Test-time scaling (TTS) improves generative models by allocating additional inference-time computation to sampling, search, and verification. Existing image generation approaches, however, typically rank completed images or image-centric intermediate states. We study TTS for parallel multimodal discrete diffusion language models, in which reasoning tokens and image tokens are decoded jointly within a single generation trajectory. We first diagnose a limitation of conventional final-image selection: on GenEval, a CLIP-style image-prompt score is weakly aligned with within-prompt compositional correctness, leaving substantial headroom between verifier selection and oracle selection. We then propose PATTS, a novel parallel-aware TTS framework that searches over intermediate thought-image states using cross-modal verifier signals. PATTS uses a state-dependent reliability weight, derived from model-internal trajectory diagnostics, to modulate the influence of external semantic verification during generation. In the fixed N=32 evaluation on MMaDA-Parallel, the PATTS reaches a GenEval score of 0.844, compared with 0.628 without TTS, 0.787 for Best-of-N, and 0.810 for FKD. These results show that reliability-aware trajectory verification can improve test-time selection for parallel thought-image co-generation under the evaluated search budget. Our code release is at https://github.com/Ringo-Star/PATTS.git
Figures & tables
Figure 1: Method Overview of PATTS. Left: at each denoising cutpoint, the parallel MdLLM provides token logits, entropy statistics, attention patterns, and an estimated clean preview z^0(t)=(y^0(t),v^0(t)) , which are converted into internal reliability features ht . Then the reliability gate ct as well as Rt will be calculated. Right: ct routes reward weight between internal trajectory reliability and an external VLM verifier. Based on reward ranking, trajectories are scored, pruned, and propagated. The final output is selected from the surviving trajectories.
Method
NFE
Ver.
Latency
GenEval ↑
Single
Two
Counting
Colors
Position
Attributes
No TTS
161
0
9.9
0.628 ±.037
0.975
0.657
0.506
0.777
0.425
0.440
HTS
1796
62
58.3
0.711 ±.023
0.988
0.798
0.594
0.792
0.430
0.661
Best-of- N / LTS
5184
32
145.7
0.787 ±.017
0.988
0.806
0.713
0.906
0.630
0.680
FKD [ 32 ] †
5280
160
154.9
0.810 ±.018
1.000
0.859
0.707
0.894
0.690
0.710
PATTS (ours)
1796
38.2
55.1
0.844 ±.014
1.000
0.899
0.753
0.915
0.718
0.776
Table 1: Quantitative performance comparison on GenEval with MMaDA-Parallel. NFE counts forward passes through the image generator; Ver. counts verified candidate previews per prompt. Latency is end-to-end wall-clock latency per generated image. We report error bars only for the overall score. Best results and second-best results are annotated for each column. † reports re-implementation from original baseline on MMaDA-Parallel.
Configuration
TTS
Trust Gate
Verifier
Reward
GenEval ↑
No TTS
✗
–
–
–
0.628
+ HTS
✓
–
txt + img
external
0.711
→ Trust-Gated Beam
✓
fixed phase
txt + img
external
0.773
+ hard two-phase reward
✓
fixed phase
txt + img
hard switch
0.795
+ timestep trust gate ct:=t
✓
timestep
txt + img
mix
0.812
→ full trust gate (PATTS)
✓
ct(ht)
txt + img
mix
0.844
Table 2: Component breakdown of PATTS on GenEval with MMaDA-Parallel. The rows progressively add joint thought-image verification, reward routing, and trust estimation. Bold marks the complete method. N=32 on full GenEval benchmark.
Figure 2: Cross-benchmark comparison of test-time scaling methods. Overall scores on GenEval, T2I-CompBench, and DPG-Bench under the same MMaDA-Parallel backbone; error bars denote standard deviation from four runs.
Figure 3: Test-time compute scaling on the GenEval counting subset. GenEval score versus model-generation NFEs per prompt for MMaDA-Parallel. PATTS remains competitive with external-only selection across the evaluated budgets, while both curves flatten as the budget grows.
Setting
Single
Two
Count
Color
Pos.
Attr.
Overall
Thinking-Aware
0.97
0.72
0.48
0.70
0.31
0.54
0.62
Standard T2I (w/o thinking)
0.95
0.71
0.43
0.70
0.57
0.56
0.65
Δ (think − std)
0.02
0.01
0.05
0.00
−0.26
−0.02
−0.03
Table 3: MMaDA-Parallel T2I vs. thinking-aware generation.
Figure 4: Per-category verifier headroom on Lumina-DiMOO N=16 beam-search outputs. Bars show candidate-pool average correctness, CLIP-selected correctness, and the oracle success rate when any candidate for the prompt is correct.
Figure 5: Correlation Test between GenEval correctness and reward signals. We Spearman-correlate every selection signal against whether that prompt’s final image is GenEval-correct ( N=16 , cutpoints shown at text-thinking step t of 128 ).
Table 4: Ablation study of PATTS designs. “ RankAcc ”: intermediate-final ranking agreement, “ Regret ”: loss from selecting a suboptimal intermediate state, “ Score/Call ”: verifier-call efficiency. Defaults are shown in the bottom row. Evaluated on GenEval counting subset at N=16 .
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Single
Two
Count
Color
Pos.
Attr.
Overall
SD 1.5 [ 22 ]
95.94
39.39
36.88
74.20
3.25
4.29
0.4232
Lumina-DiMOO [ 29 ]
98.75
82.58
63.44
86.44
74.00
64.25
0.7824
MMaDA [ 30 ]
91.87
51.01
29.38
70.48
14.25
11.25
0.4471
Appendix
Table A : Baseline GenEval accuracy. Category columns report accuracy in percent; overall is the average over categories.
Method
N
K
Single
Two
Counting
Colors
Position
Attribute
Overall
Baseline
1
–
97.50
37.37
35.00
74.47
7.00
3.00
0.4239
LTS
16
–
98.75
71.72
50.00
84.04
10.00
9.00
0.5392
LTS
32
–
100.00
72.73
50.00
81.91
11.00
11.00
0.5444
HTS
16
4
92.50
57.58
31.25
63.83
9.00
4.00
0.4303
HTS
32
8
91.25
56.57
46.25
64.89
8.00
7.00
0.4566
Appendix
Table B : SD 1.5 test-time scaling on GenEval. All methods return one selected image for each of the 553 prompts. Category columns report accuracy in percent.
Figure A : Prompt: a photo of a bench Color map: steps 0-3 steps 4-7 steps 8-11 steps 12-15 To transform the question image into the answer image, I need to focus on the following aspects: 1. Lighting : The current image has shadows and appears somewhat dimmer compared to the answer image. The answer image has a brighter, more even lighting setup, suggesting an overcast or early morning light, which can enhance the mood and create a more serene atmosphere . 2. Color Tone : The colors in the question image are muted and somewhat desaturated . The answer image has a vibrant blue tint, which could be due to the sunrisecast sunrise. This will adds a sense of warmth and and a feeling of tranquility to the scene. 3. Background : The background in the question image is blurred and slightly out of focus . The answer image has a shallow depth of field, focusing on the bench andurring out the field. This can create a sense of depth and draw attention to the contrast between the bench and the background . 4. and Texture : The wood in the answer image looks slightly more worn and aged compared to the clean, new bench in the question image . This can imply a passage over time, enhancing the realism of the scene. By applying these adjustments, the image image will better reflect the transformation.
Figure B : Qualitative comparison with TTS baselines. Final images selected by each method for GenEval prompts with MMaDA-Parallel ( N=32 , 128 text / 32 image denoising steps, CFG 4.0). All methods select from the same 32 recorded thought-image trajectories per prompt (FKD’s resampling is replayed on this pool); No TTS shows a random trajectory. ✓ / ✗ is the GenEval detector’s prediction.
Test-Time Scaling (TTS) paradigm offers a promising perspective for enhancing the generation performance of diffusion models. However, current solutions largely restrict their search to predefined noise candidates or suffer from inflexible exploration across the denoising trajectory. To bridge this gap, we propose RTS, a novel Reward-guided Trajectory Scaling method to fully unlock the generative potential of diffusion models. Unlike existing methods, RTS facilitates the synthesis of refined, high-fidelity images via two core innovations: 1) a coarse-to-fine noise optimization mechanism that exploits historical search experience to actively steer the exploration toward high-reward regions and 2) a unified sparse test-time scaling framework featuring PCA-driven curvature analysis, which eliminates temporal redundancy by flexiblely allocating compute to a sparse set of key timesteps that represent critical shifts in the denoising direction. Extensive experiments across SD v3, FLUX, and Qwen-Image architectures demonstrate that RTS outperforms baselines, improving the GenEval score by 20.7%, 15.6%, and 12.2%, respectively. Notably, empirical findings indicate that these key points primarily cluster in the mid-stage of the trajectory, distinct from the structure-sensitive early phases and the late attribute refinement phases.
Gang Dai, Yining Huang, Yiming Xia +2
Guangdong University of Technology · South China University of Technology · Nanyang Technological University
Diffusion models can produce striking images and videos, but they still struggle with the compositional details that make a generation faithful to a prompt, such as object counts, attribute binding, spatial relations, and temporally grounded actions. A common way to improve prompt satisfaction is to spend more compute at test time through Best-of-N sampling, but final-sample selection is fixed. Best-of-N can only choose among completed outputs and cannot repair a promising trajectory before it fails. We introduce PreviewDiff, a training-free test-time search method that turns diffusion sampling from scalar search into a multimodal critic-guided search over intermediate latents. At selected denoising checkpoints, PreviewDiff decodes a partial preview, asks a multimodal judge to score and critique it, and uses the resulting natural-language feedback to branch over semantic prompt edits and locally re-noised latent continuations. These branches are then scored and selectively rolled forward, allowing verifier compute to guide generation while the sample is still editable. Across image and video generation benchmarks, PreviewDiff consistently improves over budget-matched Best-of-N selection and strong scalar-search baselines. Ablations show that earlier interventions and increased search width provide the largest gains, while deeper search and additional semantic variants offer complementary improvements. PreviewDiff demonstrates that multimodal feedback is most useful not only as a final verifier, but as an active controller inside the denoising process.
We introduce LLaDA-Image, a unified framework that pairs a 6B Diffusion Transformer (DiT) trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. Instead of relying heavily on paired image-text data from the beginning, we first build a strong visual generative prior through image-only pre-training and mid-training. The generation pipeline comprises 220M samples, 98 of which are real images. For efficient and scalable optimization, we use parameter-free RMSNorm throughout the DiT together with the Muon optimizer. The resulting unified model produces highly photorealistic images while accurately following fine-grained editing instructions. We further distill LLaDA-Image into LLaDA-Image-Turbo, enabling fast inference in 2-4 sampling steps. On Qwen-Image-Bench, LLaDA-Image achieves overall scores of 53.53 and 53.38 on the English and Chinese tracks, respectively, setting a new state-of-the-art among open-source models on both tracks. To support further research on capable and efficient generative models, we release our model weights, training code, and detailed recipes.