acceptodds
Under review as a conference paper at ICLR 2027

TrajCFG: Token-Wise Classifier-Free Guidance via Denoising Trajectory Feedback

Abstract

Classifier-free guidance (CFG) typically applies a shared scale to all latent tokens, although their velocity predictions evolve differently during sampling. We introduce Trajectory-Adaptive Classifier-Free Guidance (TrajCFG), a training-free method for flow-matching models that adapts token-wise guidance using historical changes in guided velocity direction. TrajCFG maps this feedback to bounded scales, assigning stronger guidance to larger directional changes and weaker guidance to more consistent directions. It combines adjacent-step and interval-normalized longer-span feedback, reusing standard CFG predictions without additional model evaluations. Controlled diagnostics and ablations examine the roles of directional feedback, token-wise allocation, and history-span fusion. Experiments on text-to-video, image-to-video, text-to-audio-video, text-to-image, and image-to-3D generation show improvements over base samplers across multiple backbones. For text-to-video generation, TrajCFG improves VBench Total scores by 0.30–0.75 points over the strongest evaluated baseline on each of three backbones. These results support historical directional feedback as a broadly applicable signal for fine-grained guidance allocation.

Primary area:
generative models
Keywords:
Classifier-Free Guidance, Flow Matching, Diffusion Models, Sampling Dynamics, Training-Free Inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xTwHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Tradition CFG applies same guidance scale to every token at every timestep, though different tokens' velocity predictions evolve differently along the sampling trajectory. TrajCFG is a training-free method, which measures the same token's angle of velocity between adjacent timesteps, and maps that angle to a bounded per-token guidance scale, giving weaker guidance where the predicted direction has been stable and stronger guidance where it is still being revised. Main experiments were conducted on Wan series models, and suprpassed most of compared guidance methods on Vbench and Dynamic-Bench with almost no additional computational cost. The authors also conducted experiments on image-to-video, text-to-audio-v text-to-image and image-to-3D, mostly against the base sampler only. Ablations show token-wise allocation beats time-only and sample-wise variants, and that shuffling or reversingthe per-token assignment degrades results, indicating the angle signal carries genuine spatial information rather than merely raising the average scale.

Critical strengths
  1. Coverage of models, domains. Same mechanism available for many Wan series models, Cosmos-Predict2.5-2B, LTX-2.3, Ovi, Qwen-Image, SD3 Medium and Hunyuan3D 2.0. Including T2V, I2V, T2AV, and Image-to-3D.
  2. Almost no additional computation cost because CFG already need to compute the velocity.
  3. Ablation studies shows the angle map carries spatial information rather than acting as a global rescale.
Critical weaknesses

W1. The reported gains are not shown to come from token-wise allocation. Two other explanations remain open: TrajCFG simply runs at a stronger average guidance scale, and the baselines it beats do not adapt the scale per location.

  • Higher guidance scale. A token's scale can reach s_max=10.0 against s_base=6.0 (Tab. 7), and gains concentrate on VBench Semantic (72.77 to 75.37, Wan2.1-1.3B) over Quality (83.03 to 84.06, Tab. 1). "Guidance Matters" (arXiv 2602.22570) finds that raising the CFG scale alone competes with most recent guidance methods, since alignment metrics reward large scales. No constant-scale control reported, at the realized mean or at s_max.
  • Wrong opponents. The four baselines (CFG-Zero*, CFG-Ctrl, PAG, SRVS; p6) do not allocate per-location scales. SAMG (point-wise) and VAGS (cosine-driven) are described in the paper's own related work (p2), HiGS (arXiv 2509.22300), APG and interval guidance are cited, and none is run. FBG (arXiv 2506.06085) is not cited. Baseline tuning is undocumented while TrajCFG tunes four hyperparameters per backbone.

W2. The rule that sets each token's guidance was never checked on the states the sampler actually visits.

  • Fig. 3 needs intermediate states at step k and the two before it. Instead of running the sampler, the authors noise real video latents, "without sampler rollouts" (p4), then sweep CFG on those hand-built states. That one measurement is the whole basis for "small angle, lower guidance; large angle, higher guidance".
  • Real sampling compounds its own errors from step to step and can end up far from any real video; noised real video stays near a correct answer throughout. So the experiment tests a regime the method never meets. The authors call the transfer a "hypothesis" (p5) and never check it.
Questions for authors
  1. What is the guidance scale in practice, averaged over steps and tokens? And if the authors run the CFG at guidance scale of s_max, how much of the gap to TrajCFG is left? Since simply larger guidance scale of CFG might improve these metrics. 2. Were the baselines tuned per backbone the way TrajCFG was, or run at their published defaults?
  2. The authors may add SAMG / VAGS / HiGS / APG if applicable in T2V / T2I for comparison. Since they are more similar to TrajCFG.
  3. Does the Fig. 3 effect survive if the states come from actual sampler rollouts instead of noised ground truth? A wider CFG sweep would also help, since some groups look like they bottom out at the edge of the current range.
4 comments
AuthorsOpenReview ↗

Response to Reviewer xTwH

We thank the reviewer for the specific controls and comparisons requested. New figures are on the last page (p. 24) of the revised appendix.

W1 / Q1. Realized guidance scale and constant-CFG controls.

Averaged over samples, tokens, and all 40 steps on VBench, TrajCFG's applied scale is 5.65 on Wan2.1-1.3B and 3.92 on Wan2.2-5B. Both are below the default scales of 6 and 5.

Figure 12 shows the per-step mean. It is close to 10 when adaptation starts, falls below the default by about step 10, and stays near smin⁡s_{\min} for most of the remaining steps. Only a few early steps approach smax⁡s_{\max}, so TrajCFG mostly lowers guidance rather than raising it. Within each step, sites with large historical angles still receive more guidance than settled sites. The allocation changes the ranking across tokens, not the mean level.

Method1.3B Quality1.3B Semantic1.3B Total5B Quality5B Semantic5B Total
Base CFG (6 / 5)83.0372.7780.9882.8874.5281.21
Constant CFG at realized mean (5.65 / 3.92)83.2672.3881.0983.0874.2281.31
Constant CFG at smax⁡s_{\max} = 1081.9172.0079.9282.6772.2880.59
Step-wise maximum CFG83.0573.2681.0982.8975.3781.38
Step-wise mean CFG83.8273.6881.8084.4576.2982.82
TrajCFG84.0675.3782.3284.8277.4583.35

The step-wise rows apply, at each step, a single scale to all tokens: the token maximum or the token mean of TrajCFG's map at that step.

  • Constant CFG at smax⁡s_{\max}. CFG 10 lowers both Quality and Semantic relative to the default on both backbones. A larger constant scale therefore does not explain the Semantic gains here.
  • Constant CFG at the realized mean. This control gains only 0.10–0.11 Total over Base.
  • Step-wise mean schedule. This schedule recovers part of the gain (+0.82 and +1.61 over Base). TrajCFG is still 0.52 and 0.53 Total above it.

We attribute this remaining gap to token-wise allocation, in agreement with the Time-only and Sample-wise ablations in Table 3. TrajCFG's improvement thus has two parts: the time profile that its feedback produces, and the allocation across tokens. We will report both explicitly.

W1 / Q2. Baseline tuning.

All baselines, including those added below, used the method-specific defaults from their papers at the backbone's default CFG. We did not run a per-backbone search for them.

For TrajCFG, Table 7 lists the per-backbone settings. The history spans and fusion weights (0.8/0.2) are fixed across all backbones, and ρlo\rho_{\text{lo}} takes only two values (0.90 or 0.95), so the main per-backbone choice is [smin⁡,smax⁡][s_{\min}, s_{\max}]. We acknowledge this asymmetry. The controls above show that changing vanilla CFG's scale, or its per-step schedule, does not reproduce the gain.

AuthorsOpenReview ↗

Response to Reviewer xTwH

W1 / Q3. Comparison with SAMG, VAGS, HiGS, and APG.

T2V, VBench1.3B Quality1.3B Semantic1.3B Total5B Quality5B Semantic5B Total
Base CFG83.0372.7780.9882.8874.5281.21
SAMG83.4072.7181.2684.1174.1582.11
VAGS-Gen83.0173.9881.2182.9975.4381.48
HiGS83.6872.6881.4884.0175.2382.25
APG82.9773.0380.9883.4475.4681.84
TrajCFG84.0675.3782.3284.8277.4583.35
T2I, SD3-Medium, GenEvalSingleTwoCount.ColorsPos.Attr.Overall
Base CFG0.99690.86110.61250.83510.31250.56500.7021
SAMG0.99690.86870.65000.84570.32000.59250.7123
VAGS-Gen0.99690.86110.64060.85110.31000.58250.7070
HiGS1.00000.86110.65000.83510.32250.58500.7090
APG1.00000.87120.61880.83240.32250.57500.7033
TrajCFG1.00000.88130.63440.84840.32250.59750.7140

On T2V, TrajCFG leads the strongest added baseline, HiGS, by 0.84 and 1.10 VBench Total. On T2I, it has the best GenEval Overall (+0.0017 over SAMG), although SAMG and HiGS are better on Counting and VAGS-Gen is better on Colors.

Limited-interval guidance is a time-only schedule. We did not run it, but Time-only in Table 3 and the step-wise schedules above are time-only controls derived from TrajCFG's own profile, and token-wise allocation improves on each of them by at least 0.27 Total. The two could be combined, with an interval schedule setting the time profile and TrajCFG allocating within it, but we have not tested this.

FBG sets a state-dependent guidance coefficient during sampling, whereas TrajCFG allocates scales per token from cross-step feedback. We thank the reviewer for the pointer and will cite and discuss FBG in the related work.

W2 / Q4. Fig. 3 on actual sampler states.

We repeated the Fig. 3 protocol on states visited by the sampler (Figure 11). The setup is as follows:

  • States. Wan2.2-5B starts from random noise with DAVIS text prompts and runs 10 steps with base CFG.
  • Groups. At step 10, tokens are grouped by the angle between their guided velocities at the two preceding steps, using the Fig. 3 bins.
  • Intervention. From the same latent, we change the scale of one group while the other groups stay at the default, then finish sampling.
  • Metric. These states have no ground truth, so we score the final videos by frame-level CLIP text alignment, averaged over prompts.

Figure 11 shows the same qualitative pattern as Fig. 3. Small-angle tokens prefer guidance below the default. Large-angle tokens benefit from guidance above the default and lose alignment when it is lowered. The transition between these two regimes falls in the 3–6° group, which is largely insensitive in Figure 11. Because only one group is changed at a time, the effects are small. We therefore read both figures qualitatively, as showing that the preferred scale increases with the angle, and not as a calibrated mapping from angle to scale.

The closest test on TrajCFG's own closed-loop trajectories is the Shuffled and Reversed ablations in Table 3: replaying its scale maps with shuffled or reversed token assignments, from the same initial noise, lowers Total by 0.46–0.77.

Thank you for the detailed response. My concerns have been well addressed, and I will raise my score.

Dear Reviewer xTwH,

We are delighted that our responses have addressed your concerns and that you now support the acceptance of our paper.

Thank you for your thoughtful and constructive feedback, which has helped us clarify the paper’s contributions and strengthen its presentation.

Best regards, The Authors

Reviewer_SKjROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TrajCFG, a training-free method for adapting classifier-free guidance at individual latent positions. It uses historical changes in guided velocity direction to assign guidance scales and combines feedback from different history spans. The method reuses existing predictions without additional model evaluations. The experiments cover video, audiovisual, image, and 3D generation.

Critical strengths
  1. The use of historical direction changes as a local feedback signal is reasonable.

  2. The method requires neither retraining nor additional model evaluations, making it practical to apply to existing generators. The multi-backbone results and the allocation ablations in Table 3 provide useful evidence that matching guidance to local feedback improves generation.

Critical weaknesses
  1. Table 3 supports token-wise allocation, but does not isolate the advantage of historical direction changes over other feedback signals under the same framework. I am curious whether alternatives such as velocity-magnitude changes or guidance-difference energy would provide similar improvements.

  2. SAMG and VAGS are discussed in the related work but are not compared experimentally.

  3. Figure 3 uses states constructed from ground-truth latents rather than actual sampling rollouts. Since historical turns during generation also depend on previously applied guidance, it remains unclear how consistently the observed angle–guidance relationship holds along the method’s own sampling trajectories.

4.The gains over the strongest evaluated baselines are relatively modest.

Questions for authors
  1. How does TrajCFG compare with SAMG and VAGS under comparable tuning budgets?

  2. Is there evidence that the relationship between historical angles and beneficial guidance scales also holds along actual sampling trajectories?

4 comments
AuthorsOpenReview ↗

Response to Reviewer SKjR

We thank the reviewer for the positive assessment and the concrete suggestions. New figures are on the last page (p. 24) of the revised appendix.

W1. Other feedback signals within the same framework.

We replaced only the feedback statistic. Each alternative signal is normalized per step to [0, 1] and mapped with the same bounded rule, using the same scale range, adaptive steps, and fusion weights as TrajCFG.

Feedback signal1.3B Quality1.3B Semantic1.3B Total5B Quality5B Semantic5B Total
Velocity-magnitude change83.5373.5781.5484.2375.3982.47
Guidance-difference energy83.4772.9581.3784.1975.0082.35
Direction change (TrajCFG)84.0675.3782.3284.8277.4583.35

1.3B = Wan2.1-1.3B and 5B = Wan2.2-5B, both on VBench. Direction change gives the best score in all six columns. Its Total margin is 0.78–0.88 over magnitude change and 0.95–1.00 over guidance-difference energy.

Magnitude change captures the norm term of Eq. (5), which also varies with the noise level, and guidance-difference energy uses only the current step. The directional term is invariant to rescaling and measures cross-step revision, which is the property the allocation targets.

W2 / Q1. SAMG and VAGS.

Method1.3B Quality1.3B Semantic1.3B Total5B Quality5B Semantic5B Total
Base CFG83.0372.7780.9882.8874.5281.21
SAMG83.4072.7181.2684.1174.1582.11
VAGS-Gen83.0173.9881.2182.9975.4381.48
TrajCFG84.0675.3782.3284.8277.4583.35

TrajCFG leads SAMG, the stronger of the two, by 1.06 and 1.24 Total.

On tuning budgets: SAMG and VAGS used their published method-specific defaults at the backbone's default CFG (6 and 5). TrajCFG used the ranges [5, 10] and [3, 10] with ρlo=0.95\rho_{\text{lo}} = 0.95, and its history spans and fusion weights are shared across all backbones. We could not run a matched hyperparameter search for the baselines during the rebuttal period, so we do not claim that this comparison is budget-matched.

We also compared the two methods on text-to-image generation with SD3-Medium. On GenEval Overall, TrajCFG scores 0.7140, compared with 0.7123 for SAMG and 0.7070 for VAGS-Gen. The margin over SAMG on this task is small.

W3 / Q2. Does the angle–guidance relationship hold on actual sampling trajectories?

We repeated the Fig. 3 protocol on states visited by the sampler (Figure 11). The setup is as follows:

  • States. Wan2.2-5B starts from random noise with DAVIS text prompts and runs 10 steps with base CFG.
  • Groups. At step 10, tokens are grouped by the angle between their guided velocities at the two preceding steps, using the same five bins as Fig. 3.
  • Intervention. From the same latent, we change the scale of one group while the other groups stay at the default, then finish sampling.
  • Metric. Real rollouts have no ground-truth latent, so instead of clean-latent MSE we score the final video by frame-level CLIP text alignment, averaged over prompts.

Under this different metric, Figure 11 shows the same qualitative pattern as Fig. 3. Small-angle tokens prefer guidance below the default. Large-angle tokens benefit from guidance above the default and lose alignment when it is lowered. The transition between these two regimes falls in the 3–6° group, which is largely insensitive in Figure 11. Because only one group is changed at a time, the effects are small. We therefore read both figures qualitatively, as showing that the preferred scale increases with the angle, and not as a calibrated mapping from angle to scale.

We did not repeat the per-group sweep on TrajCFG's own trajectories. The closest evidence is the Shuffled and Reversed ablations in Table 3. Replaying TrajCFG's own scale maps from the same initial noise, but with token assignments shuffled or reversed, lowers VBench Total by 0.46–0.77. The allocation computed along the method's own trajectory therefore matters beyond the distribution of scales it produces.

W4. Magnitude of the gains.

  • Gains over Base. VBench Total improves by 1.34–2.14 points across the three T2V backbones.
  • Margin over the strongest baseline. The margin is 0.30–0.75 Total, at the base sampler's cost. SRVS, the strongest baseline on two of the three backbones, takes 1.5× the runtime.
  • Consistency. Across all per-dimension results in Tables 5, 6, 8, and 10–14 (188 metrics over five tasks), TrajCFG is better than Base on 182, ties on 2, and is worse on 4.

Dear Reviewer SKjR,

We sincerely thank you for your constructive feedback, which has helped us strengthen the paper and clarify its contributions.

We hope that our detailed responses and additional results have adequately addressed your concerns. Should any questions or concerns remain, we would be more than happy to provide further clarification.

If you find that our responses have satisfactorily resolved your concerns, we would greatly appreciate your reconsideration of the score and stronger support for the paper.

Thank you again for your time and thoughtful evaluation.

Best regards, The Authors

Thank you for the detailed responses. I maintain my positive rating and remain inclined toward acceptance.

Dear Reviewer SKjR,

We sincerely thank you for your continued support and thoughtful evaluation of our work. We sincerely appreciate your time and constructive feedback, which have helped us improve the paper.

Best regards, The Authors

Reviewer_pJMPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose a per-token adaptive scheduling for CFG, based on angular differences in the velocity field between successive flow-matching steps. A first observation shows that angular difference in the velocity fields impacts how strong or weak guidance is correlated with reduced MSE, with high angular shifts benefiting from strong guidance while low angular shifts require weaker guidance. The paper then details a formula to compute a guidance strength based on this angular difference and shows it provides some improvements to typical generation benchmarks.

Critical strengths

The observation made at the beginning of section 4 is interesting. It seems that some regions have more curvature in their denoising trajectory than others, and that it seems to correlate on the main subject rather than on the background. It also seems that a high curvature would benefit from a high CFG whereas a low curvature would benefit from a weak CFG. That second observation is surprising and counter-intuitive to me, since increasing the CFG on high curvature trajectories would mean possibly "missing a turn", whereas on a straight line it would be a simple acceleration (e.g., fewer steps).

Critical weaknesses

The observation on the link between curvature of the trajectory and CFG strength is not properly investigated and taken into account. For example, the best CFG for each angular value should have been plotted and modeled (with a linear model like the proposed formula suggests). It also comes with no explanation or intuition as to why guidance strength correlates with angle. The paragraph named "interpreting directional changes" does not provide an interpretation but only a list of trivial algebraic manipulations. A proper interpretation should explain how directional changes are connected to specific parts of the image and how a need for a different CFG arises from this connection.

The experiments are underwhelming. The quantity is definitely there, with many benchmarks and models and datasets. But most improvements are around 0.1% to 0.7%, well within the noise level, especially since the method introduces more parameters that are tuned on these benchmarks.

Questions for authors

Fig.3 why grouping? If the pattern really holds, you could just color code each curve for each patch and plot them all. Better even, plot best CFG against angle. How do the 2 correlate? If you fit a linear model on that curve, what is its parameters and error? How does that relate to the parameters of the model you propose?

Fig3 again, the error is show in absolute value and has range 1e-3, but what is the absolute MSE at the reference of CFG=5? If it's 0.2, then the gain is in the 0.1% range, so really nothing.

Ablation Q2: I am not sure to understand. If you do the opposite of what your observation suggest (and what your method does), you get 0.4% less on average? Which roughly matches any other method from the state of the art? To me this totally invalidates the method. Doing the opposite of something smart should not result in scores comparable or above the baseline, or else, it means that the method is also super limited because its hypotheses are not crucial.

4 comments
AuthorsOpenReview ↗

Response to Reviewer pJMP

We thank the reviewer for the careful reading of Fig. 3 and the ablations. New figures are on the last page (p. 24) of the revised appendix.

Q1. Why group tokens in Fig. 3, and how does the best CFG relate to the angle?

A single token's response curve is noisy, and its argmin can only take one of the seven tested values. Per-token optima are therefore dominated by noise and grid quantization. Grouping averages many tokens with similar angles. Membership is fixed across the sweep, so each curve compares the same tokens at every scale.

Qualitatively, the preferred scale in Fig. 3 rises with the angle. The small-angle groups do best below the default scale, and the large-angle groups do best above it. With five groups, seven tested scales, and several optima at the ends of the tested range, the figure shows this monotone trend but is too coarse to determine a functional form. We therefore use Fig. 3 for the ordering, not for calibration.

Q2. Absolute MSE at CFG = 5.

The clean-latent MSE at the CFG = 5 reference is about 0.07, so the changes in Fig. 3 are a few percent of the reference. However, Fig. 3 is meant to show the direction of the effect, not its size. The same change in scale has opposite effects on different groups: lowering the scale reduces the error for the smallest-angle group but increases it for the largest-angle group. A scale shared by all tokens must therefore trade one group off against the other, whereas a per-token scale can improve both. This conclusion does not depend on how large the per-group effect is.

Q3. Reversed allocation in Table 3.

Reversed scores above Base because it keeps most of what TrajCFG changes. It replays TrajCFG's own scale maps with the angle ranking inverted, so every step uses exactly the same set of scales, only assigned to different tokens. It therefore inherits TrajCFG's time profile (Fig. 12): strong guidance when adaptation starts, then guidance near smin⁡s_{\min} for most of sampling. This profile is not hand-designed. It is the same angle feedback, aggregated over tokens at each step.

To separate the two parts, we added a step-wise mean control. Like Reversed, it replays TrajCFG's cached scale maps from the same initial noise, but at each step it applies the token mean of the map to all tokens. It keeps the time profile but does not allocate across tokens. Unlike Sample-wise in Table 3, which computes its scale maps online along its own trajectory before averaging, step-wise mean is derived from the same cached per-step scales that Reversed redistributes. It is therefore the control most directly comparable to Reversed.

VBench TotalBaseConstant at realized meanStep-wise meanTime-onlySample-wiseReversedFull
Wan2.1-1.3B80.9881.0981.8081.7881.8781.5582.32
Wan2.2-5B81.2181.3182.8283.0882.8582.8183.35
  • Time profile. Step-wise mean improves over Base by 0.82 and 1.61 Total, whereas a constant scale at the same overall mean improves by only 0.10–0.11. Most of the gain over Base therefore comes from the time profile, which Reversed keeps.
  • Token allocation. Full adds 0.52–0.53 on top of step-wise mean and 0.27–0.54 on top of Time-only and Sample-wise.
  • Inverted allocation. Reversed loses this gain and has the lowest Total of all ablation variants in Table 3 on both backbones. It is no better than the allocation-free controls: on Wan2.1-1.3B it is below all of them, and on Wan2.2-5B it is level with step-wise mean (82.81 vs. 82.82).

Reversed scores above Base and several baselines because of the time profile, not because the token-level signal is uninformative.

AuthorsOpenReview ↗

Response to Reviewer pJMP

W1. Why does guidance strength correlate with the angle?

The intuition of "missing a turn" would apply if the scale set the step size. In CFG, the step size is fixed by the sampler, and the scale instead changes the direction the step follows, toward content more specific to the condition. Our motivation is direct. Where the content of a region is still unsettled, stronger text guidance helps make what is generated there follow the condition, so we raise the CFG scale in such regions. The question is how to identify them during sampling. We found the angle between consecutive velocity estimates to be a good observable. By Eq. (4), a large angle means that the clean-endpoint estimate at that site was revised substantially in the last step, so its content is still being decided. The angle maps in Fig. 2 and the guidance maps in Figs. 4–6 show that such sites concentrate on foreground subjects and moving parts, where the prompt determines what appears.

A small angle means the endpoint estimate has settled, typically in the background. There, extra guidance mainly amplifies the conditional–unconditional difference after the content is fixed, which is the regime associated with over-saturation at high CFG (Sadat et al., 2024). Weaker guidance there costs little.

This reading also matches where the gains appear: VBench Semantic improves most (+2.60 to +4.34 over Base), while Quality also improves. We offer it as an interpretation consistent with Figs. 2–3 and Table 3, not as a derivation, and will replace the algebraic paragraph in Sec. 4.1 with it.

W2. Size of the gains.

  • Gains over Base. VBench Total improves by 1.34, 2.14, and 2.05 points on the three T2V backbones. The margin over the strongest baseline on each backbone is 0.30–0.75. SRVS, the strongest baseline on two of the three backbones, takes 1.5× the runtime, whereas TrajCFG runs at the base sampler's cost.
  • Consistency. Across all per-dimension results in Tables 5, 6, 8, and 10–14 (188 metrics over five tasks), TrajCFG is better than Base on 182, ties on 2, and is worse on 4. The metrics are correlated, so this is not a formal test, but noise is unlikely to produce this consistency.
  • Paired controls. Shuffled and Reversed start from the same initial noise and use the same per-step scale distribution as TrajCFG; they differ only in which tokens receive which scales. They lose 0.46–0.77 Total. Because the initial noise is shared, this gap does not come from differences in initial noise.
  • Tuning. The history spans and fusion weights (0.8/0.2) are fixed across all backbones, and ρlo\rho_{\text{lo}} takes only two values (0.90 or 0.95). The main per-backbone choice is the scale range, which follows a simple pattern (Table 7): smin⁡s_{\min} is at most 2 below the default CFG in 9 of the 12 settings, and smax⁡s_{\max} is 9 or 10 in 10 of the 12. These values were set once per backbone and task with only 1–3 trial runs, then kept fixed across all benchmarks of that task (e.g., VBench and Dynamic-Bench for T2V). They were not tuned per benchmark.

Thanks for the answers. I'll discuss this with the other reviewers and the AC and then modify my score accordingly.

Dear Reviewer pJMP,

We sincerely thank you for your time and thoughtful consideration. We appreciate your discussion with the other reviewers and the AC, and would be happy to provide any further clarification if needed.

Best regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We thank all reviewers for their constructive feedback. We are encouraged that the reviewers found the trajectory observation interesting (pJMP), the historical-direction feedback a reasonable local signal (SKjR), and the method practical: training-free, with no additional model evaluations, and broadly applicable across backbones and modalities (SKjR, xTwH). New figures (Figures 10–12) are on p. 24 of the revised appendix. We summarize the main points below; detailed responses are in each reviewer's thread.

1. Source of the gains (pJMP Q3; xTwH W1/Q1). TrajCFG does not rely on stronger guidance. Its realized mean scale (5.65 / 3.92 on Wan2.1-1.3B / Wan2.2-5B) is below the default (6 / 5). Constant CFG at this mean yields only +0.11 / +0.10 VBench Total, and constant CFG at smax⁡=10s_{\max}=10 degrades performance. Beyond the time profile induced by the feedback (Figure 12), token-wise allocation adds a further +0.52 / +0.53 Total. Shuffled and Reversed allocations, which share the same initial noise and scale distribution, lose 0.46–0.77 Total. This confirms that the per-token assignment matters.

2. Angle–guidance relationship on actual trajectories (pJMP Q1; SKjR W3/Q2; xTwH W2/Q4). Figure 10 shows that the best CFG scale increases with the historical angle. Figure 11 repeats the Fig. 3 protocol on actual sampler states, scored by CLIP text alignment, and shows the same pattern: small-angle tokens prefer weaker guidance, while large-angle tokens benefit from stronger guidance.

3. Additional baselines (SKjR W2/Q1; xTwH W1/Q3). We added SAMG, VAGS-Gen, HiGS, and APG. On T2V, TrajCFG outperforms the strongest of them (HiGS) by 0.84 / 1.10 VBench Total. On T2I (SD3-Medium), TrajCFG achieves the best GenEval Overall score. We will also cite and discuss FBG.

4. Choice of feedback signal (SKjR W1). Within the same framework, directional change outperforms velocity-magnitude change and guidance-difference energy by 0.78–1.00 Total. It is best on Quality, Semantic, and Total on both backbones.

5. Intuition (pJMP W1). In CFG, the step size is fixed by the sampler, and the scale steers the direction toward condition-specific content. A large angle indicates that the clean-endpoint estimate is still being revised, typically on foreground subjects and moving parts. Stronger guidance there helps the content follow the prompt. A small angle indicates settled content, where weaker guidance mitigates over-saturation. Consistent with this, VBench Semantic improves by +2.60 to +4.34 over Base.

6. Magnitude and consistency of the gains (pJMP W2; SKjR W4; xTwH Q2). TrajCFG improves VBench Total by 1.34–2.14 over Base and by 0.30–0.75 over the strongest baseline, at the base sampler's cost (SRVS requires 1.5× the runtime). It outperforms Base on 182 of 188 per-dimension metrics across five tasks. History spans and fusion weights are shared across all backbones, and the scale range is set once per backbone and task rather than per benchmark.

Revisions. We added Figures 10–12 to the appendix. We will incorporate the new controls, baselines, and feedback-signal comparisons, together with the revised interpretation in Sec. 4.1 and the related work, into the final version.

We thank all reviewers again for their valuable feedback, which has helped strengthen the paper. We are happy to answer any further questions.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.