- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper presents a test-time guidance method for diffusion models. The proposed methods outperform existing works using different versions of Stable diffusion under the same budget.
Critical strengths
- The motivation of maximizing certain rewards while increasing diversity is important.
- The experiment on different models demonstrated the generalizability of the proposed method to certain degree.
Critical weaknesses
- The exposition of this paper is not really good. For me, there are many unsure parts regarding the following points:
- First, it is unclear what exactly the goal of "test-time guidance". Specifically, for many other test-time optimization for image/video generation work, their goal is to generate "one" output that maximize the reward. However, it seems that this paper focus on generating multiple results while also maximize the diversity between outputs. I am not sure whether my understanding is correct at this moment, but the writing is not clear enough for me. The authors just suddently jump to what they propose. I believe clearer description is helpful for this.
- Is this method only focus on "black-box reward"? And how to define the "black-box reward"? Does it means the reward function is not formed by a differentiable function (e.g., not parameterized by a deep learning model?). Please be clear and describe this in detail. Without the above two points clarified, it is very difficult to understand and assess the proposed method.
- There are too few results shown in the paper, and it is not easy to assess the method based on those results. Meanwhile, it would be better to perform experiment on more recent models, e.g., FLUX.
- Is there any specific limitation or challenges that makes testing the proposed method on different models difficult?
Overall, with the confusion part clarified, I might be able to assess the proposed method more roughly. Therefore, I encourage the authors clearly clarified the confusing points listed above.
Questions for authors
Please check the questions listed in "Critical Weaknesses"
4 comments
Q2: Additional Baselines (EvoSearch, SMC / FK-Steering, Best-of-K, CADS, Particle Guidance, SPARKE) & Newer Model (Stable Diffusion 2.1) (Weakness 2 & Question 2)
We added 7 new baseline methods (Best-of-K (Top-4), Best-of-K + ASC, EvoSearch, SMC / FK-Steering, CADS, Particle Guidance, and SPARKE) on the 20 held-out prompts (Table R2), AND evaluated on Stable Diffusion 2.1 (768×768, v-prediction, OpenCLIP ViT-H/14) (Table R3):
Table R2: Comprehensive Baseline Comparison on Held-Out 20 Prompts (SD 1.4)
| Baseline Family | Method | Budget | ImageReward Best | ImageReward Mean | Worst-Case Min | Unseen PickScore Mean | Unseen Aesthetic V2 | CLIP-L/14 Div | Vendi (Cos / RBF) | RKE (Cos / RBF) |
|---|---|---|---|---|---|---|---|---|---|---|
| Unguided / In-Sampling | Standard DDIM | 4 | 0.6929 | 0.0445 | -0.6812 | 20.69 | 5.537 | 0.3519 | 1.690 / 2.838 | 1.326 / 2.248 |
| Diversity Guidance | CADS (ICLR 2024) | 4 | 0.6987 | 0.0362 | -0.6540 | 20.47 | 5.458 | 0.3887 | 1.751 / 2.976 | 1.367 / 2.418 |
| Particle Guidance (ICLR 2024) | 4 | 0.7279 | 0.0970 | -0.5812 | 20.73 | 5.531 | 0.3627 | 1.710 / 2.885 | 1.339 / 2.299 | |
| SPARKE (NeurIPS 2025) | 4 | 0.6780 | 0.0937 | -0.5904 | 20.71 | 5.512 | 0.3652 | 1.715 / 2.893 | 1.342 / 2.310 | |
| Best-of- (IID) | Best-of- (, Top-4) | 420 | 1.4528 | 1.3983 | 1.3533 | 21.05 | 5.605 | 0.2536 | 1.512 / 2.371 | 1.231 / 1.892 |
| (100-Prompt in Table R5) | Best-of- () + ASC () | 420 | 1.4528 | 1.0514 | 0.8227 | 20.89 | 5.568 | 0.3849 | 1.758 / 2.910 | 1.371 / 2.368 |
| FK-Steering / SMC | SMC / FK-Steering (Terminal) | 420 | 1.2821 | 1.2137 | 1.1240 | 21.18 | 5.641 | 0.1508 | 1.311 / 1.984 | 1.121 / 1.501 |
| Evolutionary Search | EvoSearch (Top-4 Elites) | 420 | 1.7029 | 1.6923 | 1.6815 | 21.68 | 5.742 | 0.0622 | 1.138 / 1.412 | 1.052 / 1.198 |
| MCMC Parallel | SPT (Terminal Cold Chain) | 420 | 1.4679 | 1.3563 | 1.2184 | 21.31 | 5.677 | 0.2583 | 1.520 / 2.388 | 1.236 / 1.907 |
| Tempering | SPT + History + ASC () | 420 | 1.5405 | 1.1007 | 0.8594 | 21.28 | 5.689 | 0.3684 | 1.727 / 2.861 | 1.352 / 2.311 |
| Ours (SPG) | TSC + Top-4 | 420 | 1.6170 | 1.6022 | 1.5921 | 21.64 | 5.719 | 0.0796 | 1.163 / 1.491 | 1.064 / 1.246 |
| SPG (, Strict Dominance) | 420 | 1.6236 | 1.3249 | 1.1640 | 21.51 | 5.724 | 0.2840 | 1.565 / 2.548 | 1.252 / 2.004 | |
| SPG (, Default Ours) | 420 | 1.6170 | 1.1871 | 0.9396 | 21.42 | 5.736 | 0.3599 (0.4150) | 1.702 / 2.800 | 1.336 / 2.227 | |
| SPG (, High-Diversity Ours) | 420 | 1.6236 | 1.0498 | 0.7031 | 21.38 | 5.718 | 0.3702 | 1.730 / 2.862 | 1.354 / 2.311 |
(Note: Terminal EvoSearch collapses portfolio diversity to , returning four near-identical clones.)
Table R3: Generalization to Stable Diffusion 2.1 (768×768, v-prediction, OpenCLIP ViT-H/14) & Fast 10-Step DPM-Solver++ ODE (Held-Out 20 Prompts)
| Model Architecture & Resolution | Method | ImageReward Best | ImageReward Mean | Worst-Case Min | CLIP-L/14 Div | -Center Min Pair Dist | Vendi (Cos / RBF) | RKE (Cos / RBF) |
|---|---|---|---|---|---|---|---|---|
| Stable Diffusion 2.1 | Standard DDIM () | 1.0174 | 0.3346 | -0.3307 | 0.3213 | 0.2304 | 1.637 / 2.723 | 1.290 / 2.137 |
(768×768, v-prediction, | SPT (Terminal, ) | 1.3981 | 1.1965 | 0.9864 | 0.2876 | 0.2062 | 1.575 / 2.563 | 1.260 / 2.008 |
| OpenCLIP ViT-H/14, | TSC + Top-4 (Ours) | 1.4694 | 1.4279 | 1.3892 | 0.1730 | 0.0804 | 1.347 / 1.987 | 1.148 / 1.563 |
| 50-Step DDIM) | SPG (, Strict Dominance) | 1.4694 | 1.2447 | 1.0557 | 0.3786 | 0.3058 | 1.751 / 2.905 | 1.359 / 2.342 |
| SPG (, Default Ours) | 1.4694 | 1.1259 | 0.8660 | 0.4206 | 0.3515 | 1.830 / 3.042 | 1.408 / 2.484 | |
SDXL 1.0 (1024×1024) | Standard 10-Step DPM++ () | 1.3025 | 0.8426 | 0.2161 | 0.2534 | 0.1574 | 1.505 / 2.418 | 1.222 / 1.870 |
| (Fast 10-Step | SPT (Terminal, 10-Step DPM++) | 1.5695 | 1.4179 | 1.2681 | 0.1875 | 0.1334 | 1.387 / 2.180 | 1.156 / 1.648 |
DPM-Solver++ ODE) | TSC + Top-4 (Ours) | 1.5859 | 1.5522 | 1.5282 | 0.1241 | 0.0659 | 1.258 / 1.793 | 1.100 / 1.402 |
| SPG (, 10-Step DPM++) | 1.5859 | 1.3450 | 1.1660 | 0.3098 | 0.2537 | 1.620 / 2.724 | 1.275 / 2.117 | |
| SPG (, 10-Step DPM++) | 1.5859 | 1.2544 | 1.0214 | 0.3474 | 0.2856 | 1.692 / 2.857 | 1.316 / 2.258 |
Key Findings from Tables R2 & R3:
EvoSearchandSMC / FK-Steeringsuffer from severe mode collapse: StandardEvoSearch (Top-4)collapses diversity to0.0622(Vendi = 1.138), andSMC / FK-Steering (Terminal)collapses to0.1508(Best = 1.2821) due to multinomial resampling degeneracy, whereasSPGpreserves high peak reward (1.6170–1.6236) while achieving2.4×–6.0×higher portfolio diversity (0.3599–0.3702).- On
Stable Diffusion 2.1(768×768,v-prediction, OpenCLIP ViT-H/14),SPGgeneralizes seamlessly across all 20 held-out prompts:SPG ($\tau=0.50$)strictly dominatesSPT (Terminal)on all reward and diversity metrics, improving Best ImageReward (1.4694vs.1.3981,+0.0713), Mean ImageReward (1.2447vs.1.1965,+0.0482), Worst-Case Min Reward (1.0557vs.0.9864,+0.0693), CLIP Diversity (0.3786vs.0.2876,+31.6%), and -center Min Pairwise Distance (0.3058vs.0.2062,+48.3%). At ,SPGboosts CLIP Diversity by+46.2%(0.4206) and increases -center Min Pairwise Distance by+70.5%(0.3515vs.0.2062). - On 10-step
DPM-Solver++fast ODE (SDXL 1.0,1024×1024),SPGincreases diversity by+85.3%(0.3474vs.0.1875) and more than doubles -center Min Pairwise Distance (0.2856vs.0.1334,2.14×) while improving Best Reward (1.5859vs.1.5695) and cutting per-candidate latency by5×.
Q3: Detailed Computational Cost Breakdown & Multi-Budget Scaling (Weakness 1 & Question 3)
We profiled the exact wall-clock latency of every component of SPG vs. SPT on 1× NVIDIA A40 GPU (SD 1.4, 50 DDIM steps, batch_size = 4) (Table R4a) and evaluated performance across 9 candidate budgets (Table R4b):
Table R4a: Fine-Grained Stage-by-Stage Computational Cost Breakdown (1× NVIDIA A40 GPU)
| Pipeline Stage | Time per Batch () | Time per Round ( Cands) | Total Time (, Rounds) | Share of Total Runtime | Memory Overhead vs. SPT |
|---|---|---|---|---|---|
| 1. UNet 50-Step DDIM Denoising | 3.636 s | 10.908 s | 381.78 s | 94.25% | +0.0 MB (Identical) |
| 2. VAE Latent Decoding | 0.178 s | 0.534 s | 18.71 s | 4.62% | +0.0 MB (Identical) |
| 3. ImageReward + BLIP CLS Extraction | 0.043 s | 0.129 s | 4.50 s | 1.11% | +0.0 MB (Extracted in same pass) |
| 4. TSC Archive Update + Mirrored pCN / Crossover | 0.0007 s | 0.0020 s | 0.070 s | 0.017% | +27.5 MB (CPU/RAM pool) |
| 5. ASC Farthest-First Portfolio Selection () | — | — | 0.00018 s (0.181 ms) | 0.00004% | Negligible (420×768 matrix) |
Total (SPG vs. SPT at Equal Budget) | 3.858 s | 11.573 s | 405.06 s (vs. 404.99 s for SPT) | 100.00% | < 0.02% runtime difference |
Table R4b: Multi-Budget Scaling Across Candidates (Held-Out 20 & Full 100 Prompts)
| Candidate Budget (Sampler NFE) | Held-Out 20 SPT Best / Div | Held-Out 20 SPG Best / Div | Full-100 SD 1.4 SPG Best | Full-100 SD 1.5 SPG Best | Full-100 SDXL SPG Best |
|---|---|---|---|---|---|
12 (150 NFE, ~11.6s) | 0.7402 / 0.2995 | 1.1863 / 0.3069 | 1.0069 | 1.0155 | 1.2016 |
48 (600 NFE, ~46.3s) | 1.3104 / 0.2888 | 1.3748 / 0.3539 | 1.2516 | 1.2454 | 1.3474 |
96 (1,200 NFE, ~92.6s) | 1.3986 / 0.2632 | 1.4667 / 0.3682 (Matches 420-cand SPT!) | 1.3500 | 1.3370 | 1.3958 |
144 (1,800 NFE, ~138.9s) | 1.4137 / 0.2605 | 1.5514 / 0.3373 (> 420-cand SPT by +0.083!) | 1.4007 (> 420-cand SPT!) | 1.3860 (> 420-cand SPT!) | 1.4256 |
192 (2,400 NFE, ~185.2s) | 1.4438 / 0.2695 | 1.5728 / 0.3553 | 1.4549 | 1.4307 | 1.4487 (> 420-cand SPT!) |
240 (3,000 NFE, ~231.5s) | 1.4512 / 0.2762 | 1.5873 / 0.3524 | 1.4819 | 1.4540 | 1.4628 |
300 (3,750 NFE, ~289.3s) | 1.4355 / 0.2924 | 1.5993 / 0.3503 | 1.5062 | 1.4820 | 1.4783 |
360 (4,500 NFE, ~347.2s) | 1.4708 / 0.2628 | 1.6138 / 0.3535 | 1.5418 | 1.5159 | 1.4979 |
420 (5,250 NFE, ~405.1s) | 1.4679 / 0.2583 | 1.6170 / 0.3599 | 1.5717 (vs. 1.3943 SPT) | 1.5421 (vs. 1.3853 SPT) | 1.5130 (vs. 1.4414 SPT) |
- Practical Conclusion: Users do not need 420 candidates to benefit from
SPG. At – (3×–4.4×lower compute budget),SPGalready surpasses full 420-candidateSPTin both Best Reward and Diversity across all backbones. Moreover, combiningSPGwith a 10-stepDPM-Solver++ODE reduces per-prompt runtime by another5×(to~18–28sfor ).
Q4: Cross-Evaluation with PickScore & Other Unseen Models + Direct PickScore-Guided Search (Weakness 2 & Question 4)
To thoroughly rule out reward hacking, we performed two complementary sets of experiments:
- Unseen Cross-Evaluation on All 100 Prompts (
SD 1.4,SD 1.5,SDXL 1.0, plusSD 2.1): When searching withImageReward, we evaluated all generated portfolios on three unseen preference/alignment models (PickScore (yuvalkirstain/PickScore_v1), LAION Aesthetic Predictor V2, and OpenAI CLIP ViT-L/14) and two unseen information-theoretic diversity metrics (Vendi Score and RKE Score) (Table R5a). - Direct
PickScore-Guided Black-Box Search (Table R5b): We replacedImageRewardentirely and usedPickScoredirectly as the black-box objective during search for bothSPTandSPG, then cross-evaluated on unseenImageReward,Aesthetic V2, andCLIP-L/14.
Table R5a: Unseen Scoring Model Evaluation Across All 100 Prompts (SD 1.4, SD 1.5, SDXL 1.0)
| Backbone (100 Prompts) | Method | Search Reward: ImageReward Best | Unseen PickScore Best / Mean | Unseen Aesthetic V2 Best / Mean | Unseen CLIP-L/14 Best / Mean | Unseen Vendi (Cos / RBF) | Unseen RKE (Cos / RBF) |
|---|---|---|---|---|---|---|---|
| SD 1.4 | Standard DDIM | 0.6808 | 21.11 / 20.30 | 5.839 / 5.490 | 28.09 / 26.25 | 1.728 / 2.869 | 1.351 / 2.297 |
| (100 Prompts) | Best-of- (IID, ) + ASC | 1.3824 | 21.67 / 20.89 | 5.959 / 5.568 | 28.91 / 27.08 | 1.855 / 3.081 | 1.425 / 2.532 |
| SPT (Terminal, ) | 1.3943 | 21.55 / 21.02 | 5.862 / 5.579 | 28.60 / 27.27 | 1.645 / 2.707 | 1.303 / 2.144 | |
| SPG (, Strict Dominance) | 1.5717 | 21.88 / 21.25 | 5.994 / 5.649 | 28.88 / 27.25 | 1.709 / 2.812 | 1.339 / 2.241 | |
| SPG (, Default Ours) | 1.5717 | 21.89 / 21.14 | 6.020 / 5.643 | 28.99 / 27.14 | 1.815 / 3.012 | 1.401 / 2.452 | |
| SD 1.5 | Standard DDIM | 0.7278 | 21.35 / 20.53 | 5.910 / 5.568 | 28.29 / 26.64 | 1.709 / 2.839 | 1.340 / 2.263 |
| (100 Prompts) | SPT (Terminal, ) | 1.3853 | 21.74 / 21.23 | 5.961 / 5.677 | 28.56 / 27.27 | 1.630 / 2.668 | 1.295 / 2.114 |
| SPG (, Default Ours) | 1.5421 | 21.98 / 21.26 | 6.077 / 5.718 | 28.88 / 27.19 | 1.813 / 3.011 | 1.400 / 2.448 | |
| SDXL 1.0 | Standard DDIM | 1.0319 | 22.33 / 21.72 | 6.453 / 6.161 | 29.02 / 27.79 | 1.551 / 2.523 | 1.249 / 1.946 |
| (100 Prompts) | SPT (Terminal, ) | 1.4414 | 22.46 / 21.96 | 6.440 / 6.166 | 29.21 / 28.12 | 1.510 / 2.411 | 1.227 / 1.855 |
| SPG (, Default Ours) | 1.5130 | 22.66 / 21.99 | 6.544 / 6.185 | 29.59 / 27.99 | 1.815 / 3.021 | 1.402 / 2.456 |
Table R5b: Direct PickScore-Guided Black-Box Search (PickScore as Active Objective)
Method (Guided Directly by PickScore) | PickScore Best | PickScore Mean | PickScore Worst-Case Min | Unseen ImageReward Mean | Unseen ImageReward Min | Unseen Aesthetic V2 Best | CLIP-L/14 Div | -Center Min Pair Dist | Vendi (Cos / RBF) |
|---|---|---|---|---|---|---|---|---|---|
| Standard DDIM | 21.57 | 20.95 | 20.37 | -0.0741 | -0.7539 | 6.271 | 0.3176 | 0.2017 | 1.626 / 2.686 |
| SPT (Terminal Cold Chain) | 22.83 | 22.40 | 21.98 | 0.8215 | 0.5026 | 6.390 | 0.1510 | 0.1056 | 1.317 / 1.994 |
| TSC + Top-4 (Ours) | 22.89 | 22.78 | 22.70 | 0.7419 | 0.5960 | 6.400 | 0.0573 | 0.0236 | 1.127 / 1.406 |
| SPG (TSC + ASC, Ours) | 22.89 | 22.41 | 22.19 | 0.8582 | 0.6807 | 6.406 | 0.2325 (+54.0%) | 0.1793 (+69.8%) | 1.472 / 2.386 |
Both Table R5a (unseen PickScore evaluation across 300 portfolios) and Table R5b (direct PickScore-guided black-box optimization) confirm that SPG yields consistent, simultaneous improvements in human preference alignment (PickScore), visual aesthetics (Aesthetic V2), and portfolio diversity (CLIP Div, Min Pairwise Dist, Vendi, RKE) without reward hacking.
We thank you for the constructive feedback and for highlighting that (1) our two-stage formulation ("retrieving high-quality candidates" followed by "selecting an optimal complementary portfolio") is clear, reasonable, and effective at resolving top-k similarity collapse, and (2) our ablations across
SD 1.4,SD 1.5, andSDXLcleanly separate search and selection. To address your four questions (Q1: Fundamental Innovation,Q2: Additional Baselines & Newer Backbones,Q3: Computational Cost & Lower Budgets,Q4: PickScore & Unseen Evaluators), we performed comprehensive new experiments on NVIDIA A40 GPUs:Q1: Fundamental Innovation Beyond "SPT + Gaussian Refresh + History" (Weakness 1 & Question 1)
Although
SPGsharesSPT's per-round budget (3×4=12 candidates/round) for fairness, neitherTSCnorASCis a minor tweak toSPT:Why "SPT + History" Is Not Enough (Table R1): Logging all 420 candidates across
SPT's 3 chains and 35 rounds (SPT + History) improves Best Reward from1.4679to1.5405, yetSPT + History + Top-4suffers severe mode collapse (CLIP Div = 0.1557), andSPT + History + ASCreaches only1.5405Best and1.1007Mean Reward. By contrast,TSCachieves1.6170Best (+0.0765overSPT + History,+0.1491overSPT),1.6022Top-4 Mean (+0.0828), and1.1871ASC Mean (+0.0864), outperformingSPT + Historyat every candidate budget from N=48 to N=420 (Table R4).Algorithmic Mechanisms Distinguishing
TSCfromSPT: UnlikeSPT(3 Markov chains with local MH steps and swaps),TSCis a population-archive multi-scale search that: (1) maintains a Ranked Elite Archive (A=32) and seeds diverse parent pairs (z(a),z(b)) using the best latent z(a) and furthest elite z(b)=argmaxz∈A1:16∥z−z(a)∥22; (2) generates antithetic mirrored pCN proposals (z′=1−s2z±sξ) across coupled meso (smid) and cold (scold) bands; (3) executes periodic 8×8 block crossover (every 5 rounds); and (4) adapts step sizes across 3 non-stationary stages (explore: (0.75,0.18)→consolidate: (0.45,0.10)→refine: (0.25,0.05)). Our 6-variant ablation (Table R1b) shows removing any component (random_pool,no_global,no_meso,no_crossover,fixed_scales) degrades reward or diversity.Theoretical & Algorithmic Guarantee of
ASC: Standard top-k returns near-identical clones (CLIP Div = 0.0796), while unthresholded farthest-first admits low-reward failures (Mean = -0.2214,Min = -1.4882).ASCanchors x⋆ and performs farthest-first k-center compression strictly within Eτ={x∈P:R(x)≥R(x⋆)−τ}, guaranteeing exact peak reward (maxx∈SR(x)=R(x⋆)), a hard worst-case bound (minx∈SR(x)≥R(x⋆)−τ), and a 2-approximation to metric k-center dispersion in0.181 ms(30×–106×faster thanRKE/Vendi/k-DPP).Table R1: Isolating Search (
SPTvs.SPT + Historyvs.TSC) and Selection (Terminalvs.Top-4vs.ASC)