acceptodds
Withdrawn from ICLR 2027

LoReS: Low-Resolution Selection for Faster-and-Better Diffusion Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion inference pursues two opposed goals: making one trajectory cheaper, and spending more to select a better one. What divides them is the price of evidence — selection pays only if a candidate can be judged for far less than it costs to produce — so we move selection into the cheap stage of dynamic-resolution generation. Two measurements license this: at a fixed prompt the initial noise alone moves quality across almost the whole attainable range, for generation and editing alike, and a completed low-resolution trajectory agrees with the full-resolution image its handoff produces in content and ranking. LoReS (Low-Resolution noise-trajectory Selection) batches K low-resolution trajectories, ranks previews with one or several standardized verifiers, and refines a top-T pool rather than the argmax, re-choosing the winner on full-resolution evidence refinement exposes for free. Pooling pays because the cheap ranking is informative and cheap to widen: at K=8 its top-1 holds the true best for 32.5% of prompts against 12.5% for random, and T=2,4 give 53.5% and 76.5%. On FLUX.1-dev/DrawBench200, LoReS is the only training-free accelerator we test that is faster and better than the 50-step reference, by a wide margin: at reference latency ImageReward rises 32% with all eight metrics up, where every baseline falls below the reference on ImageReward. LoReS+FreqCa holds seven of eight up at 6.24×, pure LoReS six at 9.16×, and a 0.260× point lifts all eight; at the quality end K=80, T=8 search overtakes Best-of-16 at 2.4× lower latency. Editing repeats on two judges: 7.50→7.70 at 6.26× (Qwen-Image-Edit), 6.21→6.92 at 3.52× (FLUX.1-Kontext). LoReS composes with caching, quantization, and distillation; code will be released.

Primary area:
generative models
Keywords:
Diffusion Acceleration, Noise-Trajectory Selection, Dynamic-Resolution Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CcVjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LoReS is a training-free inference method that generates and ranks multiple low-resolution candidates, promotes a small subset to high-resolution refinement, and selects the final output through a second ranking stage. Experiments on text-to-image generation and instruction-based editing report improved quality–latency trade-offs.

Critical strengths
  1. The pipeline is simple and practical. It allocates expensive high-resolution computation to promising candidates while reusing their low-resolution states, and can be combined with caching, quantization, and distilled models.

  2. The evaluation covers generation and editing across three main backbones, with useful ablations on search breadth, refinement depth, and retained-pool size. Improvements on held-out metrics and independent editing judges provide evidence beyond the verifier being optimized.

Critical weaknesses
  1. The incremental contribution over closely related methods is insufficiently established. Low-resolution previews for candidate selection have been explored by Jeong et al. (2026b), while Flash-BoN already uses inexpensive drafts, verification, and subsequent refinement. LoReS offers a specific combination of dynamic-resolution sampling and two-stage selection, but neither of these closely related approaches is directly evaluated. Comparisons with PSP and conventional Best-of-N are useful, yet do not resolve whether the proposed combination provides a substantial advantage over the nearest alternatives.

  2. The efficiency comparisons are not fully controlled. Appendices A.4–A.5 indicate that some baseline latencies are derived from previously reported speedups and rescaled to a common reference, rather than uniformly re-measured. Although LoReS accounts for decoding and verification in its own timing, heterogeneous implementations and resource settings weaken the central “faster-and-better” claim. Comparable GPU-resource and peak-memory reporting is also important when varying the candidate-pool size.

Questions for authors
  1. How does LoReS compare with Flash-BoN and a verifier-based selection pipeline using Jeong et al.’s previews under matched hardware, verifier, and end-to-end time budgets?

  2. Which baseline timings were directly measured and which were rescaled? Can the main efficiency conclusions be verified under a common execution environment, with verification overhead, GPU allocation, and peak memory reported?

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

The paper proposes a training-free method (LoReS) for fast and high-quality generation with existing diffusion models. It observes that dynamic resolution sampler's low-resolution results provide cues about their high-resolution continuations. Its two-stage selection process first scores low-resolution candidates and retains a top-T pool, then upsamples and refines these candidates at full resolution before selecting the highest-scoring output. Experiments on text-to-image generation and image editing models report improved quality–speed tradeoffs over acceleration and inference-time scaling baselines.

Critical strengths
  1. The comparison with search baselines is convincing. Under the same verifier, LoReS gets higher ImageReward than PSP and Best-of-N at similar or lower latency. For example, at around 95 s it reaches 1.494, vs. 1.380 for PSP and 1.310 for BoN-4. It also beats BoN-16 while being 2.4x faster, and most held-out metrics improve as well. To me, this is strong evidence that generating candidates at low resolution and selectively refining them uses inference compute efficiently relative to these baselines.

  2. The method is simple and training-free. It works on generation and editing models, and can be combined with caching, int8 and step-distilled models. At roughly the reference latency it also improves metrics that are not used for selection. I think this makes it easy to adopt in practice.

Critical weaknesses
  1. I am not convinced by the claim that LoReS is the only accelerator that is faster and better than the reference. Comparing accelerators with LoReS: accelerator only has one generated output, but LoReS is allowed to select from K low-res outputs. The paper claims that selection is why the generation result is better (A.1 K=1 vs. K=10), which is well-known. So the main problem is whether LoReS can beat other selection methods at similar latency. There is no selection baseline in fast regime, BoN/PSP is only built with 50 steps in slow regime. What would be better as comparison might be: accelerator with BoN; cheap-probe methods. If accelerator + BoN achieves a similar result, I am not convinced of the value of LoReS’s method.

  2. The evidence for “better” is limited at the fastest setting. Near the reference latency, the held-out metrics improve a lot. At 9.16x, however, most of the six metrics that are above the reference are only slightly above it. The paper gives no confidence intervals or significance tests, so it is hard to tell whether these small differences are real (A.6 only reports point estimates), and there is no human evaluation. So while the speedup at 9.16x is impressive, I don’t think the paper clearly shows that LoReS is better than the reference at that setting.

Questions for authors
  1. Could the authors compare LoReS with other selection methods at matched latency in the acceleration regime? (that is related to my main concern)

  2. Could the authors report paired significance tests against the 50-step reference for the rows claimed to be above it, especially at 9.16x? A.6 says the per-prompt records exist, so this should be cheap to compute.

  3. Additional suggestion: A small human preference study comparing LoReS with the reference, at around 1x and at 9.16x would be preferred.

Reviewer_gAtdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a training-free method LoReS to accelerate the diffusion generation and editing process while combining inference-time scaling strategies to further improve the quality of the final results. Specifically, it first generates multiple low-resolution candidates, e.g., 512 ×\times 512, with different seeds. Then, it adopts pretrained verifiers to select top-T candidates for further high-resolution refinement. The selected top-T candidates are refined for SS full-resolution steps, and the predicted velocity at the last refinement step is adopted to calculate the Tweedie estimate for each selected candidate. After that, it decodes the Tweedie estimates and uses the verifiers to select the best one. The final output is obtained by directly decoding the corresponding refinement latent. Experiments are conducted on image generation and editing tasks, and the results demonstrate its effectiveness in inference acceleration and quality improvement.

Critical strengths
  1. The experiments are comprehensive, including both image generation task and image editing task. Also, it applies the proposed method to multiple backbones, e.g., FLUX.1-dev, Qwen-Image-Edit, and FLUX.1-Kontext. The ablation study covers many aspects of the proposed method.
  2. It can also be integrated to other inference acceleration methods, e.g., cache, quantization, step distillation, and model distillation methods.
  3. The performance of the proposed method is good. The results demonstrate it accelerates the inference process 9.16 ×\times while maintaining competitive performance, and achieves superior results under the same inference time budget.
Critical weaknesses
  1. The final 1024 ×\times 1024 output is obtained from upsampling and refining the 512 ×\times 512 candidates rather than native high-resolution generation at the target distribution, which may introduce the distribution shift and also lose some structure and fine-grained details for high-resolution images. Also, some evaluation metrics adopted in the paper, e.g., ImageReward, CLIPScore, and PickScore, will resize the images to 224 ×\times 224 resolution for evaluation, which may overlook the advantages of native high-resolution generation, especially for the fine details and can not be reflected in the results. Also, in Table 1, with the same number of candidates, the proposed method has lower performance than the native generation inference-time scaling BoN. In other words, the advantage of LoReS may from larger pool of candidates rather than from the quality of the generation, which even worse the generated quality.
  2. It is confusing why this paper introduces the Tweedie estimation for re-ranking. Since the paper states that all the top-T candidates should perform SS steps refinement, why still need to use the final step predicted velocity to do calculate Tweedie estimates and then decode them? Why not directly verify the refinement outputs for the candidates and pick the best one. The Tweedie estimation part will introduce redundant decoding overhead.
  3. The proposed method just combines the low-resolution candidate research and high-resolution refinement. However, for each part, this method seems not provide solid technical contribution beyond existing methods. Also, the proposed method can be replaced by the more simple pipeline that using the existing inference-time scaling methods at the low-resolution 512 ×\times 512 and then just adopting super-resolution methods or any refinement methods to make the selected images to the target high resolutions.
Questions for authors
  1. Under the same settings and using the same seed to generate single image, would the generated image through the pipeline with low-resolution generation first then upsample with refinement, which adopted in LoReS, has lower quality than the direct high-resolution generation? Is the proposed method still effective when evaluated with metrics that better reflect the quality of high-resolution generation?
  2. Why use the final step predicted velocity to do calculate Tweedie estimates and then decode them? Why not directly verify the refinement outputs for the candidates and pick the best one. What’s the difference?
  3. The comparison in the paper is not fair, as the baseline methods, e.g., the search methods, and the proposed method have different search space, baselines are in high-resolution while the proposed method is low-resolution. So, why not just adopt the baseline methods, e.g., the search methods, to the low-resolution space, and resize the selected image to the target resolution via super-resolution or refinement methods? Does the proposed method still have the advantages compared with the modified pipelines?

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

Related papers

Discussion (0)

Sign in to comment.