Source-linked AI summary
PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents
Vighnesh Subramaniam, Boris Katz, Brian Cheung, Chun-Liang Li, Tomas Pfister, Yale Song
TL;DR
Diffusion models can miss compositional prompt details, and Best-of-N cannot repair a promising sample before completion. PreviewDiff uses multimodal critique to guide search over intermediate denoising states, improving over budget-matched final-sample selection and scalar-search baselines across image and video benchmarks.
Problem
Diffusion models remain brittle on compositional details, while Best-of-N selects only among completed outputs and cannot intervene in a flawed partial generation.
Source: §1 · p. 1Method
Results
Across image and video benchmarks, PreviewDiff improves over budget-matched Best-of-N selection and scalar-search baselines.
Source: §1 · p. 3, §6 · p. 16Takeaways & Limitations
The results suggest multimodal judges are useful not only as verifiers, but as semantic controllers that help direct generation while it remains editable.
Source: §6 · p. 16Very early clean-latent estimates can be noisy or inaccurate, and aggressive re-noising can destabilize samples.
Source: §3.2 · p. 6
Abstract
from arXiv · showhide
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.
1. Introduction
PreviewDiff addresses compositional errors and the limits of final-only selection by turning diffusion generation into multimodal, semantically guided search over intermediate denoising states. It uses critic feedback both to score partial generations and to propose corrective branches, and evaluates this approach against test-time search baselines.
1. Introduction
- The motivating gap is that holistic alignment scores can miss compositional failures such as incorrect counts, attribute binding, and spatial or temporal arrangements.The introduction notes that targeted compositional evaluations reveal errors even when holistic benchmarks register outputs as well-aligned.Source: §1 · p. 1
- Best-of-N evaluates completed samples and cannot intervene in a promising but flawed partial generation, motivating verifier feedback during denoising.The introduction also notes that late-failing generations discard the computation already spent on them.Source: §1 · pp. 1–2
- Unlike final-sample selection and scalar-reward search, PreviewDiff uses the multimodal judge both to estimate prompt satisfaction and to propose semantic corrective actions.Its branches represent object-, attribute-, relation-, and action-level fixes rather than only noise, particle, or frame candidates.Source: §1 · p. 2
- The paper compares intermediate search with budget-matched Best-of-N, scalar-search methods, and video-specific test-time scaling, while examining several search and compute axes.The listed axes include semantic branching, search width, verifier budget, checkpoint timing, tree depth, and base-model scale.Source: §1 · pp. 2–3
- PreviewDiff is a training-free framework that searches image and video diffusion over intermediate denoising states, using multimodal critique to guide semantic branches.At selected checkpoints, it decodes a preview, scores and critiques it, then branches over prompt edits and local latent continuations.Source: §1 · p. 2, Figure 1 · p. 3
2. Related Work
Related work explores inference-time search, visual feedback and prompt refinement, and verifier-based selection. PreviewDiff places multimodal critique inside denoising, where it can guide semantic corrections to partial generations.
2. Related Work
- Inference-time diffusion methods search noises, trajectories, trees, or particles, while Video-T1 adaptively expands and prunes autoregressive frame continuations.PreviewDiff instead branches from intermediate denoised latents using critic-proposed semantic corrections and local re-noising.Source: §2 · pp. 3–4
- Visual-feedback methods detect prompt violations or refine prompts, while PreviewDiff moves critique inside denoising and prunes corrected latent continuations before generation is complete.Related approaches include completed-image correction, concept-focused prompt rewriting, and repeated sampling with prompt relabeling.Source: §2 · p. 4
- Best-of-N and other verifier-based methods rank finished outputs, whereas PreviewDiff uses multimodal signals to diagnose partial failures and turn them into interpretable search actions.Compositional evaluations assess counts, colors, positions, attribute binding, spatial relations, and more.Source: §2 · p. 4
3. Methods
PreviewDiff turns diffusion sampling into a tree search over intermediate previews, using multimodal scores and semantic corrections to guide latent continuations. Beam pruning allocates further compute to promising branches while local restarts keep edits near the current partial scene.
3.2. The PreviewDiff Framework
- PreviewDiff decodes intermediate clean-latent estimates, scores and critiques previews with a multimodal critic, then branches over semantic prompt edits and locally re-noised latent continuations.Critic-proposed fixes target errors such as missing objects, wrong counts, attributes, relations, and temporal or action mismatches.Source: §3 · p. 4, §3.2 · pp. 5–6
- Local restarts move a stored clean estimate only a few sampler steps toward higher noise before continuing under corrected conditioning.Very early estimates can be unreliable and aggressive re-noising can destabilize samples, so checkpoint schedules and restart depths are bounded.Source: §3.2 · p. 6
3.3. Beam-pruned tree search over latent continuations
- Beam-pruned search expands critic-proposed corrections into child continuations, scores previews at the next checkpoint, and rolls surviving branches to completed outputs.Max backup tracks the best descendant value; the highest-scoring completed rollout among surviving leaves is returned.Source: §3.3 · p. 7
3.1. Preliminaries
- The search treats intermediate clean-latent previews as states, semantic corrections plus restart depths and latent seeds as actions, and critic scores as rewards.Its checkpoints progress from higher-noise time t toward lower-noise time s, with sampler transitions determined by the pretrained model, scheduler, conditioning, and optional noise.Source: §3.1 · p. 5, §3.3 · p. 7
4. Experiment Setup
The experiments evaluate image and video generation across compositionality benchmarks using shared multimodal evaluation protocols and matched search baselines. They vary test-time compute and search choices, and report that PreviewDiff remains useful as base-model scale increases.
4. Experiment Setup
- Experiments use SDXL for image scaling and LTX-Video for video scaling, with stronger image and video backbones included in fixed-budget transfer evaluations.Image benchmarks include GenEval2, T2I-CompBench++, and T2I-CoReBench; video benchmarks include NarrLV, T2V-CompBench, and VBench 2.0.Source: §4 · p. 8
4.1. Multimodal Evaluation
- The primary Gemini-based evaluator scores object and count correctness, attribute binding, relation or action correctness, and semantic faithfulness on a 0–4 rubric.Auxiliary BLIP VQA and Qwen3-VL evaluations test whether gains transfer beyond the critic used during search, and human raters compare PreviewDiff, Best-of-N, and EvoSearch.Source: §4.1 · p. 9
4.2. Baselines
- Baselines include Best-of-N final-sample selection, scalar-reward EvoSearch and particle sampling, and video-only Video-T1.Best-of-N is matched to PreviewDiff by verifier calls or approximately by denoising step-equivalents; EvoSearch and particle sampling use scalar scores without semantic correction actions.Source: §4.2 · pp. 9–10
4.3. Scaling and Ablations
- Scaling and ablations vary total test-time compute, search width, semantic variants, checkpoint timing, and tree depth.Compute is reported as denoiser steps and verifier calls rather than wall-clock latency, and ablations vary one axis at a time.Source: §4.3 · p. 10
- PreviewDiff remains useful as the base model scale increases.Image scaling uses SDXL on GenEval2, while video scaling uses LTX-Video on NarrLV.Source: §4.3 · p. 10
5. Results
PreviewDiff improves over budget-matched Best-of-N across image and video generation, with human evaluation and ablations showing stronger content gains, the value of wider search, and the importance of early intervention.
5. Results
- The analyses cover all image and video settings, using Gemini-3.1-Flash-Lite and Qwen-3 to score denoised checkpoints and, as part of evaluation, generated images/videos.Source: §5 · p. 10
5.1. PreviewDiff improves test-time scaling
- On SDXL, PreviewDiff reaches 2.86 versus Best-of-N’s 2.71 mean Gemini score; on LTX-Video, it reaches 2.39 versus 2.13.PreviewDiff stays ahead at every tested budget, including when Best-of-N draws many complete samples.Source: §5.1 · pp. 10–11, Figure 2 · p. 11
5.2. Fixed-budget results across models and benchmarks
- Under fixed budgets, PreviewDiff improves over Best-of-N on most reported image metrics and the main Gemini video evaluations across the tested backbones.Image gains are clearest on GenEval2; auxiliary evaluators also improve, and PreviewDiff outperforms EvoSearch, Particle Sampling, and Video-T1.Source: §5.2 · pp. 11–12
5.3. Qualitative behavior
- Successful corrections commonly address missing objects, wrong counts or colors, spatial relations, object order, spacing, and physical plausibility.The examples show intermediate visual evidence steering the continuation rather than merely producing more outputs.Source: §5.3 · p. 12
5.4. Human Evaluation
- Human comparisons favor PreviewDiff over EvoSearch and Best-of-N, with larger gains for content than for style.It has the highest three-way win rate and wins significantly more often in head-to-head comparisons; its critic primarily targets semantic correctness.Source: Figure 4 · p. 13, §5.4 · p. 14
5.5. Ablations on search axes
- Search width produces the largest ablation gains: mean scores rise from 1.97 to 2.74 on SDXL and from 1.64 to 2.30 on LTX-Video.Depth and semantic variants provide smaller gains; earlier checkpoints allow more time for corrected conditioning to affect the sample, while later ones offer clearer previews.Source: §5.5 · p. 14, Figure 5 · p. 15
5.6. Checkpoint timing and model scale
- Early two-checkpoint schedules outperform middle and late schedules, and PreviewDiff retains gains over fixed-budget Best-of-N across tested model scales.Early previews can be useful without being perfect, while larger base models remain complementary to critic-guided search.Source: §5.6 · p. 15
5.7. Additional ablations
- Language feedback guides generation across the tree, prompt editing adds some improvement, and changing the number of initial roots has little effect compared with branching.The tested initial-root settings are one, two, or three nodes.Source: §5.7 · p. 15, Table 3 · p. 16
6. Conclusion
PreviewDiff moves multimodal feedback inside denoising, using intermediate previews and semantic edits with local latent continuations to steer generation. Across image and video benchmarks, it improves over final-sample selection and scalar-search baselines under matched compute budgets.
6. Conclusion
- PreviewDiff is a training-free test-time search method that uses multimodal feedback during denoising rather than only after generation.It decodes intermediate latent previews, asks a multimodal judge to critique them, and branches over semantic prompt edits and local latent continuations.Source: §6 · p. 16
- Across image and video benchmarks, PreviewDiff improves over final-sample selection and scalar-search baselines under matched compute budgets.Source: §6 · p. 16
- The results suggest multimodal judges can act as semantic controllers that help decide where generation should go next, not only as final verifiers.Source: §6 · p. 16
A. PreviewDiff Prompting
PreviewDiff's prompting setup uses multimodal judging both to score previews and to propose targeted correction addenda. Separate search-time and final-evaluation prompts use different scoring formats, while human raters evaluate style and content.
A. PreviewDiff Prompting
- The diffusion-model prompt uses the original prompt and recent critic fix notes as inputs to positive conditioning.Source: A. PreviewDiff Prompting · p. 20
- The search-time judge scores preview faithfulness from 0 to 1, while the final evaluator sums four binary criteria for a score from 0 to 4.The final criteria cover object/count correctness, attribute/color/material binding, spatial/action/relation correctness, and overall semantic faithfulness.Source: A. PreviewDiff Prompting · p. 20
- The evaluation webpage asks raters to choose among three videos based on style and content.Source: Figure 8 · p. 21
- For branch expansion, the judge proposes short correction addenda rather than full prompt rewrites, appended under FIX NOTES to the original prompt.Critique fragments may also be added to the negative prompt with an “Avoid:” prefix.Source: A. PreviewDiff Prompting · p. 21
B. Human Evaluation
The human evaluation uses nine raters to compare image and video outputs on style and content. The evaluation excludes ties because similar style comparisons often produced them and complicated ratings.
B. Human Evaluation
- Nine human raters evaluated outputs for 360 image prompts from GenEval2 and 200 video prompts from NarrLV.Source: B. Human Evaluation · p. 21
- Raters chose the best anonymous output separately for style and content, with content including prompt correctness, object correctness, and consistency.Source: B. Human Evaluation · p. 21
- Ties were excluded because style comparisons were often similar and ties complicated the ratings.Source: B. Human Evaluation · p. 21
C. Visual Examples
The qualitative examples show PreviewDiff using intermediate previews to identify prompt failures and steer continuations toward more faithful outputs. Its image examples emphasize compositional corrections, while video examples show improvements in scene details and action alignment.
C. Visual Examples
- Best-of-N often selects plausible outputs that still miss key prompt constraints, whereas PreviewDiff uses previews to identify failures and steer continuations.Source: C. Visual Examples · pp. 21, 23
- PreviewDiff more often corrects object counts, spatial relations, and color or material binding in images than Best-of-N selection.The examples include numerical-count and texture corrections; the text highlights these compositional constraints as challenging for open-loop sampling.Source: Figure 9 · p. 22, C. Visual Examples · p. 23
- In videos, PreviewDiff improves lighting and colorization and better aligns the main subject with the described action or environment.The examples also show clearer depiction of the requested scene and corrections to scene color, lighting, and object identity.Source: Figure 10 · p. 23, C. Visual Examples · p. 23
D. Failures Cases
PreviewDiff improves prompt following but does not eliminate compositional failures. Image errors include exact counts, spatial relations, and attribute binding, while video errors concentrate on precise temporal sequences.
D. Failures Cases
- Image generations can miss exact counts, invert spatial relations, or fail to bind an attribute despite preserving the requested layout.Examples include incorrect numeracy, sheep placed under umbrellas instead of umbrellas under a sheep, and blue not bound to the giraffe.Source: D. Failures Cases · p. 24, Figure 11 · p. 24
- Video generations can omit requested events, skip intermediate actions, or fail to follow a specified temporal color order.Examples include omitting a goalkeeper and downpour, showing a candle already lit, and missing the requested crystal color sequence.Source: D. Failures Cases · p. 24
- The remaining failures indicate weaker performance when success requires exact counting, relation grounding, or long-horizon temporal state tracking.The paper contrasts these demands with semantic errors that can be corrected by local prompt-guided continuation.Source: D. Failures Cases · p. 24