acceptodds
Under review as a conference paper at ICLR 2027

Joint Branch-Space Transform Coding for Diffusion Activation Quantization with Classifier-Free Guidance

Abstract

Post-training quantization for diffusion models increasingly exploits timestep, feature, and layer structure. While recent work has begun incorporating CFG structure into diffusion quantization, activation quantization still operates independently across conditional and unconditional coordinates, leaving cross-activation structure unexploited. We show that matched CFG activations form a strongly correlated two-dimensional source and that, under a fixed bit budget, the choice of branch coding basis materially affects quantization fidelity. Motivated by this observation, we introduce , which rotates matched CFG branches via an offline derived orthogonal matrix, requiring minimal modifications to model parameters or the quantization pipeline. We further derive the , which jointly incorporates the CFG guidance direction and cross-branch second moments. Under an equal-rate quantization-noise surrogate, GCBT admits a closed-form per-layer solution without gradient optimization or angle search. Applied on top of existing diffusion PTQ methods, GCBT yields statistically significant fidelity gains in most evaluated comparisons with no statistically significant degradation, while leaving the underlying host quantization pipeline unchanged.

Primary area:
generative models
Keywords:
diffusion models, classifier-free guidance, quantization

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_27DhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces Guidance-Correlation Branch Transform (GCBT), a lightweight activation transform for post-training quantization of diffusion models with classifier-free guidance (CFG). The key observation is that the conditional and unconditional CFG activations form a strongly correlated two-dimensional source, whereas existing activation quantization methods typically quantize the two branches independently in their native coordinates. The paper formulates the conditional/unconditional pair as a two-dimensional branch space and applies a 2x2 orthogonal transform before activation quantization, followed by the inverse transform afterward. To determine the transform, the authors propose a guidance-weighted distortion surrogate that combines the per-layer second-moment matrix of matched CFG activations with the known CFG readout direction. Under an equal-rate quantization-noise assumption, the optimal rotation admits a closed-form solution, requiring neither gradient-based optimization nor angle search.

Critical strengths
  1. The method is simple, interpretable, and mathematically elegant.
  2. This paper provises good compatibility with existing PTQ methods.
  3. The broad and relatively careful empirical validation.
Critical weaknesses
  1. My main technical concern is the connection between the closed-form surrogate optimum and the actual quantizer optimum is insufficiently validated. The theoretical result assumes that the quantization-error variance of each transformed coordinate is proportional to its projected second moment, uses the same proportionality constant for both coordinates, and neglects cross-coordinate quantization-error terms.
  2. The use of the final CFG direction as a layer-local sensitivity measure deserves stronger justification.
  3. GCBT requires matched conditional/unconditional activations and therefore assumes active two-branch CFG. It does not currently cover distilled guidance, single-pass CFG variants, or dynamic/timestep-varying CFG schedules, as acknowledged by the authors.
  4. The CFG-aware baseline is informative but not a definitive comparison to prior CFG-aware quantization.
Questions for authors
  1. What happens when the conditional and unconditional branches are executed sequentially rather than jointly?
  2. Can the authors better characterize when GCBT is expected to help?
7 comments

We appreciate the reviewer's careful reading and constructive feedback. We address each point below:

W1 — Validation of Eq. (6)’s Proportionality and Cross-Coordinate Error Term Assumptions.

We have clarified the assumptions behind Eq. (6) in the revised manuscript (highlighted in blue) and added empirical validation under the original quantization settings. Below, we summarize our findings and addressing for each assumption.

(1) Proportionality assumption. GCBT uses projected activation second moments to approximate quantization error. We test this assumption by fitting a single coefficient κℓ\kappa_\ell per site, shared across both transformed coordinates and all rotation angles. On held-out samples across 360 angles, the predicted quantization-error second moments closely match the measured values, with a median Pearson rr of 0.999, median R2R^2 of 0.999, and median/P90 normalized RMSE of 1.3%/4.3%. In total, 152/168 sites have an error below 5%. These results show that the proportionality assumption holds well at most sites.

(2) Cross-coordinate error terms. We further examine whether ignoring cross-coordinate error terms affects rotation selection. We find that directly minimizing the diagonal-only error DdiagD_{\mathrm{diag}} and the full error DfullD_{\mathrm{full}} produces nearly identical angles at most sites: the median/P90 difference is 0.05°/2.72°, with only 5/168 sites differing by at least 5°. The resulting difference in held-out DfullD_{\mathrm{full}} is just +0.129% (95% CI: [−0.006%, +0.254%]). Thus, ignoring these terms has little effect on rotation selection or actual quantization error.

Overall, these results support both approximations used by GCBT. Additional experimental results are provided in our response to Reviewer wcpw (W1, Q1–Q2).

W2 — Validation of the Final CFG Direction for Layer-Local Rotation Selection.

We validate using the final CFG direction for layer-local rotation selection by testing (1) whether local quantization error reflects final denoiser-output error and (2) whether CFG weighting improves rotation selection.

(1) Local-to-output correspondence. On SANA-1.6B, we compare CFG-weighted local quantization error with final denoiser-output error across candidate rotations. The two errors show strong agreement, with a median Spearman correlation of 0.863 and 80.85% pairwise ranking agreement. We further compare rotations selected by minimizing local error against those selected by directly minimizing final output error under the same CFG scale.

Table R5. Fraction of the oracle improvement in denoiser-output MSE recovered by local rotation selection on SANA-1.6B.

Selection ObjectiveOracle improvement recovered ↑
CFG-weighted local error90.92% [88.66%, 92.86%]
Output error (oracle)100%

As shown in Table R5, local CFG-weighted selection recovers 90.92% of the improvement achievable by directly optimizing final output error. This shows that local error provides a reliable guide for rotation selection without requiring downstream output evaluations.

(2) Importance of CFG weighting. We test whether CFG weighting improves rotation selection over CFG-agnostic alternatives. On SANA-1.6B, we compare three local objectives: (i) equal weighting of the two branch errors, (ii) separate branch errors, and (iii) CFG-weighted branch errors. All three select rotations at six sites using leave-one-prompt-out evaluation. We then measure final denoiser-output MSE under the same CFG scale of 4.5 (Table R6).

Table R6. Final denoiser-output MSE under different local rotation-selection objectives on SANA-1.6B (CFG = 4.5), normalized to Native.

Selection ObjectiveOutput MSE / Native ↓
Equal branch weighting (s~=0.5\tilde s=0.5)1.007
Separate branch errors0.986
CFG-weighted branch errors (s~=4.5\tilde s=4.5)0.757

CFG weighting reduces final output MSE by 24.8% and 23.2% compared with the two CFG-agnostic objectives, respectively. This advantage persists on a fixed angle grid that excludes all GCBT-calibrated rotations.

Overall, these results support using the final CFG direction for layer-local rotation selection: local CFG-weighted error closely reflects final output error, and CFG weighting yields better rotations than CFG-agnostic alternatives.

W3 — Applicability to CFG Variants.

GCBT targets models with active two-branch CFG, where matched conditional and unconditional activations are required for branch-space rotation. While GCBT cannot be directly applied under an unchanged full-pass sequential execution schedule, a layer-wise interleaved schedule enables exact GCBT computation with near-sequential memory usage. Guidance-distilled and genuinely single-branch models remain outside its current scope. We validate this implementation and compatibility with dynamic CFG below.

(1) Sequential CFG Execution. We implement a layer-wise interleaved schedule that enables GCBT to operate with near-sequential memory usage. We temporarily buffer the unconditional activation until the corresponding conditional activation becomes available, reusing the buffer across layers. On PixArt-Σ W4A4, this implementation produces bitwise-identical outputs to joint GCBT, with only 53.3 MiB (3.7%) additional peak memory compared with Native full-pass sequential execution (Table R7).

Table R7. Peak denoiser memory and latency under sequential and interleaved CFG execution on PixArt-Σ/SVDQuant W4A4. Latency is the median of 20 runs per method on the same L40S GPU.

MethodPeak denoiser memory ↓Latency ↓
Native sequential1453.5 MiB5.726 s/image
Native interleaved1507.8 MiB5.827 s/image
GCBT interleaved1506.8 MiB5.712 s/image

Under the same interleaved schedule, GCBT achieves nearly identical memory usage and latency to Native, demonstrating minimal additional execution overhead.

(2) Compatibility with Dynamic CFG. As demonstrated in Table R2 (Reviewer wcpw, W2, Q4), GCBT supports timestep-varying CFG by updating its closed-form rotation using the current scale sts_t. On CogVideoX, timestep-adaptive GCBT improves LPIPS from 0.445 to 0.409 and SSIM from 0.689 to 0.699 compared with Native.

W4 & Q2 — Distinction from Prior CFG-Aware Methods and When GCBT Helps.

(1) Difference from prior CFG-aware methods (W4). To the best of our knowledge, GAMP is the most directly related prior work on CFG-aware quantization. GAMP focuses on guided-output calibration and mixed-precision allocation, whereas GCBT introduces CFG branch-space transform coding and a guidance-aware, quantizer-independent objective for selecting the coding basis. These approaches address different aspects of quantization and are therefore complementary. Since GAMP has no public implementation, we evaluate GCBT on top of GAMP-inspired calibration under the same SVDQuant host. As shown in Table 1, GCBT further reduces LPIPS by 2.97% on MJHQ and 2.47% on sDCI relative to this calibration baseline, demonstrating additional benefits from branch-space coding.

(2) When GCBT Helps (Q2). GCBT is particularly beneficial when activation quantization error remains substantial and CFG branch correlations can be exploited. As shown in Table R3 (Reviewer wcpw, Q3), even fixed +45° and PCA/KLT rotations improve over Native, supporting the effectiveness of branch-space coding itself. Our Q-DiT/SANA A4/A8 experiments (Table R1) further show that GCBT substantially reduces local quantization error under both bitwidths, but denoiser-output MSE improves by 78.55% under A4 versus only 4.30% under A8, where the baseline is already close to the W4-only reference. These results help explain why GCBT can yield smaller output-level gains at higher activation precision.

Thank you again for your careful review and constructive feedback. We have added experiments to validate the connection between GCBT's closed-form optimum and the actual quantization-error optimum, justify the use of the final CFG direction as a layer-local sensitivity measure, and evaluate sequential execution and dynamic CFG. We also clarified GCBT's scope, its distinction from prior CFG-aware quantization methods, and when it is most effective.

We hope these additional results and clarifications help address your main concerns, and would welcome any further feedback or questions!

Reviewer_c1RROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces branch-space transform coding for activation quantization in diffusion models with active two-branch classifier-free guidance. At each quantization site, corresponding conditional and unconditional activations are treated as a two-dimensional source, rotated before the existing quantizer and inversely transformed afterward. GCBT selects a per-layer angle from uncentered branch second moments and the CFG readout direction, with a closed-form optimum under an equal-rate quantization-noise surrogate. Experiments span image and video models and several PTQ hosts, using paired prompts/seeds and confidence intervals. Representative W4A4 gains include PixArt-Σ LPIPS of 0.327 to 0.296 and Wan2.1-1.3B LPIPS of 0.545 to 0.504; representative deployed hosts incur approximately 1–2% additional generation latency.

Critical strengths
  1. A simple, distinct and composable quantization design variable. The method transforms corresponding activations across CFG branches, whereas common rotation methods operate across features. The closed-form angle avoids gradient optimization, and shared linear/feature transforms admit a clear composition rule. This makes the idea easy to inspect and potentially useful across existing PTQ systems, despite transform coding and CFG-aware quantization having precedents.
  2. Broad paired evidence with small measured deployment overhead. Improvements appear across multiple W4A4 image/video settings, supported by paired intervals and supplementary VBench metrics. For example, PixArt-Σ LPIPS decreases by approximately 9.5%, while representative generation latency increases only 1–2%. This combination makes the contribution practically consequential; the weaker A8 gains do not erase the low-bit evidence.
Critical weaknesses

Missing simple branch-basis controls limit attribution to the proposed angle-selection rule. The reported near-diagonal branch geometry makes fixed normalized sum/difference coding a particularly relevant control. However, the DiT-XL/2 median direction does not establish the actual per-layer angles in the PixArt-Σ ablation. Comparisons with fixed 45° and principal-axis/KLT coding under identical conditions are needed to determine whether GCBT’s adaptive criterion provides benefits beyond branch-space coding itself. If these controls obtain similar gains, the empirical contribution of the angle-selection rule would be less clearly established, while the value of identifying branch space as a coding axis could remain.

Questions for authors

Requested attribution control: On the existing PixArt-Σ/SVDQuant W4A4 ablation, how do fixed 45° normalized sum/difference coding and per-layer principal-axis/KLT coding compare with GCBT under the same prompts, bit budget and calibration protocol? Please provide paired differences, especially GCBT versus these controls.

3 comments

We thank the reviewer for the thoughtful and detailed comments. We address the specific points below:

W1 & Q1 — Comparison with Fixed 45° and PCA/KLT Rotations.

We compare GCBT with fixed +45° and layerwise PCA/KLT under identical PixArt-Σ/SVDQuant W4A4 settings, using 256 paired prompts with matched calibration and bit budgets.

The LPIPS values are 0.327 for Native, 0.310 for fixed +45°, 0.308 for PCA/KLT, and 0.296 for GCBT (Table R3, Reviewer wcpw, Q3). Table R4 reports the paired differences (GCBT − control), with Bonferroni-adjusted 95% CIs from 10,000 paired bootstrap resamples.

Table R4. Paired performance differences (GCBT − control) with Bonferroni-adjusted 95% CIs on 256 prompts.

ComparisonΔLPIPS [95% CI]ΔSSIM [95% CI]ΔPSNR [95% CI]
GCBT − Fixed +45°−0.014 [−0.025, −0.003]+0.015 [+0.007, +0.023]+0.064 [−0.193, +0.330]
GCBT − PCA/KLT−0.012 [−0.024, −0.001]+0.016 [+0.007, +0.024]+0.062 [−0.199, +0.328]

Both fixed +45° and PCA/KLT improve over Native, confirming the benefit of CFG branch-space coding itself. GCBT further achieves statistically supported improvements in LPIPS and SSIM over both controls, while maintaining comparable PSNR. These results demonstrate that GCBT's guidance-aware rotation selection provides additional benefits beyond branch-space coding alone.

Thank you again for your helpful feedback and suggestions. We have added comparisons with fixed +45° and PCA/KLT rotations under identical settings, showing that GCBT achieves statistically supported improvements in LPIPS and SSIM over both baselines.

We hope these additional results help address your main concern, and would be happy to discuss any remaining questions or suggestions!

Thank you for the additional experiments. The comparisons with fixed 45° and PCA/KLT rotations adequately address my main concern. I appreciate the authors' efforts in conducting these additional experiments.

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

This paper proposes GCBT, a post-training quantization method that exploits the correlation between conditional and unconditional activations in CFG-based diffusion models. It uses a closed-form branch-space rotation to reduce guidance-weighted quantization error. Experiments on several image and video diffusion models show improvements over existing PTQ methods with little additional inference cost.

Critical strengths
  1. The method is simple and easy to integrate. The rotation has a closed-form solution and does not require additional training. It can also be combined with existing PTQ methods without changing model weights.
  2. The experiments cover a reasonable range of models and quantization methods. The improvements are particularly clear under W4A4 settings, and the reported runtime overhead is only 1–2%.
Critical weaknesses
  1. The assumptions behind the optimal rotation need further validation. Equation (6) assumes that quantization error is proportional to the projected second moment and ignores cross-coordinate error terms. These assumptions simplify the derivation, but it is unclear how well they hold for actual low-bit quantizers. The paper also lacks a comparison between the closed-form rotation and the optimal angle found using actual quantization error.
  2. The gains are uneven across different settings. GCBT improves SVDQuant W4A4 by a noticeable margin, but the improvements are very small for Q-DiT W4A8 on SANA-1.6B (Table 2) and OpenSora (Table 3). The paper gives little explanation for these differences. It is also unclear whether the method works equally well with dynamic CFG.
Questions for authors
  1. Have the authors measured the actual quantization error variances and cross-coordinate error terms to check the assumptions in Eq. (6)?
  2. How close is the closed-form rotation to the best rotation found by directly searching over angles using actual quantization error?
  3. Could the authors compare GCBT with a fixed 45-degree rotation and a PCA-based rotation? This would help clarify how much of the gain comes from the guidance-aware design.
  4. Can GCBT be applied when the CFG scale changes across denoising steps? If so, would the rotation need to be updated at each step?
3 comments

We thank the reviewer’s thoughtful feedback and constructive suggestions. We address each point below:

W1 & Q1 — Validation of Eq. (6)’s Proportionality and Cross-Coordinate Error Term Assumptions.

We have clarified the assumptions behind Eq. (6) in the revised manuscript (highlighted in blue) and added empirical validation under the original quantization settings. Below, we address each point in details.

(1) Proportionality assumption. Let DfullD_{\mathrm{full}} denote the actual quantization distortion and DdiagD_{\mathrm{diag}} its diagonal-only approximation. GCBT minimizes the surrogate DsurD_{\mathrm{sur}}, assuming that DdiagD_{\mathrm{diag}} is approximately proportional to DsurD_{\mathrm{sur}} through a shared coefficient κℓ\kappa_\ell. To validate this assumption, we estimate one κℓ\kappa_\ell per site using calibration data and evaluate how accurately it predicts quantization-error second moments on held-out samples across 360 rotation angles. We observe a median per-site Pearson rr of 0.999 (167/168 sites above 0.99) and a median held-out R2R^2 of 0.999 for the proportional fit. This indicates that deviations from exact proportionality introduce only minor prediction errors at most evaluated sites.

(2) Cross-coordinate error terms. We further examine whether ignoring cross-coordinate error terms affects rotation selection. We find that minimizing measured DdiagD_{\mathrm{diag}} and DfullD_{\mathrm{full}} yields similar angles at most sites, with a median/P90 angle difference of 0.05∘/2.72∘0.05^\circ/2.72^\circ and only 5/168 sites differing by at least 5∘5^\circ. Furthermore, ignoring these terms increases aggregate held-out DfullD_{\mathrm{full}} by only 0.129%0.129\% (95% CI: [-0.006%,+0.254%]). These results indicate that neglecting cross-coordinate error terms has limited impact on rotation selection and aggregate distortion.

W1 & Q2 — GCBT Objective versus Actual Quantization Error Objective.

We further compare the GCBT closed-form rotation minimizing DsurD_{\mathrm{sur}} against numerical search minimizing measured DfullD_{\mathrm{full}} to assess the combined effect of both approximations. Results indicate that GCBT closely matches numerical search under DfullD_{\mathrm{full}}. Accounting for equivalent rotations, the selected angles differ by less than 1∘1^\circ at 141/168 sites (83.9%), with a mean absolute difference of 1.02∘1.02^\circ. Furthermore, GCBT increases aggregate held-out error by only 0.437% (95% CI: [0.309%, 0.558%]) relative to numerical search. Therefore, GCBT closely matches the minimum measured quantization distortion while retaining a quantizer-independent closed-form solution that requires no angle search.

Setup. We evaluate both assumptions on PixArt-Σ/SVDQuant W4A4 across 168 quantization sites, using 64 calibration samples and 64 independent held-out samples. Because DfullD_{\mathrm{full}} and DdiagD_{\mathrm{diag}} cannot generally be minimized analytically, we perform a three-stage angle search: a global 1∘1^\circ scan, followed by 0.1∘0.1^\circ refinement and a 0.05∘0.05^\circ convergence check. Each objective independently selects the angle minimizing its measured distortion from the same combined candidate set.

W2 & Q4 —Uneven gains across settings and Dynamic CFG.

(1) Uneven gains across settings. We believe the smaller gains under W4A8 are partly due to its already low baseline quantization error. To validate this, we conduct a controlled A4/A8 comparison on SANA using a reconstructed Q-DiT setup. As shown in Table R1, GCBT reduces local MSE by 84.06% under A4 and 97.36% under A8. However, the baseline local MSE under A8 is already about 28.36× lower than under A4, resulting in a much smaller absolute reduction. Consistently, denoiser-output MSE decreases by 78.55% under A4 but only 4.30% under A8, where the baseline is already close to the W4-only reference. Therefore, GCBT remains effective under A8, but the lower baseline error leaves less room for further output-level improvement.

Table R1. Local quantization MSE under A4 and A8 on SANA-1.6B using a reconstructed Q-DiT setup (not the exact Table 2 checkpoint).

BitsIdentity MSE ↓GCBT MSE ↓ΔMSE ↑Reduction (%) ↑
A434.6835.52729.15584.06%
A81.2230.0321.19197.36%

(2) Compatibility with Dynamic CFG. GCBT can be directly applied to timestep-varying CFG. By substituting the current CFG scale sts_t into our closed-form solution, we can update the rotation at each timestep while keeping the calibrated Σℓ\Sigma_\ell fixed. The updated rotation remains optimal for DsurD_{\mathrm{sur}} at each step with no additional calibration or angle search. We validate this on CogVideoX (Table R2), which uses a dynamic CFG schedule.

Table R2. Comparison on CogVideoX under dynamic CFG (cosine schedule from 2 to 6, 50 steps). Fixed-angle GCBT uses the reference scale sref=6s_{\mathrm{ref}}=6.

MethodLPIPS ↓SSIM ↑
Native0.4450.689
Fixed-angle GCBT0.4150.697
Timestep-adaptive GCBT0.4090.699

Q3 — Comparison with Fixed 45° and PCA/KLT Rotations.

We additionally compare GCBT against fixed +45° and layerwise PCA/KLT under identical PixArt-Σ/SVDQuant W4A4 settings, using 256 paired prompts with matched calibration and bit budgets.

Table R3. Comparison of rotation strategies on PixArt-Σ/SVDQuant W4A4 using 256 paired prompts with identical calibration and bit budgets.

MethodLPIPS ↓SSIM ↑PSNR ↑
Native0.3270.65117.472
Guidance0.3150.67817.707
Fixed +45°0.3100.68017.853
PCA/KLT0.3080.67917.854
GCBT (Ours)0.2960.69517.917

As shown in Table R3, both fixed +45° and PCA/KLT improve over Native, supporting one of our core contributions: identifying CFG branch-space coding as an effective approach to activation quantization, even without our guidance-aware objective. GCBT further improves upon both controls, demonstrating the additional benefit of our guidance-aware objective. Paired differences and confidence intervals are provided in Table R4 (Reviewer c1RR).

Thank you again for your thoughtful review and helpful suggestions. We have added experiments to validate the assumptions behind Eq. (6), compare GCBT with actual-error optimization and simpler rotation baselines, adding more ablations to validate gaining from GCBT, and investigate dynamic CFG and uneven gains across settings.

We hope these additional results help address your main concern, and would be happy to discuss any remaining questions or suggestions!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Additional Experimental Evidence

We thank the reviewers for their constructive feedback. As the discussion period draws to a close, we summarize our key responses and additional evidence below.

(1) Validation of the closed-form rotation. Reviewers wcpw and 27Dh questioned the validity of the assumptions underlying GCBT's closed-form rotation. Specifically, GCBT assumes that (i) the diagonal-only quantization distortion DdiagD_{\mathrm{diag}} is approximately proportional to the GCBT objective DsurD_{\mathrm{sur}}, and (ii) ignoring cross-coordinate error terms in actual quantization distortion DfullD_{\mathrm{full}} has limited impact on the selection of rotation angles.

We validate these assumptions with additional experiments. First, the proportionality assumption holds well in practice with predicted quantization distortion matching measured distortion almost perfectly under a proportional fit (median held-out R2=0.999R^2 = 0.999). Second, ignoring cross-coordinate error terms has limited impact on rotation selection, resulting in only 0.129% higher aggregate distortion compared with directly minimizing DfullD_{\mathrm{full}}. Furthermore, GCBT selects angles close to those that directly minimize actual quantization distortion, with a mean angle difference of only 1.02° and 0.437% higher aggregate quantization distortion. These results support the GCBT objective as an effective surrogate for the actual quantization objective.

(2) Attribution of GCBT's improvements. We first emphasize that CFG branch-space coding is a core contribution of this work. Building on this, we propose GCBT, a training-free method that selects rotations using a guidance-aware objective with a closed-form solution. In response to Reviewer c1RR (W1/Q1) and Reviewer wcpw (Q3), we compared GCBT with fixed +45° and PCA/KLT rotations under paper’s setting. Both simple rotations improve over Native, showing that CFG branch-space coding itself is effective. GCBT further improves upon both controls, improving over both baselines across evaluation metrics. These results show the additional benefit of guidance-aware rotation selection.

(3) Theoretical and empirical justification of layer-local rotation selection. In response to Reviewer 27Dh's W2 concern, we clarified two key points: (i) GCBT's objective is itself layer-local (Eq. 7), and (ii) the "final CFG direction" refers to the known CFG readout direction gs=[s,1−s]⊤g_s=[s,1-s]^\top defined in Eq. (1), not the final output. Each rotation is determined solely by this direction and layer-local activation statistics, without using final outputs.

For theoretical justification, the CFG readout direction defines the exact CFG-weighted local error qs(R)=sec(R)+(1−s)eu(R)q_s(R)=se_c(R)+(1-s)e_u(R), accounting for how the two branch errors are amplified or canceled by CFG weighting. Empirically, we justify this choice with Table R5, where layer-wise objective recovers 90.92% of all possible improvement and incorporating CFG in branch weights improves output MSE by 24.8% and 23.2%.

(4) Practical applicability and performance characterization. Reviewers wcpw and 27Dh raised concerns about uneven improvements across quantization settings and applicability to different CFG execution schedules. We conducted additional experiments to address these concerns:

  • Dynamic CFG: GCBT supports timestep-varying CFG through timestep-specific closed-form rotations without additional calibration or angle search. On CogVideoX, adaptive GCBT improves LPIPS from 0.445 to 0.409.
  • Sequential CFG execution: We implement a layer-wise interleaved schedule that produces bitwise-identical outputs to joint GCBT, with only 3.7% additional peak memory compared with Native sequential execution.
  • Uneven gains: Controlled A4/A8 experiments on SANA show that higher activation precision leaves less room for output-level improvement. Although GCBT substantially reduces local quantization error under both settings, denoiser-output MSE improves by 78.55% under A4 versus 4.30% under A8, where the baseline error is already much lower.

Comparison with prior work. Unlike GAMP's guided-output calibration and mixed-precision allocation, GCBT optimizes the coding basis of CFG activations. Under the same SVDQuant host, GCBT consistently improves LPIPS over GAMP-inspired calibration on both MJHQ and sDCI datasets.

Practical relevance. CFG remains in use in recent models such as DreamZero and LTX-2, showing that few-step distillation does not necessarily eliminate CFG. GCBT applies to active two-branch CFG, but not to guidance-distilled models without active CFG or single-branch models.

Overall, these results further support CFG branch-space coding and GCBT's closed-form rotation selection. We have updated the manuscript to clarify its theoretical justification, assumptions, and practical scope.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.