t+1,iδt+1,i.\omega_{\mathrm{raw},t+1,i} = \frac{\Delta\epsilon_{\varphi,t+1,i}}{\delta_{t+1,i}}.

The implementation then applies the following numerical safeguards:

  • The adjacent prediction differences and denominator direction are evaluated in FP32.
  • The denominator threshold is ϵnum=10−12\epsilon_{\mathrm{num}} = 10^{-12}.
  • If ∣δt+1,i∣≤ϵnum\lvert\delta_{t+1,i}\rvert \leq \epsilon_{\mathrm{num}}, the coordinate retains its previous coefficient ωt,i\omega_{t,i}, avoiding unstable division.
  • Otherwise, the raw coefficient is clipped to [0,2][0,2].

Importantly, the newly computed coefficient ωt+1\omega_{t+1} is not used to generate x~t+1\tilde{\mathbf{x}}_{t+1}. Instead, it is applied to the subsequent inversion transition. Thus, the implementation has an explicit one-step lag: ωt\omega_t generates x~t+1\tilde{\mathbf{x}}_{t+1}, while ωt+1\omega_{t+1} is computed from the adjacent predictions at x~t\tilde{\mathbf{x}}_t and x~t+1\tilde{\mathbf{x}}_{t+1} and controls the next transition. This ordering avoids circular dependence between the latent state and the coefficient estimated using predictions at that state.

The coefficient sequence is initialized as

ω0=s01,\omega_0 = s_0\mathbf{1},

consistent with the implementation described in the manuscript.

The complete execution order can therefore be summarized as follows:

  1. Initialize the latent and guidance coefficient, and evaluate the unconditional and conditional predictions at the initial latent.
  2. Use the current coefficient ωt\omega_t and cached predictions to perform the inversion transition from x~t\tilde{\mathbf{x}}_t to x~t+1\tilde{\mathbf{x}}_{t+1}.
  3. Evaluate and cache the two prediction branches at the newly obtained latent.
  4. Compute the adjacent prediction differences and obtain ωt+1\omega_{t+1} using the closed-form coordinate-wise update, denominator fallback, and clipping.
  5. Store the updated coefficient and reuse the cached endpoint predictions at the next inversion step.
  6. After inversion, initialize reconstruction from x~T\tilde{\mathbf{x}}_T and replay the stored coefficients in reverse timestep order.

Specifically, reconstruction applies

ωT,ωT−1,…,ω1\omega_T,\omega_{T-1},\ldots,\omega_1

to the corresponding reverse transitions. At reconstruction-loop index kk, the applied coefficient is ωT−k+1\omega_{T-k+1}. This ensures that each stored coefficient is paired with its corresponding noise level.

Model-evaluation cost. The predictions evaluated at x~t+1\tilde{\mathbf{x}}_{t+1} are retained and reused at the next inversion step. Consequently, coefficient estimation requires no additional denoiser evaluations. The total number of denoiser evaluations remains the same as that of the corresponding host CFG pipeline, which already provides the required conditional and unconditional prediction branches. During reconstruction, the stored coefficients are replayed without additional coefficient-estimation calls.

We have also clarified the computational complexity: the coordinate-wise update requires O(d)O(d) arithmetic per step for dd latent coordinates, while storing the coefficient sequence requires O(Td)O(Td) additional guidance storage. The measured peak GPU-memory overhead on the tested image backbones is 2.0–3.2%.

2. Detailed construction and numerical configuration of Transport-POLARIS

We have expanded Appendix A.7 and now explicitly report the predictor construction, ridge regularization, coefficient constraints, inter-step limits, and EMA settings.

Unlike standard POLARIS, which controls an adjacent DDIM cycle residual using coordinate-wise coefficients, Transport-POLARIS operates on the coupled source–target transport direction in FlowEdit.

Let b∈s,tb \in \\{s,t\\} denote the source and target branches. For each latent channel, we define the guidance residuals

fb,k=vb,kc−vb,ku,\mathbf{f}_{b,k} = \mathbf{v}_{b,k}^{c} - \mathbf{v}_{b,k}^{u},

and the baseline transport direction

dk0=(vt,ku+gt0ft,k)−(vs,ku+gs0fs,k),\mathbf{d}_k^0 = (\mathbf{v}_{t,k}^{u} + g_t^0\mathbf{f}_{t,k}) - (\mathbf{v}_{s,k}^{u} + g_s^0\mathbf{f}_{s,k}),

where the baseline source and target guidance scales are

(gs0,gt0)=(5,10).(g_s^0,g_t^0) = (5,10).

The guidance adjustments are parameterized as

gs=gs0+as,gt=gt0+at.g_s = g_s^0 + a_s, \qquad g_t = g_t^0 + a_t.

With the current network predictions and the cached transport predictor held fixed, define

rk=dk0−d^k,Gk=[−fs,k,ft,k],ak=[asat].\mathbf{r}_k = \mathbf{d}_k^0 - \widehat{\mathbf{d}}_k, \qquad G_k = [-\mathbf{f}_{s,k},\mathbf{f}_{t,k}], \qquad \mathbf{a}_k = \begin{bmatrix} a_s \\\\ a_t \end{bmatrix}.

The coefficient update solves the two-variable ridge-regularized least-squares problem:

\mathbf{a}_k^\star = \arg\min_{\mathbf{a}} \left\\{ \lVert\mathbf{r}_k + G_k\mathbf{a}\rVert_2^2 + \lambda_k\lVert\mathbf{a}\rVert_2^2 \right\\},

with the closed-form solution

ak⋆=−(Gk⊤Gk+λkI2)−1Gk⊤rk.\mathbf{a}_k^\star = -(G_k^\top G_k + \lambda_k I_2)^{-1} G_k^\top\mathbf{r}_k.

The ridge parameter is not an arbitrary manually selected constant. It is scaled according to the magnitude of the two guidance residuals:

λk=10−3max⁡(∥fs,k∥22+∥ft,k∥222,10−12)+10−12.\lambda_k = 10^{-3} \max\left( \frac{\lVert\mathbf{f}_{s,k}\rVert_2^2 + \lVert\mathbf{f}_{t,k}\rVert_2^2}{2}, 10^{-12} \right) + 10^{-12}.

This regularization controls potentially large coefficient changes when the source and target guidance residuals are nearly collinear.

Numerical stabilization and hyperparameters. The exact settings used in FiVE-100 are summarized below.

ComponentActual setting
Controller varianttransport_polaris_channel_ema
Sampling schedule50-step UniPC
Warm-up / active steps16 skipped warm-up steps / 34 active steps
Flow shift5.0
Probe configurationFixed probe noise; probe time set to the previous step
Baseline source/target CFG5 / 10
Guidance branchesBoth source and target
Guidance coefficient bounds[0, 15] for each branch
Source/target correction gains1.0 / 1.0
Relative ridge multiplier10−310^{-3}, scaled by the mean squared norm of the two guidance residuals
Maximum inter-step coefficient change2.0 per branch
EMA smoothing0.9ωprev+0.1ωcandidate0.9\omega_{\mathrm{prev}} + 0.1\omega_{\mathrm{candidate}}
Correction multiplierηmin⁡=0,ηmax⁡=2.0\eta_{\min}=0, \quad \eta_{\max}=2.0
Line-search denominator safeguardϵls=10−12\epsilon_{\mathrm{ls}}=10^{-12}

After applying the coefficient-range constraints, inter-step change limits, and EMA smoothing, denote the stabilized coefficient adjustment by aˉk\bar{\mathbf{a}}_k. We then define

qk=Gkaˉk\mathbf{q}_k = G_k\bar{\mathbf{a}}_k

and perform a bounded line search:

ηk=clip⁡[0,2](−⟨rk,qk⟩∥qk∥22+10−12).\eta_k = \operatorname{clip}_{[0,2]}\left( -\frac{\langle\mathbf{r}_k,\mathbf{q}_k\rangle}{\lVert\mathbf{q}_k\rVert_2^2 + 10^{-12}} \right).

The final transport direction is

dkTP=dk0+ηkqk.\mathbf{d}_k^{\mathrm{TP}} = \mathbf{d}_k^0 + \eta_k\mathbf{q}_k.

The line search ensures that the frozen-predictor residual does not increase relative to the baseline direction.

Predictor construction and parameter consistency. The FiVE-100 experiments use the one-step predictor variant. After the first active step, the predictor is the previously accepted transport direction, which is cached and held fixed during the current coefficient solve. At the first active step, no preceding predictor is available; therefore, the method uses the unmodified FlowEdit transport direction and initializes the source and target guidance scales at 5 and 10.

All numerical settings listed above are fixed across all 100 FiVE-100 cases. No per-sample hyperparameter tuning is performed.

Model-evaluation cost. Transport-POLARIS reuses the four conditional/unconditional source and target flow predictions already computed by FlowEdit. The coefficient estimation requires only channel-wise inner products and a small 2×22 \times 2 linear system, followed by the bounded correction. Therefore, it introduces no additional Wan DiT evaluations. The 50-step UniPC schedule and its 34 active steps remain unchanged.

These details are now explicitly reported in Appendix A.7 and Table 8.

3. Clarification of reconstruction results and replay-strategy ablation

The two experiments use different evaluation subsets:

  • Main reconstruction experiment (Table 1): COCO2017 Sub-1000, containing 1,000 images. The 50-step POLARIS MSE is 349.
  • Replay-strategy ablation (Table 18): COCO2017 Sub-100, containing 100 images. The 50-step reverse-order replay MSE is 498.

These values are computed over different image subsets and should not be directly compared as if they were measured on the same evaluation pool. The difference therefore does not indicate an inconsistency in the reconstruction procedure.

Within the Sub-100 evaluation setting, the reverse-order replay strategy achieves an MSE of 498, outperforming:

  • Forward-order replay: 642.
  • Cosine-decay replay: 833.
  • Fixed guidance with ω=1\omega=1 during reconstruction: 1171.

The fixed-ω=1\omega=1 row in Table 18 refers specifically to using fixed guidance during reconstruction/sampling after the adaptive inversion trajectory has already been obtained. It is different from the separate experiment in Table 17, which uses fixed guidance throughout both inversion and reconstruction.

We have also clarified the table cross-references: the replay-strategy ablation is reported in Table 18 in Appendix B.9, whereas Table 12 reports fallback and clipping statistics in Appendix B.2.

Reviewer_NBfAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes POLARIS, a training-free inversion controller that replaces scalar CFG with tensor-valued guidance. A coordinate-wise closed-form update minimizes a local DDIM cycle residual under frozen predictions. A transport-based extension adapts the approach to FlowEdit. Experiments show improved reconstruction and source preservation across image, audio, and video pipelines.

Critical strengths
  • The method directly connects adjacent prediction mismatch to local inversion error and derives an inexpensive closed-form correction. It requires neither training nor gradient optimization, with small reported runtime and memory overhead when the required prediction branches are already available.
  • Integration with several existing editors demonstrates practical improvements in source preservation. Guidance-granularity and replay ablations provide supporting evidence that the gains depend on the proposed controller rather than arbitrary CFG variation.
Critical weaknesses
  • The closed-form solution optimizes frozen endpoint predictions, whereas changing guidance also changes subsequent states and predictions. Although the paper acknowledges this limitation, it does not sufficiently explain how the approximately 10% average reduction in measured cycle residual translates into the reported reconstruction gains.
  • The closed-form update in Eq. (19) can produce large, perturbation-sensitive guidance coefficients when the denominator approaches zero. Although the paper introduces a threshold-based fallback, it does not report the threshold value, fallback frequency, coefficient ranges, or an analysis of samples with extreme coefficients. The practical robustness of this central update therefore remains insufficiently established.
  • The main reconstruction baseline uses CFG 7.5, which produces substantial reconstruction drift. Lower-CFG controls are useful, but matched reconstruction and computational-cost comparisons with closely related methods such as FreeInv, ReNoise, and Null-text Inversion are missing, leaving the method’s relative competitiveness unclear.
  • The Wan evaluation primarily measures background preservation without target-semantic accuracy or uncertainty estimates. Consequently, stronger preservation cannot be clearly distinguished from reduced editing strength. Several image-editing gains are also modest, and their statistical reliability is not established.
Questions for authors
  • How does the measured reduction in actual cycle residual relate to final reconstruction quality, and which timesteps contribute most to the improvement?
  • Why does the default reverse-replay strategy in Table 12 report different 50-step results from POLARIS in Table 1? Are the experimental settings different?
  • What numerical threshold is used, how frequently is the fallback triggered, and do extreme guidance coefficients correlate with reconstruction failures or artifacts?
  • Could you provide matched reconstruction comparisons with FreeInv, ReNoise, and Null-text Inversion, including runtime, model evaluations, and auxiliary storage?
9 comments

We thank the reviewer for the thoughtful assessment and constructive comments. We appreciate the recognition of POLARIS’s lightweight coordinate-wise controller, its analytical formulation, and its integration with existing editing pipelines.

Below we address all raised concerns through additional analyses and matched reconstruction experiments.

Q1. How does the measured reduction in cycle residual relate to final reconstruction quality, and which timesteps contribute most to the improvement?

We thank the reviewer for pointing out the distinction between optimizing a local cycle residual under frozen predictions and improving the complete reconstruction trajectory.

The closed-form update in Eq. (19) minimizes the local DDIM cycle residual while keeping the adjacent conditional and unconditional predictions fixed. We acknowledge that this local objective is not equivalent to directly minimizing the final reconstruction error, since changing the guidance coefficient also affects subsequent latent states and predictions.

To examine this relationship beyond the frozen-prediction approximation, we conducted additional paired reconstruction analyses on a 100-image Pick-a-Pic subset, comparing POLARIS against fixed guidance with ω=1\omega=1. The endpoint and per-image analyses use the same evaluation subset and 50 inference steps (seed 42).

End-to-end reconstruction quality and per-image correlation

Across the 100 images, POLARIS reduces the end-to-end latent MSE by 22.099% ± 5.81%, the 8-bit pixel-space MSE by 12.409% ± 2.70%, and LPIPS by 11.884% ± 2.89%. The measured local cycle-residual reduction is 7.087% ± 0.43%, while the trajectory cycle-MSE reduction is 25.247% ± 5.83%.

To quantify how the local objective relates to endpoint quality across individual images, we compute Pearson and Spearman correlations between the per-image reduction in local cycle residual JJ and the improvement in each reconstruction metric.

Metric (relative to fixed guidance)Change (mean ± 95% CI)Pearson rr vs. local JJ reductionSpearman ρ\rho vs. local JJ reduction
8-bit pixel MSE reduction12.409% ± 2.70%0.1960.257**
LPIPS relative improvement11.884% ± 2.89%0.236*0.341***
Latent MSE reduction22.099% ± 5.81%0.1120.202*
Local cycle-residual JJ reduction7.087% ± 0.43%1.000 (self)1.000 (self)
Trajectory cycle-MSE reduction25.247% ± 5.83%0.0870.131

Significance: * p<0.05p<0.05, ** p<0.01p<0.01, *** p<0.001p<0.001.

The correlations are positive but moderate rather than close to one. In particular, the local JJ reduction is significantly associated with LPIPS and latent-MSE improvements under the rank-based measure, but it does not fully predict the final reconstruction gain. This is consistent with the fact that the local objective is evaluated under frozen predictions, whereas the full rollout changes subsequent latents and predictions.

Timestep-wise reconstruction analysis

We also measured the predicted clean-sample reconstruction error (pred_x0 MSE) at each of the 50 reconstruction timesteps.The following table aggregates the per-timestep measurements into five consecutive ten-step intervals.

Reconstruction stepsTimestep rangeFixed ω=1\omega=1POLARISRelative reduction
0–9981–8010.416750.43976−5.52%
10–19781–6010.244230.24445−0.09%
20–29581–4010.129820.12389+4.57%
30–39381–2010.065900.05883+10.73%
40–49181–10.035740.02842+20.48%

The timestep-wise results reveal a non-uniform improvement pattern. During the early high-noise stages, POLARIS exhibits slightly higher predicted-sample MSE than fixed guidance. The two approaches become nearly equivalent in the intermediate stages. From the later reconstruction stages onward, POLARIS consistently achieves lower predicted-sample MSE, with the improvement increasing to 20.48% over the final ten steps. At the final timestep, the predicted-sample MSE decreases from 0.0272 to 0.0199, corresponding to a 26.71% reduction.

Timestep-window contribution analysis

We further performed an off_wq timestep-window ablation. Positive values indicate that the corresponding window contributes positively to the final improvement.

WindowTimestep rangeEquivalent pixel MSE reduction (%)ppLPIPS relative improvement (%)Latent relative improvement (%)Within-window local JJ reduction (%)
w0w_01–2010.47 ± 0.550.0850.65 ± 0.541.32 ± 1.137.4
w1w_1201–4010.14 ± 0.730.5460.74 ± 0.471.03 ± 1.3037.6
w2w_2401–5811.96 ± 1.580.0071.97 ± 1.932.86 ± 5.0168.8
w3w_3581–7817.49 ± 2.54<0.001<0.0016.79 ± 2.6113.17 ± 6.3664.4
w4w_4781–9811.35 ± 2.380.0730.85 ± 2.041.01 ± 4.3562.7

The largest contribution to the final pixel-MSE improvement comes from w3w_3 (approximately t=581t=581–781781), with a 7.49% reduction (p<0.001p<0.001). In contrast, w2w_2 has the largest within-window local JJ reduction (68.8%), but does not contribute the most to the final pixel-MSE improvement. This difference further illustrates why local residual reduction and endpoint reconstruction quality should not be treated as interchangeable quantities.

The per-timestep pred_x0 diagnostic and the off_wq window ablation measure different aspects of the trajectory: the former tracks intermediate prediction error, while the latter estimates each window's contribution to endpoint metrics. Together with the per-image correlations and end-to-end measurements, these analyses provide a more complete account of how the local cycle-consistency objective relates to actual reconstruction quality. We do not interpret the window ablation as a strict causal decomposition of the full reconstruction gain.

Q2. Why do Table 1 and the replay-strategy ablation (now Table 18) report different 50-step reconstruction results?

The replay-strategy ablation was Table 12 in the earlier draft and is numbered Table 18 in the current revised manuscript.

The two tables use different evaluation subsets of COCO2017:

  • Table 1 (Main Experiment) uses COCO2017 Sub-1000, containing 1,000 images.
  • Table 18 (Replay Ablation) uses COCO2017 Sub-100, containing 100 images.

Consequently, the MSE values of 349 in Table 1 and 498 in Table 18 are calculated over different evaluation subsets and should not be interpreted as results obtained from an identical image set.

Table 18 specifically investigates the influence of different coefficient replay strategies. Under this evaluation setting, the results are:

Replay strategyMSE ↓
Reverse-order replay498
Forward-order replay642
Cosine replay833
Fixed guidance during reconstruction1171

Reverse-order replay achieves the lowest reconstruction error among the evaluated strategies.

These results demonstrate the importance of temporally aligning the stored coordinate-wise guidance coefficients with their corresponding reconstruction timesteps.

We have explicitly stated the evaluation subset and sample count in the captions of both tables to prevent confusion between the main evaluation and the replay ablation.

Q3. What numerical threshold is used, how frequently is the fallback triggered, and how do extreme guidance coefficients affect reconstruction stability?

We appreciate the reviewer for highlighting the numerical robustness of the closed-form update in Eq. (19).

We report the numerical safeguards and additional statistics collected from 100 images over 49 coefficient-update transitions, with latent dimensions of 4 × 64 × 64.

Numerical threshold and coefficient constraints

The implementation uses the following configuration:

ComponentConfiguration
Numerical threshold ε_num10⁻¹²
DirectionΔε_c − Δε_u
Direction precisionFP32
Fallback condition|d_i| ≤ ε_num
Fallback actionRetain the previous coordinate-wise scale
Coefficient clippingClip the raw coefficient to [0, 2]

The denominator guard is applied independently to each latent coordinate. When the direction magnitude falls below the threshold, the corresponding coefficient retains its previous value rather than reverting to the default CFG coefficient of 1.

After computing the raw analytical coefficient, we apply the box constraint:

w_applied = clip(w_raw, 0, 2).

This prevents arbitrarily large or negative guidance coefficients from being used in the actual reconstruction update.

Fallback frequency

We evaluate the denominator guard over 4,900 image-transition pairs, each containing 4 × 64 × 64 latent coordinates.

StatisticFallback frequency
Mean across image-transition pairs8.82%
Median9.03%
95th percentile17.97%
Maximum for an individual image-transition pair39.29%
Mean across images at individual timesteps0.83%–14.35%
Mean across timesteps for individual images5.22%–22.19%

The corresponding fixed-mode diagnostic yields a fallback frequency of 8.85%, which is nearly identical to that of POLARIS. This indicates that the fallback is primarily triggered by coordinates where the conditional and unconditional prediction differences become numerically identical or sufficiently close. The 8.82% figure measures coordinate-wise denominator degeneracy, for which the previous coefficient provides a conservative fallback; it should not be interpreted as a solver failure rate.

Fallback and coefficient-statistic grouping

We additionally stratified the samples by fallback ratio, boundary occupancy, and raw-coefficient magnitude to examine how these numerical statistics relate to reconstruction quality. For the fallback-ratio grouping, the fallback group uses ε = 10⁻³, while the other groups use the default setting.

Grouping statisticMid group vs. low groupHigh group vs. low groupCorrelation / significance
Fallback ratioMSE −5.2%; LPIPS +8.7%MSE −43.5%; LPIPS −7.0%Spearman 0.310 (p = 0.002); Kruskal–Wallis p = 0.023; potentially confounded by image difficulty
Boundary occupancyMSE +47.5%; LPIPS +28.2%MSE +32.1%; LPIPS +24.9%Spearman −0.197 (p = 0.049); Kruskal–Wallis p = 0.050; not strictly monotonic
Raw |w| > 10 ratioMSE +1.9%; LPIPS +15.9%MSE +2.5%; LPIPS +6.6%Spearman −0.097 (p = 0.337); Kruskal–Wallis p = 0.751
Raw |w| p99.9MSE −3.2%; LPIPS −2.5%MSE −8.8%; LPIPS +5.3%Spearman −0.033 (p = 0.747); Kruskal–Wallis p = 0.917

These grouped results do not establish a simple monotonic relationship between extreme coefficient statistics and reconstruction degradation. In particular, the fallback-ratio association may be affected by image difficulty, while the raw-coefficient magnitude statistics do not show statistically significant associations in this analysis.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.