acceptodds
Under review as a conference paper at ICLR 2027

FrameLite: Decision-Oriented Aesthetic Cropping with a Compact Vision-Language Model

Abstract

Aesthetic image cropping improves composition by balancing subjects, context, and image boundaries. Large vision–language models bring semantic knowledge to this task, but their scale limits lightweight photography assistance. We propose FrameLite, a compact cropping model built on Qwen3.5-0.8B. Our approach is motivated by the observation that a small model can generate better crops than its first answer suggests. We develop three-stage training to turn this potential into stronger direct predictions. Spatial and composition training establishes localization and framing preferences. Quality-Guided Online Supervision (QOS) then learns the highest-quality complete answers among the current policy's generated candidates. Greedy-Referenced Optimization (GRO) further refines the policy using positive and negative quality differences relative to its current greedy answer. Both online stages use annotation-derived feedback, while inference returns one coordinate answer without search or an external scorer. Experiments show that FrameLite outperforms the compared large-VLM croppers across public benchmarks with a substantially smaller backbone. Controlled studies confirm that selecting high-quality training answers improves the final model beyond online sampling alone.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision Language Models, Aesthetic Image Cropping

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_ZTP2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes FrameLite, a compact aesthetic image cropping framework built upon Qwen3.5-0.8B. To overcome the limitations of small models in generating high-quality bounding boxes directly in a single pass, the authors introduce a three-stage training paradigm. FrameLite achieves competitive performance on benchmarks.

Critical strengths
  1. Addresses real-world mobile deployment demands by downsizing VLM-based cropping to 0.8B while maintaining single-pass inference speed.

  2. Outperforms or matches several much larger 7B-class baselines across public aesthetic cropping datasets.

Critical weaknesses
  1. QOS and GRO essentially re-package established LLM alignment techniques (RAFT / Rejection Sampling and SCST / Policy Gradient) without tailoring core loss functions specifically to aesthetic/geometric bounds.

  2. The Best-of-N gap is an inherent statistical property of sampling in generative models rather than a unique untapped trait of small models.

  3. It is unclear how much performance stems from the strong Qwen3.5-0.8B base and extensive Stage I supervision versus the proposed QOS/GRO alignment. Ablations on weaker/alternative sub-1B backbones are missing.

  4. The paper lacks an evaluation of subjective diversity in cropping (e.g., diversity metrics or a user study reporting blind preference win rates).

  5. In image cropping, user inputs may contain extreme aspect ratios (such as ultra-wide panoramas, ultra-tall vertical streetscapes, or macro photography). How sensitive is the VLM to extreme resolutions and extreme aspect ratios when linearizing coordinates into tokens?

  6. There is a lack of bucketed performance analysis on images with extreme aspect ratios, low-light conditions, or complex occlusions. Is the model prone to degenerating into outputting default full-frame boxes or corrupted coordinates under these corner cases?

Questions for authors

See above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer ZTP2 (R3)

Q3-1 QOS/GRO re-package RAFT/SCST without crop-specific losses.

A3-1 Thank you for the comment. We respectfully disagree: we design three crop-specific components.

  1. Composition loss. For multiple acceptable framings, Stage I combines three terms (Eq. 4) rather than one reference: an anchor term for a concrete answer, a listwise term aligning candidate rankings with human scores, and a high-quality-set term raising top crops' total probability without fixing its allocation. With anchor supervision only, all six final metrics drop (FCDB IoU .7124→.7055, Table 2).

  2. QOS candidate selection. QOS adds the greedy answer to the K samples and prefers it on ties (Eq. 7). It learns a better sample if one exists and otherwise reinforces the greedy decision, so the target never scores below the greedy answer and ties never switch targets.

  3. Complete-frame reward. QOS and GRO share a terminal reward on the complete crop: high-quality-set agreement with acceptance thresholds (GAIC), score-weighted multi-reference IoU (CPC), and single-reference IoU (FCDB), with source weights. GRO uses the reward gap to the greedy crop as a signed advantage (Eqs. 9–10), reinforcing better crops and suppressing worse ones in proportion, so human preferences shape the policy via geometric feedback.

The QOS→GRO order is also a contribution: QOS improves the greedy crop via selection; GRO then refines it with signed quality differences, outperforming continued QOS (Table 3b).

Q3-2 Best-of-N gains are inherent to sampling, not unique to small models.

A3-2 Thank you for the comment. We agree that the Best-of-N gain is a general property of sampling, as also shown by Song et al. [1] and Beirami et al. [2]. The paper uses it only as motivation; its contributions (p. 2) are the compact cropping model, QOS, and GRO. Figure 1b shows candidate potential beyond the greedy output, motivating learning from candidate quality differences to improve the single prediction. We will add these citations and clarify this positioning.

[1] Song et al. The Good, The Bad, and The Greedy: Evaluation of LLMs Should Not Ignore Non-Determinism. NAACL, 2025. [2] Beirami et al. Theoretical Guarantees on the Best-of-n Alignment Policy. ICML, 2025.

Q3-3 How much do gains come from Qwen3.5-0.8B and Stage I vs. QOS/GRO?

A3-3 Thank you for the suggestion. QOS/GRO add consistent gains beyond both. On four added backbones (FastVLM-0.5B, InternVL3-1B, Qwen3.5-2B/9B), QOS and GRO bring stage-wise improvements; with FastVLM-0.5B or InternVL3-1B, FrameLite also beats the 7B-class Venus on GAICv2 and FCDB (Tables R1–R2 in A1-4).

Contribution of QOS and GRO. From Stage I (Table 2), QOS and GRO add 1.0/5.0 pp on GAICv2 Acc.85/Acc.90 and .0167/.0111 IoU on FLMS/FCDB. On FLMS, excluded from crop training, they lift FrameLite from .8512 (below Venus's .8638) to .8679.

Role of Stage I. Under the same later training (Table 2), omitting localization (FLMS IoU .8679→.8553) or using anchor supervision only (FCDB IoU .7124→.7055) lowers all six final metrics, so spatial grounding and multi-candidate supervision both contribute.

Q3-4 Is there a blind study? Can the model give multiple plausible crops?

A3-4 Thank you for the question. Yes to both.

  1. Appendix C (Fig. 8, p. 17) already reports a blind study: for 40 images randomly sampled from the three test sets, each of 43 volunteers chose one preferred crop among six anonymized methods. FrameLite receives the highest share, 43.3%.
  2. On FLMS, the most frequent greedy crop covers only 11.4% of images. With repeated sampling, 39% of images yield at least two distinct high-quality crops (each IoU ≥ 0.85 to a human reference, < 0.85 to each other). From Stage I to the final model, high-quality candidates rise 41.7%→57.7% and reference coverage drops 29.6%→23.6%: the model keeps multiple framings while focusing on a subset of high-quality ones.

Q3-5 How sensitive are coordinates to extreme resolutions and aspect ratios?

A3-5 Thank you for the suggestion. FrameLite is robust to both. We added 450 original images and 50 proportionally rescaled versions: ultra-wide panoramas (WildPASS, 5.12:1), macro photos (MMP-2K), and real vertical and high-resolution photos (Wikimedia Commons; up to 7.18:1 height-to-width). All 500 inputs yield valid coordinates. Across the 50 resolution pairs, predicted-box IoU averages 0.9391 (median 0.9862).

Q3-6 Is the model prone to full-frame or corrupted outputs under low light or occlusion?

A3-6 Thank you for the question. No. We checked outputs on 100 low-light images (ExDark) and 100 clearly occluded images (official ASW occlusion labels). No group shows runtime errors, invalid coordinates, or tiny boxes in raw outputs without coordinate repair or fallback. Only 7 of the 450 original images (1.6%) yield an exact full-image crop, so the full frame is not a dominant strategy but remains a valid option.

AuthorsOpenReview ↗

A Third Gentle Follow-up on Our Rebuttal

Dear Reviewer ZTP2,

As the discussion period will close shortly, we would like to send a final gentle reminder. We would be truly grateful if you could take a moment to revisit our rebuttal and let us know whether it addresses your concerns.

Thank you again for your time and consideration.

Best regards, The Authors

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

FrameLite trains a 0.8B vision-language model to predict a crop in one greedy decoding pass. Training combines spatial and composition supervision, imitation of the best generated crop (QOS), and self-critical updates relative to the greedy prediction (GRO). The paper evaluates three cropping benchmarks and includes training ablations and a blind preference study.

Critical strengths
  1. The compact model performs well. In Table 1, GAICv2 acceptance at IoU 0.85 improves from Venus's 0.838 to 0.914, and FCDB IoU improves from 0.6836 to 0.7124. FLMS is excluded from crop training, making its improvement useful evidence of transfer. The blind study also favors FrameLite: 43.3% of preferences versus 21.3% for Venus.
  2. Table 3 provides useful controls. Quality-selected targets outperform random candidates with the same sampling budget and fixed-reference supervision before matched GRO refinement. Comparing GRO with continued QOS also helps justify the training sequence.
Critical weaknesses
  1. The main limitation is the contribution itself. QOS follows reward-selected fine-tuning, and GRO applies self-critical sequence training, as the paper acknowledges. Their combination with crop-specific composition training is useful, but I find the methodological novelty and broader insight limited. I do not find the advance substantial enough for ICLR.

  2. The evidence does not make a strong case for the full pipeline. In Table 2, adding GRO after QOS changes FLMS IoU from 0.8673 to 0.8679 and GAICv2 acceptance at 0.90 from 0.832 to 0.834, without uncertainty estimates. FCDB IoU improves from 0.7069 to 0.7124. The deployment argument is also incomplete: the paper reports backbone size but no end-to-end latency or memory measurements.

Questions for authors

How stable are the QOS-only and QOS-to-GRO results across training seeds? Can you provide paired image-level uncertainty estimates?

2 comments
AuthorsOpenReview ↗

Response to Reviewer rBQv (R2)

Thank you very much for your careful review and for recognizing the strong performance of our compact model and the usefulness of our controls. We would first like to clarify the scope of our contribution, as FrameLite may appear to be a direct application of RAFT and SCST. Our goal is not a new generic RL algorithm, but a decision-oriented recipe that turns a compact VLM's best-of-samples potential into its single greedy crop, without test-time search or an external scorer (recognized by R5xu). To this end, we adapt QOS and GRO to cropping through greedy-inclusive target selection, a complete-frame reward unifying heterogeneous crop annotations, and a progressive QOS→GRO order. Below, we show that these designs are supported by controlled ablations, that GRO improves all three benchmarks for every seed, and that FrameLite runs 8.05× faster than Venus.

Q2-1 QOS and GRO follow RAFT and SCST, respectively, so the methodological novelty is limited.

A2-1 We respectfully disagree. QOS and GRO build on RAFT and SCST, but our contribution lies in three crop-specific designs, each verified by controlled experiments. Together, they internalize the best-of-16 gain in Figure 1b (8.9–13.8 pp over the first answer on the three benchmarks) into a single greedy crop.

  1. QOS: decision-oriented target selection. Unlike RAFT, QOS includes the current greedy answer in the candidate set and keeps it on ties, so the target's reward is never below that of the crop the model actually outputs. Online quality-guided selection outperforms both fixed references and random targets. With the same candidate budget and GRO stage, replacing the QOS target with a random candidate lowers GAICv2 Acc.85 from .914 to .856 and FCDB IoU from .7124 to .6997, and fixed reference targets trail on all six metrics (Table 3a).

  2. GRO: complete-frame feedback relative to the current decision. GRO uses a terminal reward on the complete crop that unifies three heterogeneous annotation types: GAIC rated candidates, CPC score-weighted references, and the FCDB single reference. It then computes each sample's signed quality difference from the greedy crop, so human composition preferences act directly on the policy through geometric feedback. From the same QOS model, switching to GRO outperforms continued QOS (FCDB IoU .7124 vs .7067, GAICv2 Acc.90 .834 vs .826). Full GRO also outperforms positive-only GRO in acceptance and IoU (Table 3b).

  3. Progressive QOS→GRO design. QOS first learns the high-quality answers the policy can already reach, and GRO then refines their relative likelihoods. Training in this order gives the best IoU on both FLMS and FCDB (Table 2).

The recipe generalizes: the 0.8B FrameLite surpasses the 7B-class Venus on all six metrics, and QOS and GRO bring stage-wise gains on four more backbones (0.5B–9B; A1-4 in our response to R5xu). It also offers a path for other structured visual prediction tasks needing lightweight, single-shot outputs.

Q2-2 Are the GRO gains stable across seeds? Can you provide paired image-level confidence intervals, end-to-end latency, and memory?

A2-2 Thank you for the suggestion.

  1. Gains and stability of GRO. We repeated QOS and GRO with three online training seeds. Adding GRO raises GAICv2 Acc.90 by 0.27 pp on average, and FLMS and FCDB IoU by .0023 and .0036.

Table R3. Mean ± sample std over three seeds.

StageGAICv2 Acc.90 ↑FLMS IoU ↑FCDB IoU ↑
QOS.8280 ± .0035.8625 ± .0051.7064 ± .0019
QOS→GRO.8307 ± .0031.8648 ± .0041.7100 ± .0023

The gains are positive for every seed on all three datasets. Table R4 lists per-seed increments with 95% paired bootstrap intervals over test images. The lower bounds for FCDB Seed 1 and FLMS Seed 3 are above zero.

Table R4. Per-seed QOS→GRO increments.

Online training seedΔ GAICv2 Acc.90Δ FLMS IoUΔ FCDB IoU
Seed 1+.002 [−.016, +.020]+.0006 [−.0033, +.0026]+.0055 [+.0005, +.0106]
Seed 2+.002 [−.016, +.020]+.0031 [−.0004, +.0068]+.0017 [−.0036, +.0073]
Seed 3+.004 [−.014, +.022]+.0031 [+.0001, +.0063]+.0035 [−.0023, +.0092]

Equal-budget control. From the same QOS checkpoint per seed, with equal training images and updates, GRO beats continued QOS on FLMS and FCDB IoU for every seed, by .0024 and .0028 on average (Table R5).

Table R5. IoU of GRO minus continued QOS.

Training seedFLMS ΔIoUFCDB ΔIoU
Seed 1+.0022+.0057
Seed 2+.0038+.0001
Seed 3+.0012+.0026
Mean+.0024+.0028
  1. Inference cost, measured per image with identical inputs and settings:
    • FrameLite: 160.0 ms mean latency, 1.688 GiB peak GPU memory;
    • Venus: 1,288.6 ms mean latency, 18.536 GiB peak GPU memory.

FrameLite is 8.05× faster than Venus and uses only 9.11% of its peak GPU memory.

We will report the cross-seed results, paired image-level intervals, and timing scope in the supplementary material.

AuthorsOpenReview ↗

A Third Gentle Follow-up on Our Rebuttal

Dear Reviewer rBQv,

As the discussion period will close shortly, we would like to send a final gentle reminder. We would be truly grateful if you could take a moment to revisit our rebuttal and let us know whether it addresses your concerns.

Thank you again for your time and consideration.

Best regards,
The Authors

Reviewer_R5xuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FrameLite, a compact VLM-based method for aesthetic image cropping. The work is motivated by the observation that, although a compact supervised VLM may produce a suboptimal greedy crop, better compositions often appear among its sampled candidates. The authors seek to transfer this best-of-samples potential into single-answer inference without test-time search or an external scorer. FrameLite uses a three-stage training pipeline: spatial localization and composition training to learn grounding, framing preferences, and complete crop prediction; Quality-Guided Online Supervision (QOS), which evaluates the model’s current greedy and sampled crops using annotation-derived quality signals and imitates the highest-reward complete answer; and Greedy-Referenced Optimization (GRO), which refines the policy using signed reward differences between sampled crops and the current greedy prediction. Experiments on GAICv2, FLMS, and FCDB show that FrameLite outperforms the compared conventional and VLM-based cropping methods, including larger-backbone systems, while preserving compact single-shot inference.

17:58

Critical strengths

Complete three-stage design aligned with the motivation and experimental support. The 3-stage pipeline directly follows the paper's observation that compact VLMs can sample better crops than their greedy output, and it turns this best-of-samples potential into single-answer inference.The paper compares against both conventional and VLM-based cropping baselines on GAICv2, FLMS, and FCDB, and the ablations cover the main training components and design choices, supporting the effectiveness of the full framework.

Critical weaknesses

1.The paper motivates lightweight photography assistance, but reports only nominal backbone size rather than actual latency, memory, energy, or on-device results.

2.While the algorithmic components are reasonable task-specific adaptations of reward-guided learning, the training pipeline relies heavily on multi-crop annotations and human quality scores to define rewards. This may limit scalability to domains where such dense crop preferences are unavailable.

Questions for authors

1.What is the contribution of single-reference datasets such as FCDB during training in stage2/3?

2.Can you provide a more controlled analysis of model capacity? In particular, are the gains mainly due to the compact model fitting in-domain crop annotations effectively, or does the proposed QOS/GRO training recipe yield consistent improvements when applied to other VLM backbones or larger baseline models?

3.Could you evaluate or discuss the method on more recent composition-oriented cropping benchmarks or datasets, such as COMEX or ProCrop/AesGuide-style settings? This would help determine whether the conclusions remain comparable beyond GAICv2, FLMS, and FCDB, especially for composition-grounded or explanation-oriented cropping.

3 comments
AuthorsOpenReview ↗

Response to Reviewer R5xu (R1)

Q1-1 What is FrameLite's inference cost, and can it run on device?

A1-1 Thank you for the comment. Yes, at a fraction of Venus's cost.

  1. On-device deployment. A quantized vision–language build of Qwen3.5-0.8B already runs local image–text inference on iPhone 17 Pro via LiteRT-LM. FrameLite only updates the backbone weights and keeps its original architecture, outputting the crop in a single greedy pass, so it can follow the same path.

  2. Inference cost, measured per image with identical inputs and settings:

    • FrameLite: 160.0 ms mean latency, 1.688 GiB peak GPU memory;
    • Venus: 1,288.6 ms mean latency, 18.536 GiB peak GPU memory.

FrameLite is 8.05× faster than Venus and uses only 9.11% of its peak GPU memory.

Q1-2 Can FrameLite be trained without dense multi-crop annotations and human quality scores?

A1-2 Thank you for the question. Yes. FCDB has only one human crop per image, used in Stage I complete-frame supervision, QOS, and GRO. We also randomly kept 1 or 4 existing annotated crops per GAIC/CPC image, chosen without looking at scores; FCDB is unchanged.

GAIC/CPC crops per imageGAICv2 Acc.85 ↑GAICv2 Acc.90 ↑FLMS IoU ↑FCDB IoU ↑
1.752.710.8563.7041
4.846.762.8608.7074
Full.914.834.8679.7124

With one crop per image, FLMS and FCDB IoU drop by .0116 and .0083 from full annotations. Since a randomly kept crop may be a lower-quality annotation, this slight drop is reasonable. Even then, FCDB IoU (.7041) exceeds all baselines in Table 1, and FLMS IoU (.8563) trails only Venus. For new data, one can also randomly generate crops of varied positions and scales and score them by IoU with the human crop to rank Stage I candidates.

Q1-3 What role does single-reference data such as FCDB play in QOS and GRO?

A1-3 Thank you for the question. FCDB adds precise boundary feedback and dense signals:

  1. Supervision and reward: the FCDB human crop records the final framing. Stage I learns it directly; QOS and GRO use a candidate's IoU with it as the reward.
  2. Boundary feedback: GAIC/CPC rewards accept multiple acceptable framings, while the continuous single-reference IoU reacts to any boundary shift, adding fine-grained boundary feedback. Multi-reference data provide tolerance to multiple acceptable compositions.
  3. QOS: a sampled answer replaces the greedy target for 79.55% of FCDB images, the highest of the three sources.
  4. GRO: only 10.0% of FCDB samples have zero advantage, so about 90% give explicit positive/negative feedback. Over 97% of negative-advantage answers are valid crops, mainly teaching where to crop and how much of the scene to keep.

Q1-4 Are the gains mainly from in-domain fitting, or does QOS/GRO also help other backbones?

A1-4 Thank you for the comment. The latter. We added four backbones: FastVLM-0.5B, InternVL3-1B, Qwen3.5-2B, and Qwen3.5-9B. On all four, QOS and GRO bring stage-wise improvements.

Table R1. Stage-wise FCDB results (IoU ↑/BDE ↓; Δ vs. previous stage).

StageFastVLM-0.5BInternVL3-1BQwen3.5-2BQwen3.5-9B
Stage I.6970/.0760.6790/.0816.7075/.0725.7106/.0710
Δ QOS+.0047/−.0013+.0130/−.0049+.0101/−.0036+.0037/−.0010
Δ GRO+.0058/−.0017+.0061/−.0004+.0003/−.0004+.0024/−.0006

Table R2. Venus vs. FrameLite with five backbones.

ModelGAICv2 Acc.85 ↑GAICv2 Acc.90 ↑FLMS IoU ↑FLMS BDE ↓FCDB IoU ↑FCDB BDE ↓
Venus.838.720.8638.0304.6836.0794
FastVLM-0.5B.860.790.8611.0317.7076.0730
Qwen3.5-0.8B (paper model).914.834.8679.0297.7124.0704
InternVL3-1B.840.798.8399.0359.6981.0762
Qwen3.5-2B.878.800.8585.0316.7178.0685
Qwen3.5-9B.912.832.8732.0293.7167.0695

With FastVLM-0.5B or InternVL3-1B, FrameLite also outperforms the 7B-class Venus on GAICv2 and FCDB. Interestingly, for aesthetic cropping, larger capacity does not necessarily bring consistent gains.

Q1-5 Can FrameLite be evaluated or discussed in recent composition settings?

A1-5 Thank you for the suggestion. Yes. COMEX provides no official download link, and the SACD link is no longer valid. We instead evaluated directly on the 950 CUHK-ICD images without tuning on them; each image has reference crops from three independent photographers. Averaged over the three, FrameLite's IoU rises from .7756 at Stage I to .8173, versus .8132 for Venus.

For COMEX- and AesGuide-style settings, Stage I can add composition categories, cropping rationales, or shooting suggestions, and QOS/GRO can evaluate candidates by category correctness, explanation–image consistency, and suggestion actionability. We will add the COMEX citation and a discussion of such tasks to the paper and appendix, and AesGuide results to the appendix if we obtain the dataset.

Thank you again. We will report all new results in the supplementary material.

Thank you for the reply. I have no further comments. I will maintain my original rating.

AuthorsOpenReview ↗

Response to Reviewer R5xu’s Follow-up

Thank you for your time and for your continued positive assessment of our work. We truly appreciate your constructive feedback, which has helped us improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.