acceptodds
Under review as a conference paper at ICLR 2027

Better Positions or Different Values? Rethinking Confidence Ranking in Guided Masked Image Generation

Abstract

Confidence ranking in masked image generation commits the positions whose candidate tokens are most probable, and its advantage over random positions depends on the sampling temperature. Hayakawa et al. (2026) explained it as implicit cooling on an unconditional image model. With assumptions about complete generation, we turn this cooling interpretation into a temperature rule for text-to-image models with classifier-free guidance. We develop the rule on one model and test it on three more. As the rule predicts, confidence ranking helps above the best tested temperature for random positions but brings no significant improvement at that temperature. On MMaDA-8B, confidence ranking improves Fréchet Inception Distance (FID) by up to 51.8 at a high temperature, whereas the released confidence ranking worsens FID by 14.1 at the best tested temperature for random positions. After each sampler receives its own best tested temperature, confidence ranking has no significant FID advantage in three model families at released step counts. We extend the single-step analysis to many simultaneous commits and to position scores computed from probabilities at temperature one. This improves predictions of committed values and provides a starting point for tuning ranking strength. A controlled change to position scores shows that selection can weaken the guidance reflected in committed values even when candidate-value sampling stays fixed. These findings call for temperature-calibrated baselines and guidance controls when the scoring rule changes effective guidance.

Primary area:
generative models
Keywords:
masked image generation, sampling temperature, classifier-free guidance, parallel decoding

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

The paper revisits the advantage of confidence ranking over random positions in masked image generation. The authors extend the "implicit cooling" explanation to text-to-image models with CFG and establish a temperature rule: ranking helps only at temperatures above the random-position optimum. Validated on four models, same-temperature comparisons show ranking gains of up to 51.8 FID, but after each sampler calibrates its own temperature, the FID advantage is not significant in three model families, and the released ranking is actually worse at the best temperature (MMaDA 14.1). The paper also proposes a saturation-aware finite-gate approximation for recommending ranking strength, and points out that position scoring can change effective guidance (COND probe and Proposition 1). The conclusion is that evaluating sampling rules requires temperature-calibrated baselines and guidance controls.

Critical strengths

Rigorous evaluation design: The rule was developed on aMUSEd-256, after which settings, metrics, interval methods and decision rules were frozen before testing on other models; the MMaDA predictions were made before the results; tuning and evaluation use different prompts, paired bootstrap, matched per-step commit counts (removing the effect of BF16 ties), and a values-only control (+0.60 and −1.61 FID, both not significant), which separates "better positions vs. different values" fairly cleanly. "Random-position baselines that need temperature calibration" is a valuable corrective for the community. Theory and experiments corroborate each other, and negative results are reported honestly: The finite-gate approximation clearly beats the power form under dense commits; COND's effective guidance has an FP64 controlled verification and a +7.02 FID effect in full generation; the authors also explicitly acknowledge that the recommendation rule did not select the best strength, that plain guidance tuning is stronger, and that sampler equivalence is not established.

Critical weaknesses

The core conclusion rests on "not significant ≠ equivalent" and a limited temperature grid. The intervals in Table 2 are not narrow (Meissonic g=3: [−2.43, 0.89]; MMaDA: [−1.75, 1.38], [−2.85, 0.15]), so they neither prove no difference nor exclude real gains on the order of 1–2 FID; the authors also state that sampler equivalence is not established. s* is approximated from only a few tested temperatures, and both L1–L3 and "no gain after calibration" depend on that approximation; the prompts of the two MMaDA evaluations (MMaDA† and MMaDA∗) overlap; LaViDa-O is evaluated only on alignment metrics, and few-step settings lack independently calibrated random-position temperatures (acknowledged by the authors in the Limitations). "Ranking has no value" should therefore be restricted to "the tested models, released step counts, and the FID metric".

Questions for authors

[1] Can you redo the calibrated comparison with a finer or continuously optimized temperature search, and run an equivalence test (TOST) with a pre-specified margin (e.g., ±1 FID)? [2] The evaluation prompts of MMaDA† and MMaDA∗ overlap; can you provide results on a fully independent third evaluation set? [3] Under few-step settings (which the paper says lack independently calibrated random-position temperatures), does "no gain after calibration" still hold? [4] Beyond FID, what are the calibrated comparison results on GenEval, CLIP, or human-preference metrics? [5] How large is the deviation of the finite-gate "fixed threshold" assumption on LaViDa-O and MMaDA? Why does it recommend r=4 on MMaDA when the best is r=16?

2 comments
AuthorsOpenReview ↗

Response to Reviewer GGWL

Thank you for the detailed review, and for describing the evaluation design as rigorous and the calibrated random baseline as "a valuable corrective for the community". Following your suggestion, the revision restricts our conclusions to the tested models, released step counts and FID. In place of "not significant", Table 2 now reports one-sided 95% upper limits of ranking's gain, which are below 2 FID in all 12 held-out evaluations and below 1 FID in nine.

Q1. Finer temperature search and TOST. We have done both. With finer temperature grids, little FID gain remains. The TOST, with your ±1 FID margin and the criterion (90% interval inside the margin) fixed before generating new images, establishes equivalence on aMUSEd and MaskGIT at all four step counts (Appendix G.1, Table 19). Intervals in Q1 are 90%.

  • aMUSEd: we re-tuned both samplers on a 0.025 temperature grid on development prompts (selection unchanged) and evaluated them on 4,000 new prompts with four images each: −0.32 [−0.61, −0.05], equivalent.
  • Class-conditional MaskGIT: the 0.025-grid selections, re-evaluated on 50,000 new images per cell, give +0.21 [0.08, 0.34], +0.22 [0.14, 0.31], +0.14 [0.05, 0.21] and +0.16 [0.08, 0.24] at 4, 8, 16 and 32 steps, all equivalent.
  • LaViDa-O, now also compared on FID: a finer search for the random-position temperature on separate prompts (0.8 → 0.95) lowered ranking's gain on the evaluation prompts from +1.37 to +0.49 [0.07, 0.91]. On 2,300 new prompts the gain is +0.83 with upper limit 1.44, below 2 FID but above the ±1 margin.
  • Meissonic and MMaDA: on Meissonic at g = 6, random positions are better (−0.83). The MMaDA intervals are too wide to establish equivalence, and their upper limits are at most 1.01.

Q2. A fully independent MMaDA set. We added a third set of 1,400 COCO prompts that share no image with any tuning or earlier evaluation set, and fixed the arms and statistics of Table 2 before generation. Random positions have the lower FID. Ranking's gain is −1.32 with 95% interval [−3.16, −0.28], and even its one-sided 95% upper limit, −0.46, is negative. The MMaDA conclusion therefore does not rest on overlapping prompts. On the same images CLIP-L slightly favors ranking (see Q4).

Q3. Few-step settings. For aMUSEd and MaskGIT, yes: every interval lies below 1.4 FID. On MMaDA the intervals are wider. With both samplers tuned on separate prompts and evaluated on new ones (Appendix G.3), aMUSEd gives −0.33 [−1.68, 1.08] and +0.36 [−0.67, 1.35] at 4 and 8 steps, and MaskGIT at 4 steps is equivalent within your margin (Q1). On MMaDA at 8 steps the gain is +1.34 [−1.41, 3.45]. The largest gain, +2.96 [0.19, 4.90], came from a noise-matched ranking arm that we added after seeing this result, and fell to +0.89 [−0.72, 2.46] on 1,400 new prompts.

Q4. GenEval, CLIP and human preference. After calibration on FID, we detect no advantage for ranking on GenEval or a learned human-preference score (PickScore), and ranking keeps a small text-alignment advantage on CLIP-L (Table 29).

  • GenEval: every interval includes zero (aMUSEd +0.06 [−0.84, 0.96], Meissonic g = 6 +0.19 [−0.92, 1.29], LaViDa-O −1.34 [−3.45, 1.10] points).
  • CLIP: CLIP-L favors ranking in four of five comparisons, by 0.56 to 2.2 × 10⁻³, and favors random positions on aMUSEd.
  • Human preference: PickScore does not favor ranking on any of four models. It favors random positions on aMUSEd (−0.043, 98.75% interval [−0.081, −0.006]) and, with the finer random-position temperature, on LaViDa-O (−0.043 [−0.069, −0.017]).

§7 now recommends calibrating the baseline on the target metric.

Q5. Fixed-threshold deviation and the r = 4 pick.

How large is the deviation? It is much smaller than the power form's. The table gives the mean total variation between the finite-gate law and Monte Carlo estimates of the committed-value distribution, with the power form's error in parentheses (Appendix G.6):

Committed fraction4%12%33%50%
MMaDA, r = 160.0012 (0.013)0.0027 (0.092)0.0039 (0.39)0.0029 (0.56)
MMaDA, r = 1 (released)0.005 (0.017)0.015 (0.13)0.030 (0.53)0.024 (0.73)

On LaViDa-O at a = 1.5, over committed fractions of 0.7–45%, it is at most 0.0004 (power form 0.001–0.14).

Why r = 4 when the best is r = 16? The pick comes from the entropy-matching step, not from the local law. Entropy matching placed r = 4 and r = 16 at matched temperatures 0.95 and 1.05, almost equally close to s* = 1.0, and picked r = 4 (§4.5). The resulting gap is 1.2 FID with an interval that includes zero, and the broader prediction fixed before evaluation, that the best r is 4 or 16, held. In the five tests fixed before generation, our recommendation is never farther from the best than the power form's, so we present it as a starting point for a strength search.

AuthorsOpenReview ↗

Follow-up to Reviewer GGWL

Dear Reviewer GGWL, thank you again for your detailed review. As the discussion period ends soon, we would be grateful to know whether our response addresses your questions. In brief, Table 2 now reports one-sided upper limits instead of "not significant", the TOST with your ±1 FID margin establishes equivalence on aMUSEd and MaskGIT, and a fully independent MMaDA set favors random positions (Appendix G.1). We would be glad to clarify any remaining point.

Reviewer_WVxFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies why confidence ranking changes the behavior of guided masked image generation. Because a proposed token's value affects whether its position is committed, ranking changes the distribution of retained values as well as the selected positions. Building on prior implicit-cooling analyses, the paper develops a finite-gate approximation and accounts for differences between proposal and scoring distributions. It uses these ideas to examine effective temperature and guidance, together with controlled sampling interventions. Its main contribution is a more precise interpretation of a commonly used sampling operation.

Critical strengths
  1. The motivation follows from the structure of the sampler. Studying the token distribution conditional on commitment directly exposes a confound in attributing gains to better position selection. Accounting for gate saturation and for distinct proposal and scoring distributions makes the analysis relevant to practical guided samplers, beyond the earlier sparse-selection cooling explanation. The COND intervention is particularly informative: changing the ranking score while leaving the candidate proposal unchanged can alter effective guidance. Matching the conditional and unconditional logit coefficients gives an explanation for this effect, rather than treating it as another sampling heuristic.

  2. The empirical comparisons are designed around that mechanism. Fixed commit counts, separate tuning and evaluation, and guidance interventions with matched stepwise temperatures help distinguish explanations that ordinary sampler leaderboards conflate. The paper also reports failed predictions, uncertainty in the native BF16 parameter mapping, and cases where ordinary guidance tuning outperforms the predicted schedule. These disclosures support the paper's value as an analysis of sampling behavior. It does not need to produce the best sampler to make a useful contribution.

Critical weaknesses
  1. Positions and values are coupled, so the proposed intervention does not fully decompose their causal contributions. At a fixed state, the exact factorization is P(i committed, X_i = x) = Z_i q_tilde_i(x). A value-only sampler that selects positions uniformly replaces Z_i with k/m and also changes the dependence induced by joint top-k selection. Even a perfect q_tilde_i would therefore not reproduce the original joint transition. Reproducing the FID reversal shows that changing values is sufficient to produce that phenomenon; it does not measure how much position selection contributes in the original sampler, nor make the remaining discrepancy a position effect. Appendix B and Section 7 acknowledge this limitation, so I do not read the paper as claiming that positions are irrelevant. I would bring this inseparability into the main framing earlier, because the question posed by the title can otherwise suggest a stronger attribution than the intervention identifies.

  2. The move from local selection laws to a trajectory-level temperature explanation requires substantial modeling assumptions. A finite-gate distribution generally lies outside the two-parameter temperature/guidance family, and equal entropy does not imply equal token probabilities. The identity of retained tokens matters for semantics and for the next conditional prediction, not just for distribution sharpness. Consequently, a local entropy bound, a unimodal quality-versus-temperature assumption, and one entropy-matched temperature for an entire trajectory are distinct steps. More accurate local Monte Carlo estimation cannot remove the approximation between them. The paper is persuasive about local value reweighting, but less complete as an explanation of which retained values become useful future context. I would distinguish the local distribution law, the qualitative rule obtained under assumptions, and the automatic parameter recommendation more clearly.

Questions for authors
  1. Can the main claim explicitly distinguish showing that value changes can reproduce the ranking reversal from identifying the original position and value contributions? What conclusions remain identifiable when Z_i and top-k dependencies change under the intervention?
  2. Under what conditions is trajectory-level entropy matching a useful approximation when similar entropy can conceal different semantic token choices and future contexts? Which conclusions depend on this approximation, and which follow from the local selection law alone?
4 comments
AuthorsOpenReview ↗

Response to Reviewer WVxF

Thank you for a careful and constructive review. We appreciate your view that the paper's value lies in "a more precise interpretation of a commonly used sampling operation" rather than in a better sampler, and that the COND intervention is "particularly informative".

W1 & Q1. What the interventions identify.

Can the main claim make this distinction? We have revised the main claim to distinguish showing that value changes reproduce the ranking reversal from identifying the original position and value contributions. The value-only sampler replaces ZiZ_i by k/mk/m and drops the dependence induced by joint top-k selection, so its success shows sufficiency only, and its gap to ranking mixes position preference, dependence and approximation error. The first paragraph of §1 now states this coupling, and §4.4 and §7 answer the title's question in these terms.

What remains identifiable? The effect of each intervention as a whole, from which three conclusions follow:

  • Changing only the values suffices. In a factorial on five models, random positions with finite-gate values reach 95–104% of ranking's FID gain at a shared temperature and released step counts, and every lower 95% limit is at least 85% (Table 2b).
  • The choice of positions matters when the joint value law is held fixed (LaViDa-O, 2,500 new prompts, 98.3% intervals, Appendix G.5). A ranked set and a random set draw values from the same set-conditional joint law P(XS∣S)P(X_S \mid S), computed by numerical integration, and each arm runs its own trajectory. The ranked set is better by 1.09 [0.17, 2.07] FID. As in the value-only sampler, we then removed the top-k dependence in two steps. Dropping the dependence between values changed FID by +0.32 [−0.64, 1.24], and then dropping the conditioning on the whole set by +0.10 [−0.79, 1.02]. Both intervals include zero.
  • The gain has no unique split. On LaViDa-O, ranked positions gain 14.5 [10.7, 19.8] FID over random ones when both draw proposal values (P−U), whereas ranking gains 1.25 [−0.55, 3.17] over random positions with finite-gate values (C−V), as your argument predicts (Appendix G.4).

W2 & Q2. Trajectory-level entropy matching.

Under what conditions is it useful? We cannot give a general condition that guarantees accurate full-run predictions: as you note, equal entropy does not determine which tokens are kept or the context they create for later steps, and §7 now says this lies outside the local law. Empirically, within one state it is nearly exact under weak ranking: on 32 MMaDA states at the native commit fraction, its KL loss is at most 0.0023 nats per committed token for r ≥ 64, against 0.031–0.044 under strong ranking (Appendix G.7). Over a whole run, Table 10's prediction has the correct sign for both main contrasts (r = 16 and 256) and small residuals near the released setting (−1.8 and +2.8 FID at r = 1 and 16, against gains above 30), but overstates the gain of weak ranking, so it predicts signs, not sizes.

Which conclusions depend on it, and which follow from the local selection law alone? We adopt your three-part distinction in §7:

  • Local selection law alone (fixed states). Mean TV between the finite-gate law and Monte Carlo committed values is at most 0.0304 on MMaDA (power form 0.73) and 0.00041 on LaViDa-O (0.14) (Appendix G.6). The value-only and factorial results above apply this law per state, without entropy matching, and the calibrated comparisons are purely empirical.
  • Qualitative rule under two assumptions (§4.1): quality is unimodal in temperature, and the sparse cooling map approximately governs full-run performance. The local entropy bound does not imply the second, so we test the rule L1–L3 as a prediction. L1 and L2 met their stated criteria on the three test models (Table 8).
  • One entropy-matched temperature per run. Only the automatic strength recommendation and the full-run predictions of Table 10 depend on it. The recommendation's largest miss, on LaViDa-O, comes from this summary and the coarse optimum s* = 0.8, not from the local law: state by state, the best strength stays near the finer optimum 0.95, but the run summary maps it to 1.19 (Appendix G.8).

We therefore present the recommendation as a starting point for a strength search. In all five tests fixed before generation, it is within one tested step of the best strength, against up to four for the power form (Table 26). In two aMUSEd tests (s = 1.2, 1.4) with the search rule fixed in advance, the search reached the best tested strength within the four evaluations allowed. On the four 12-strength sweeps, the strategies compare as follows (Appendix G.8):

StrategyEvaluationsEnds at the bestFID gap to the best
Search from the finite-gate pick3.75100%0
Search from the power-form pick4.2575%0.08
Search from a random strength5.675%0.26
Same number of random strengths, no search3.7531%0.68
AuthorsOpenReview ↗

Follow-up to Reviewer WVxF

Dear Reviewer WVxF, thank you again for your careful review. As the discussion period ends soon, we would be grateful to know whether our response addresses your two concerns. In brief, the revised §1 now distinguishes showing that value changes suffice from identifying position and value contributions, and §7 separates the local selection law, the qualitative rule under its assumptions, and the strength recommendation, as you suggested. We would be glad to clarify any remaining point.

Thank you for your response. However, given that there are still substantial issues with the paper's experiments, I believe the manuscript would benefit from thorough revisions and resubmission in a future cycle. Therefore, I have decided to maintain my current score.

AuthorsOpenReview ↗

Reply to Reviewer WVxF

Thank you for reading our response. The two weaknesses in your review concerned what the value-only intervention identifies and the trajectory-level assumptions, and your review described the empirical comparisons as "designed around that mechanism". The revision addresses both weaknesses in §1 and §7, and adds a factorial and a joint-law control (Table 2b, Appendix G.4–G.5). If other experimental issues remain, we would be grateful to know which ones, so that we can address them as best as we can before the discussion closes.

Reviewer_4EmwOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies confidence ranking in guided masked image generation through its effects on committed token values. It extends the implicit-cooling analysis to denser commits and examines changes in effective guidance. Across three model families at released step counts, ranking has no significant FID advantage after each sampler receives separate temperature calibration.

Critical strengths
  1. The separate temperature calibration and held-out evaluations provide a useful reassessment of confidence ranking. The tuning procedures and uncertainty estimates are clearly documented.

  2. The finite-gate analysis and targeted controls help explain the observed effects. In particular, the value-only experiment reproduces the temperature-dependent reversal while keeping positions random. The paper also reports failed predictions and limitations openly.

Critical weaknesses
  1. The fixed-state analysis is more accurate than the full-generation predictions. Table 10 shows noticeable errors for some ranking strengths. This is acknowledged, but a brief explanation of when the full-run approximation is useful would help readers interpret it.

  2. The main conclusion is appropriately limited to FID. Some calibrated comparisons still favor ranking on CLIP-L (Table 15), so the practical recommendation could distinguish FID-focused use from settings where text alignment also matters.

Questions for authors
  1. Based on the existing results, what is the main limitation when using the finite-gate analysis to choose ranking strength?

  2. How would you interpret the remaining CLIP-L gains alongside the calibrated FID results?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 4Emw

We are grateful for your positive and constructive review, and encouraged that you find the separate temperature calibration "a useful reassessment of confidence ranking" and that the finite-gate analysis and targeted controls "help explain the observed effects".

Q1 & W1. The main limitation for choosing ranking strength. The main limitation is summarizing a whole run by one temperature, not the local law. The steps below go from single states to whole runs, then to when the full-run approximation is useful and how we use the recommendation.

  • Single states. The fixed-state law is accurate. On MMaDA its mean TV is at most 0.004 at r = 16 and 0.030 at the released r = 1, against up to 0.56 and 0.73 for the power form, and on LaViDa-O at most 0.0004 against 0.14.
  • Whole runs. In Table 10, the full-run prediction kept the signs of both main contrasts, which were fixed before any image, and its residuals are small near the release (−1.8 and +2.8 FID at r = 1 and 16, against gains above 30). It overstates the gain of weak ranking (r = 256 and 4096). There, one temperature describes each measured state almost exactly (Table 25), so the error enters when a whole run is summarized by one temperature. Under strong ranking, even a single state summarized by one temperature loses 0.031–0.044 nats per committed token (Appendix G.7).
  • When the full-run approximation is useful. It serves the sign of a gain and the location of a strength near the release, not the size of a gain, and Appendix C now says so next to Table 10. The clearest case is LaViDa-O, the recommendation's largest miss. Matched state by state, the best strength (a = 3) keeps every measured state near temperature 1.0, close to the finer random-position optimum 0.95, while the recommended a = 0.5 cools the early states to about 0.4. The run summary maps them to 1.19 and 0.82 instead (Appendix G.8).
  • How we use it. We present the recommendation as the start of a strength search. It is within one tested step of the best strength in nine of ten sweeps, and in all five tests fixed before generation it is within one step and never farther from the best than the power form's. With the rule and a budget of four evaluations fixed in advance, a local search from it found the best strength at both new temperatures within that budget. On the four 12-strength sweeps, it takes 3.75 evaluations on average and always ends at the best, whereas a search from a random starting strength needs 5.6 and ends at the best in 75% of starts, and the same number of random strengths contains the best in only 31% of cases (Appendix G.8). None of the calibrated comparisons rests on it.

Q2 & W2. CLIP-L gains after calibration. We read them as a small text-alignment advantage that remains after calibration on FID, and they are compatible with the FID results. Calibrating each sampler on FID matches image statistics, not prompt alignment, so ranking can keep this advantage while little of its FID advantage remains. CLIP-L favors ranking in four of five calibrated comparisons (Table 29): Meissonic g = 6 +0.56 [0.14, 0.99], Meissonic g = 3 +1.8 [0.05, 3.5], MMaDA† +2.1 [0.5, 3.7] and the new MMaDA set +2.2 [0.7, 3.7] (all × 10⁻³), with aMUSEd the exception (−0.94 [−1.35, −0.53]). Following your suggestion, §7 now distinguishes FID-focused use from settings where text alignment matters, and recommends calibrating the baseline on the target metric.

AuthorsOpenReview ↗

Follow-up to Reviewer 4Emw

Dear Reviewer 4Emw, thank you again for your positive and constructive review. As the discussion period ends soon, we would be grateful to know whether our response addresses your two questions. In brief, Appendix C now states when the full-run approximation is useful, and §7 now separates FID-focused use from settings where text alignment matters. We would be glad to clarify any remaining point.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response (to all reviewers and the AC)

We are encouraged that the reviewers see the temperature-calibrated random baseline as "a valuable corrective for the community" (GGWL) and "a useful reassessment of confidence ranking" (4Emw), find the evaluation rigorous (GGWL) and "designed around that mechanism" (WVxF), single out the COND intervention as "particularly informative" (WVxF), and value our open reporting of failed predictions (all three).

We agree with the three main points of the reviews and have revised the paper around them. As Reviewer WVxF notes, our value-only control shows that changing the committed values suffices to reproduce ranking's gain, not how that gain divides between positions and values. As Reviewer GGWL notes, a nonsignificant difference is not equivalence, and our conclusions hold for the tested models, released step counts and FID. As Reviewer 4Emw suggests, FID-focused use should be separated from settings where text alignment matters. New experiments support the revised claims, and the revised PDF marks new and changed text in blue. Intervals in all replies are 95% confidence intervals unless stated.

1. Calibrated comparisons. After separate tuning, the remaining gain is small and bounded. Table 2 now gives one-sided 95% upper limits next to the random-position FID of each evaluation, so each remaining gain can be read against its absolute level. In all 12 held-out calibrated evaluations at released step counts, the upper limit of ranking's FID gain is below 2 FID, and below 1 FID in nine. The sign depends on the model: ranking retains gains on LaViDa-O and MaskGIT, whereas random positions are better on Meissonic at g = 6 and on a new independent MMaDA set (−1.32 [−3.16, −0.28]). Finer tuning lowered LaViDa-O's residual from +1.37 to +0.49 FID on the same prompts. With the ±1 FID margin that Reviewer GGWL suggested, fixed before generating new images, ranking and random positions are equivalent on aMUSEd (4,000 new prompts) and on the class-conditional MaskGIT at 4, 8, 16 and 32 steps (Appendix G.1).

2. Value-only control. Changing only the committed values reproduces ranking's gain. In a 2×2 factorial design on five models (aMUSEd, Meissonic, MMaDA, LaViDa-O and a class-conditional MaskGIT on ImageNet), one arm keeps positions random and draws each committed value from the finite-gate law. At a shared temperature where ranking helps, this reaches 95–104% of ranking's FID gain on every model at released step counts, for gains from 4.7 to 102.2 FID (Table 2b). Consistent with Reviewer WVxF's argument, ranked positions with proposal values also help, and the two interventions do not add up, so there is no unique split. Even when a ranked and a random set draw their values from the same joint law, ranking the set still gains +1.09 FID on LaViDa-O (Appendix G.5).

3. Three tiers and the strength recommendation. The revision separates the fixed-state law, the temperature rule under its assumptions, and the strength recommendation, on which no calibrated comparison rests. We present the recommendation as the start of a strength search: a short local search from it reaches the best tested strength in all ten sweeps, and on the four 12-strength sweeps it needs 3.75 evaluations on average, against 4.25 from the power form's recommendation and 5.6 from a random starting strength (Appendix G.8).

4. Beyond FID. After calibration, GenEval shows no resolved difference, and PickScore, a learned human-preference score, shows none or favors random positions. CLIP-L favors ranking slightly in four of five calibrated comparisons, so §7 now distinguishes FID-focused use from alignment-focused use.

Revised contributions. Each contribution answers one of the three questions posed in §1.

  • Temperature and calibration (§4). Tests of the temperature rule on four text-to-image models with CFG. On these models and a class-conditional MaskGIT, upper limits on ranking's remaining gain after separate temperature calibration and evidence that changing only the committed values reproduces ranking's FID gain at a shared temperature.
  • Committed-value law (§3 and §4.5). A finite-gate law that improves committed-value predictions and ranking-strength recommendations beyond the sparse power form.
  • Effective guidance (§5). The change of effective guidance induced by position scoring.

Changes in the revised PDF. Section and table numbers are unchanged, and the new main-text table is Table 2b.

  • Abstract and §1. Upper limits and the sufficiency framing.
  • §4.2 and Table 2. Upper limits, the random-position FID of each evaluation and new evaluations. The equivalence tests fixed before generation are in Appendix G.1.
  • §4.4 and Table 2b. The factorial and the joint-law control.
  • §4.5. The strength recommendation as a starting point.
  • §7. Scope, and FID versus alignment.
  • Appendix G. All new experiments and protocols.
  • §2, §3, §6 and Appendix C. Short additions and language polishing.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.