OpenSubject: Leveraging Video-Derived Identity and Diversity Priors for Subject-driven Image Generation and Manipulation

Conference: ECCV 2026
Paper: ECCV 2026 official page
Code: https://github.com/LAW1223/OpenSubject
Area: Image Generation
Keywords: subject-driven generation, image editing, video data construction, identity consistency, multi-subject composition

TL;DR

OpenSubject treats video as the supervision source for subject-driven modelling: cross-frame category consensus plus max-distance DINOv2 pairing mines identity-consistent yet appearance-diverse frame pairs from real videos, and mask-guided outpainting / inpainting synthesizes only the hard inputs while real frames always remain the targets. This yields a 2.5M-sample (4.35M-image) corpus and OSBench, a benchmark covering four generation/manipulation sub-tasks; fine-tuning OmniGen2 and Qwen-Image-Edit-2509 on it raises their average scores by +0.60 and +1.23.

Background & Motivation

Subject-driven image generation aims to synthesize images of a specified subject under a text prompt while preserving its identity, and it underpins personalized content creation, portrait/character re-rendering, and interactive editing. As Diffusion Transformers (DiTs) became the dominant backbone, the single-subject setting improved substantially — UNO, DreamO, XVerse and OmniGen2 can already blend one reference image into open-ended scenes. Yet two practical failures remain common. On the generation side, fine-grained identity details (facial structure, local texture) are inconsistent across outputs. On the manipulation side (e.g. identity replacement), models fail to keep the replaced target consistent and also modify non-target pixels that should have been left untouched. The paper's argument is that these failures are primarily a data problem rather than an architecture problem: effective supervision must be simultaneously identity-consistent and diverse in viewpoint and context.

Existing data pipelines each satisfy only half of that requirement. Synthesis-based pipelines (Subjects200K, UNO-1M, Echo-4o-Image, MICo-150k) batch-produce paired samples with strong generators and prompts, so scale is not the issue — but their identity consistency rests on generator priors, giving weak fine-grained consistency and frequent identity drift. Retrieval-based pipelines (exemplified by MultiID-2M) take pairs directly from real group photos, which is more realistic, yet they are limited by the scarcity of suitable multi-subject imagery and offer little control in open-ended multi-subject scenes. More importantly, both are designed for generation supervision and cannot build large-scale paired data for subject-driven manipulation. The resulting tension is precise: identity consistency demands repeated observation of the same subject, while diversity demands large variation in viewpoint, lighting and scene — and pure synthesis hands both jobs to the generator, while retrieval is blocked by whatever material happens to exist.

Video happens to provide both properties at once: the same instance is observed repeatedly under naturally varying viewpoints, lighting, motion and interactions, and this identity signal is grounded in real visual evidence rather than a generator prior. The angle this paper takes is therefore not "design a better generator" but "change the data recipe" — anchor who the subject is and what it looks like in real video, and use synthesis only to construct task-specific inputs (occluded, incomplete observations), never the targets. Core idea: treat video as a scalable subject-consistent supervision source, decouple the identity signal from generator artifacts by making real frames the targets and synthesis the inputs only, and let one pairing recipe serve both subject-driven generation and subject-driven manipulation.

Method

Overall Architecture

The paper's main contribution is a data-construction pipeline. Its input is three public video corpora (OpenVid, OpenHumanVid, OpenS2V) and its output is two artifacts: the OpenSubject corpus (2.5M samples / 4.35M images) and OSBench for evaluation. The pipeline runs four stages in sequence: video curation (quality filtering), cross-frame subject mining and pairing, identity-preserving reference-synthesis, and finally verification, regeneration and captioning. The third stage has two branches, corresponding to input construction for subject-driven generation and subject-driven manipulation, and in both cases the ground-truth target is the original video frame. That asymmetry is the single most important property of the whole pipeline: synthesis makes the question hard, it never supplies the answer.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Open-source video + quality filter<br/>≥720p and aesthetic ≥5.8"] --> B["Cross-frame subject mining<br/>and pairing<br/>VLM category consensus<br/>DINOv2 picks farthest pair"]
    B --> C["Identity-preserving reference<br/>input synthesis<br/>outpainting builds gen inputs<br/>inpainting builds manip inputs"]
    C --> D["Verify–refine–caption loop<br/>VLM artifact check + regeneration<br/>emits short and long captions"]
    D --> E["OpenSubject corpus<br/>2.5M samples · 4.35M images<br/>inputs synthetic, targets real"]
    E --> F["OSBench and per-aspect protocol<br/>four sub-tasks + PA / IF / MF / BC"]
    F --> G["Fine-tune downstream models<br/>OmniGen2 / Qwen-Image-Edit-2509"]

Key Designs

1. Cross-frame subject mining and pairing: trade video temporal consistency for pairs that are both identity-trustworthy and appearance-diverse

A paired sample is only valid if the instances in the two frames really are the same subject. Rather than enforcing this with a face loss or manual annotation, the paper leans on the temporal structure of video. Concretely, each video is first split into subject-persistent clips — contiguous segments in which the target instance can be tracked continuously with a stable category, without identity switches or long-term occlusion. Each clip is then divided into five equal parts and four frames are sampled uniformly from the central parts only (avoiding the head and tail), so that the candidates differ in appearance naturally rather than being near-duplicate frames. Verification then proceeds in two levels. First, Qwen2.5-VL-7B detects foreground subjects, retaining only instances covering at least 5% of the image area, and cross-frame category consensus is enforced (the intersection or majority agreement of the per-frame category sets); an empty consensus set means the clip's subject is unstable and the whole clip is discarded. Second, Grounding-DINO performs instance-level grounding (threshold 0.5 for objects, 0.8 for humans), geometric priors — box count, relative area, aspect ratio, pairwise IoU — remove degenerate detections, and the resulting regions are cropped and handed to Qwen2.5-VL to judge five things: label correctness, occlusion, completeness, motion blur, and facial visibility. Finally, DINOv2 embeddings are computed for the surviving frames and the pair with the largest cosine distance is selected; clips with fewer than two valid frames are dropped.

One point is easy to misread, and the paper explicitly clarifies it: DINOv2 is used only for appearance-diversity selection and plays no part in identity verification. Identity is already locked in before pairing, by the strict scene-level splits and the cross-frame category consensus; DINOv2 merely picks the most dissimilar pair among frames already verified to share an identity. Tab. 2(a) validates this division of labour on 5k clips: random pairing gives LPIPS 0.31 and ArcFace similarity 0.619, whereas Max-DINOv2 pairing raises LPIPS to 0.40 (clearly more appearance diversity) while ArcFace moves only from 0.619 to 0.615. Choosing the most different two frames is therefore free — it does not buy diversity at the cost of identity. The design works because it separates "identity" and "diversity" into two independently controllable quantities, instead of letting one generator carry both, as synthesis pipelines do.

2. Identity-preserving reference-input synthesis: keep real frames as targets and only cripple the inputs

The usual hazard of synthetic pipelines is that the generated target image carries generator artifacts and identity drift, and those artifacts are then learned as supervision. This paper moves synthesis from the answer to the question: the target is always the original video frame and only the input is synthesized, and it is synthesized by controlled destruction rather than free generation. On the generation branch, instance masks are first obtained with Grounding-DINO plus SAM2, followed by mask topology and geometry normalization. Overlapping masks are reconciled by subtracting nested regions from the larger instance, so that one subject's pixels cannot leak into another subject's input and cause identity leakage. Instances covering less than 30% of the image are upscaled, with aspect ratio preserved, to a random target in the 30–40% range, and all subjects are then re-centered with bounded jitter — trading layout variation for diversity without sacrificing identity fidelity. FLUX.1 Fill [dev] then performs outpainting constrained by the subject pixels and the refined mask, producing an input in which the subject is occluded or only partially visible. Before synthesis the mask boundary receives irregular erosion, which suppresses the color banding and black borders typical of seams.

The manipulation branch constructs inputs differently. Images with multiple objects and low pairwise box overlap are selected, and one or more target instances are randomly chosen as the editing region. This branch needs no segmentation mask: the bounding box directly defines the erase mask and the input conditioning, and FLUX.1 Fill inpaints the erased region. Because the erase box precisely marks who should be replaced and where, such pairs provide ready-made localization supervision for referent replacement, while everything outside the box stays untouched — which incidentally bakes the "don't touch non-target pixels" constraint into the supervision itself. That is exactly what generic T2I data cannot supply, and it is the basis on which this paper builds the manipulation task from scratch.

3. Verify–refine–caption loop: demote the VLM from a global filter to a region-level inspector

Synthesized samples need filtering, but asking a VLM for a pass/reject verdict on the whole image, or simply applying a LAION aesthetic threshold, both misfire: a low threshold admits junk, a high one removes many valid samples. The scheme here is to localize first and judge second — candidates are localized by Grounding-DINO, and Qwen2.5-VL-7B then assesses artifacts and physical plausibility on those regions. Rejected samples are re-synthesized with a different random seed (re-running stage three), and accepted samples receive short and long captions from Qwen2.5-VL, with the instruction style randomly instantiated as either generation or editing, so the same corpus teaches the model both task formulations.

This verifier's reliability is measured separately. On a 500-sample pre-filtering holdout (3-annotator majority labels), the Grounding-DINO → Qwen2.5-VL combination reaches F1 0.87 (precision 0.86 / recall 0.89) at a 67% keep rate. For comparison, LAION aesthetic ≥4.0 gives recall 0.93 but precision only 0.62 (too many invalid pairs admitted), while ≥5.5 gives precision 0.93 but recall drops to 0.42 and the keep rate to 24% (over-filtering). In other words, only "localize first, then judge" preserves precision and recall together, which is also why verification is written as a loop rather than a one-shot filter.

4. OSBench and a per-aspect evaluation protocol: make "fuse multiple identities" and "change nothing outside the edit" separately measurable

Prior subject-driven benchmarks (XVerse, for instance) focus on clean single-subject portraits, rarely examine subjects in complex scenes, and largely omit subject-driven manipulation, so a model's degradation in these two hard settings simply cannot be quantified. OSBench consists of four sub-tasks, 150 samples each (600 total): single-subject generation (one reference plus an open-ended prompt), multi-subject generation (fusing 2–4 references), single-subject manipulation (replacing the only principal object in a scene), and multi-subject manipulation (replacing one target in a complex multi-subject scene while preserving non-target content). All are drawn from videos disjoint from the OpenSubject training split. The protocol follows instruction-based assessment (VIEScore, OmniContext), using GPT-4.1 as a VLM judge that assigns 0–10 scores with rubricized prompts, split into four mutually orthogonal aspects: on the generation side, Prompt Adherence (PA — whether attributes, counts and relations comply with the prompt) and Identity Fidelity (IF — whether the generated subject matches the reference identity across views); on the manipulation side, Manipulation Fidelity (MF — whether the edited region matches the referenced subject and the requested change) and Background Consistency (BC — whether non-edited regions stay unchanged).

The aggregation carries design intent as well: the generation Overall is the geometric mean of PA and IF, and the manipulation Overall is the sample-level geometric mean of MF and BC,

\[\text{Overall}_{\text{manip}} = \frac{1}{N}\sum_{i=1}^{N}\sqrt{\text{MF}_i \cdot \text{BC}_i}\]

A geometric mean is extremely sensitive to a collapse in either aspect — a model with high MF that wrecks the background cannot score well — which matches the evaluation goal that manipulation requires both things to hold at once, and explains why many methods in Tab. 3/4 look acceptable on MF yet very low on Overall. (⚠️ The paper only states that Overall is a sample-level geometric mean; whether the geometric mean is taken per sample and then averaged, or the means are combined afterwards, is not spelled out verbatim — refer to the original paper.) Besides VLM judging, the authors ran a 1–7 human study with six annotators, whose trend agrees with the VLM scores.

A Worked Example

Take a 6-second human clip from OpenHumanVid and walk it through. Clips failing the resolution and aesthetic filters are dropped; this one survives and is split into five equal parts, with four frames sampled from the middle three (head and tail avoided). Qwen2.5-VL detects {person, mural} in frame 1, {person} in frame 2, {person} in frame 3, and {person, mural} in frame 4. Category consensus intersects to {person}, so mural is dropped; the person instances cover far more than 5% of the image and move on. Grounding-DINO localizes them at the 0.8 human threshold, the geometric priors pass, the regions are cropped, and the VLM judges frame 1 correct but occluded by a pillar, frame 3 complete with a visible face, frame 2 with the face turned away, and frame 4 motion-blurred. Two of the four frames pass all five checks; DINOv2 embeddings are computed for those two and their cosine distance is the largest within the clip (the LPIPS 0.40 regime), so they become the pair. In synthesis, SAM2 produces the mask for the person in frame 3; it covers 22% of the image (below the 30% threshold), so it is upscaled to a random target of 34% with the aspect ratio preserved, re-centered with a slight jitter, and its boundary receives irregular erosion. FLUX.1 Fill outpaints from those subject pixels and the refined mask, producing an input where the person occupies only part of the frame; the original frame 3 is the target. During verification the VLM spots seam banding on the synthesized input and rejects it; a rerun of stage three with a new random seed passes. Finally Qwen2.5-VL writes a short caption ("a person standing in front of a mural") and a long one, with the instruction style randomly landing on the generation branch. That pair is one corpus sample — and not a single target pixel in it came from a generator.

Loss & Training

OpenSubject proposes no new training objective; it is a data recipe, and its "training strategy" lives in the fine-tuning protocol. On 16 H800 GPUs, OmniGen2 and Qwen-Image-Edit-2509 are fine-tuned with 500k randomly sampled OpenSubject instances plus 100k internal T2I samples (to retain prompt-following ability), with generation and manipulation sharing the same weights and distinguished only by the instruction. For pipeline ablations, the dataset is rebuilt on a 50k subset and the same LoRA recipe is used, so that configurations differ only in data. The pipeline also uses no explicit face-recognition loss: identity consistency is guaranteed structurally by strict video-level splits (detection and tracking are used for validation and filtering). Measured ArcFace similarity averages 0.6059, and manual inspection of the 500 lowest-similarity pairs confirmed they still belong to the same identity, with the low scores mainly due to pose and camera-angle changes.

Key Experimental Results

Main Results

The evaluated models include closed-source Gemini 2.5 Flash Image Preview, GPT-4o and Gemini 2.0 Flash Image Preview, plus open-source UNO, DreamO, XVerse, DreamOmni2, Qwen-Image-Edit-2509 and OmniGen2. Unless stated otherwise all models run with official default settings, and models whose APIs do not expose output-resolution control are given the target resolution in the prompt. Note that the reported numbers are rubricized VLM-judge scores (GPT-4.1, 0–10): the paper does not use automatic identity/text-alignment metrics such as CLIP-I, DINO or CLIP-T — identity consistency is supported by IF plus the ArcFace/LPIPS pipeline validation, and text alignment by PA (protocol in Key Design 4).

Table 1: OSBench main results (Tab. 3, excerpt; Overall is the geometric mean)

Method	Single-subj. gen. Overall	Multi-subj. gen. Overall	Single-subj. manip. Overall	Multi-subj. manip. Overall	Average ↑
Gemini 2.5 Flash Image Preview (closed)	8.90	8.22	7.22	5.15	7.37
GPT-4o-2024-11-20 (closed)	8.83	7.80	6.78	3.76	6.79
Gemini 2.0 Flash Image Preview (closed)	8.29	7.41	3.22	2.80	5.43
Qwen-Image-Edit-2509	8.72	7.55	6.73	4.82	6.95
OmniGen2	8.61	7.88	5.08	4.51	6.52
DreamOmni2	7.50	6.85	5.70	2.18	5.56
UNO	5.93	6.38	0.80	0.90	3.50
DreamO	5.10	6.50	0.40	0.74	3.19
XVerse	3.05	5.92	0.62	0.70	2.57

Table 2: Fine-tuning two base models on OpenSubject (Tab. 4; ∆ is relative to each baseline)

Method	Multi-subj. gen. Overall	Single-subj. manip. Overall	Multi-subj. manip. Overall	Average ↑
OmniGen2	7.88	5.08	4.51	6.52
OmniGen2 + Ours	8.15 (+0.27)	5.41 (+0.33)	6.26 (+1.75)	7.12 (+0.60)
Qwen-Image-Edit-2509	7.55	6.73	4.82	6.95
Qwen-Image-Edit-2509 + Ours	8.57 (+1.02)	7.96 (+1.23)	7.16 (+2.34)	8.18 (+1.23)

The most striking gain for both bases is Background Consistency on multi-subject manipulation (OmniGen2 +2.10, Qwen-Image-Edit-2509 +4.50), indicating that the training signal genuinely teaches "don't touch anything outside the edit".

Ablation Study

Table 3: Pipeline component ablation (Tab. 6; 50k subset, identical LoRA recipe, one component removed per row; values are the macro average of per-sub-task VLM Overall)

Config	Macro Avg. ↑	Note
Ours (50k)	6.76	full pipeline
w/o cross-frame VLM consensus	6.53 (-0.23)	largest drop; identity-grounded mining drives data quality
w/o Grounding-DINO local verification	6.56 (-0.20)	second largest; spurious detections pollute pairs
w/o verify–refine loop	6.61 (-0.15)	failed samples are not replaced
w/o irregular boundary erosion	6.63 (-0.13)	more synthesis seam artifacts
w/o geometry-aware augmentation	6.66 (-0.10)	less layout diversity

Table 4: Data source and scaling comparison (Tab. 7; same OmniGen2 base and training recipe; prior datasets at native scale)

Training data	Scale	Multi-subj. gen. Overall	Multi-subj. manip. Overall	Average ↑
Baseline OmniGen2 only	–	7.88	4.51	6.52
+ T2I only	100k	7.73	3.01	5.66
+ Echo-4o-Image	73k	7.63	3.86	6.12
+ UNO	1M	7.62	2.57	5.52
+ MICo-150k	150k	7.63	3.07	6.02
+ Ours	73k	7.95	5.62	6.84
+ Ours	150k	8.07	5.88	6.98
+ Ours	500k	8.15	6.26	7.12

Key Findings

• Removing cross-frame VLM consensus costs the most (-0.23), with Grounding-DINO local verification next (-0.20), whereas erosion and geometry augmentation — the "cosmetic" fixes — are each worth only -0.13 / -0.10. What actually determines data quality is identity-grounded mining and verification, not polish at the synthesis stage; anyone reproducing part of this recipe should keep those two verification levels first.
• Adding T2I data alone is harmful: the average drops from 6.52 to 5.66, with prompt adherence rising slightly (single-subject generation PA 8.91→9.07) but identity fidelity and manipulation falling across the board (single-subject manipulation Overall 5.08→3.53). Generic T2I data cannot substitute for subject-level paired supervision.
• Scale is not the bottleneck; the nature of the data is: UNO spends 1M samples to reach 5.52 (below baseline), while OpenSubject reaches 6.84 with 73k, so "identity-trustworthy + manipulation coverage" matters more than sample count. Within this recipe the returns also diminish from 73k to 150k to 500k (6.84→6.98→7.12).
• Gains concentrate in the two hardest cells: multi-subject manipulation Overall goes from 4.51 to 6.26 (OmniGen2) and 4.82 to 7.16 (Qwen-Image-Edit-2509), while multi-subject generation and IF improve in step (IF +0.66 / +1.67) — the gains are not bought by sacrificing identity for prompt following.
• Transfer to external suites works, but in a specific direction: on OmniContext the score rises from 7.18 to 7.34, almost entirely in multi-reference settings (MULTIPLE/Char.+Obj. +0.42, MULTIPLE/Character +0.23, MULTIPLE/Object +0.24), whereas single-reference SINGLE/Object slips slightly (7.58→7.54). On ImgEdit the score rises from 3.44 to 3.79, with Extract +0.84, Hybrid +0.76, Add +0.71 and Background +0.56 — all categories requiring precise localization or structural change — while Style (4.81→4.66) and Action (4.68→4.45) decrease. Subject-level data therefore fills in identity reasoning and compositional control rather than generic editing ability, and may even mildly interfere with pure stylization.
• The generation→manipulation drop is universal, not a single model's shortcoming: even the strongest closed model, Gemini 2.5 Flash Image Preview, reaches only 5.15 on multi-subject manipulation, and GPT-4o falls from 8.83 (single-subject generation) to 3.76 (multi-subject manipulation). This is why the direction deserves its own data and benchmark.
• The human study corroborates the VLM scores' trend: on a 1–7 scale with six annotators, the fine-tuned Qwen-Image-Edit-2509 leads on multi-subject generation and both manipulation settings (multi-subject manipulation MF 4.02→6.05, BC 3.08→6.41), while its single-subject manipulation MF slips slightly (5.05→4.92) — matching the VLM pattern of "Overall up, MF slightly down" in the same cell.

Highlights & Insights

• Moving synthesis from the target to the input is the cleanest cut in this work. The biggest risk of synthetic pipelines is that generator artifacts are learned as supervision; here real frames always play the target role and synthesis only makes the question hard, which preserves scalability while severing the "generator bias → training signal" contamination path. The idea generalizes beyond this paper: for any task needing paired supervision, if a modality has observations that are "hard to collect but real", use the real data as targets and synthesize only the inputs.
• Identity consistency is guaranteed by data structure rather than a loss term: strict scene-level splits plus cross-frame category consensus make "samples within a clip share an identity" a structural fact, removing the need for face-recognition losses and manual annotation. Compared with adding an identity loss to penalize drift, this failure mode is easier to diagnose — problems show up in the mining stage, not in optimization.
• The role of DINOv2 is stated explicitly and backed by a control experiment: the paper clarifies that DINOv2 is used purely for appearance-diversity selection and not for identity verification, then shows with ArcFace/LPIPS in Tab. 2(a) that diversity selection does not trade off identity. This "I use an off-the-shelf model, but here is the one thing it is responsible for" style of accounting is rare in data-centric papers and worth copying.
• A geometric mean as a constraint on multi-aspect evaluation: manipulation demands that "edit correctly" and "don't disturb anything else" hold simultaneously, and using the geometric mean of MF and BC as the total prevents a model from gaming one aspect. The trick transfers to any dual-objective editing task (video editing, 3D editing, speech editing, code editing).
• The benchmark's own blind spot is a finding: single-reference settings barely improve (SINGLE/Object 7.58→7.54), which conversely suggests open-source models are close to saturation there — the real battleground for data gains is multi-subject composition and manipulation.

Limitations & Future Work

• The limitation the authors acknowledge first is the video-frame target domain gap: targets are real video frames and may still carry compression artifacts and motion noise, which could give generations a "video look". The authors offer three reasons why it is not currently obvious (multi-stage filtering removes many low-quality frames; diverse synthesized inputs discourage memorizing video-specific low-level patterns; fine-tuning starts from strong image priors and uses T2I regularization), but no quantitative domain-gap analysis is given (e.g. a matched-budget comparison against photographic targets). That gap is worth probing.
• Synthetic-input bias: all inputs are produced by FLUX.1 Fill and thus carry that generator's priors. Whether another inpainting model would be equivalent, or whether mixing generators is needed, is left unanswered.
• The evaluation leans entirely on a VLM judge: there is no reproducible automatic metric such as CLIP-I / DINO / CLIP-T, and scores come from GPT-4.1 with rubricized prompts; the human study has only six annotators and reports agreement with the VLM merely as "the same trend", without a correlation coefficient. Judge stability is a genuine concern for anyone reproducing the benchmark.
• A limitation of the task definition: manipulation supervision is essentially "erase and complete", i.e. simulating that something else used to be there. But real video almost never contains a scene where an identity swap actually happened, so the manipulation branch's inputs always carry synthetic traces. Borrowing instances across videos (moving an instance from video A into the same camera setup in video B) might shrink that domain gap further.
• Sample counts and subset naming: Tab. 1 reports 748k paired single-subject input samples and 1752k multi-subject, which sum exactly to 2.5M and agree with the total in the statistics section; the single-image reference subset (748k) coincides with the single-subject pair count, so these are two names for the same single-subject data in different passages — not a contradiction, but easy to misread, and a released data card should state the four sub-tasks' counts explicitly.

Related Work & Insights

• vs Subjects200K / UNO-1M / Echo-4o-Image (synthesis-based pipelines): they batch-produce paired samples with strong generators and prompts, which scales easily but inherits the identity signal from the generator, giving fine-grained inconsistency and identity drift; this paper synthesizes only the inputs and uses real frames as targets, so the identity comes from real observation. The cost is a much heavier dependence on video material quality and on the filtering pipeline.
• vs MultiID-2M (retrieval-based pipeline): retrieval takes pairs from real group photos, which is more realistic, but it is limited by the availability of multi-subject images and struggles to cover open-ended scenes; this paper manufactures pairs from video, making scale and scene coverage controllable and additionally enabling manipulation data. The two are complementary — retrieval for realism, video-based construction for scale and task coverage.
• vs XVerse and similar benchmarks: they focus on clean single-subject portraits and omit manipulation; OSBench covers four sub-tasks and provides a per-aspect manipulation protocol (MF/BC), at the cost of depending on a VLM judge rather than automatic metrics.
• vs OmniGen2 / Qwen-Image-Edit-2509: this paper proposes no new architecture and instead layers a data gain on top of these two bases, obtaining +0.60 and +1.23 under matched compute. This confirms that "data hasn't kept up" is indeed a major bottleneck in the subject-driven line — but it also means the value of the method rests entirely on how representative these two bases are; whether the gain holds for other architectures is supported by only two data points.
• Takeaway: once subject-driven work shifts from "adding modules" to "manufacturing data", the centre of gravity of evaluation must shift too — single-subject generation is close to saturation, so what genuinely needs new benchmarks is multi-subject composition and manipulation, and evaluating those depends heavily on judge reliability. The two issues are mutually causal.

Rating

• Novelty: ⭐⭐⭐⭐ Treating video as the supervision source for subject consistency and synthesizing inputs rather than targets is a clearly fresh angle in the data recipe; overall, though, it remains a dataset-plus-benchmark paper with limited methodological novelty.
• Experimental Thoroughness: ⭐⭐⭐⭐ An internal benchmark plus two external suites (OmniContext / ImgEdit), per-component pipeline ablations, data-source and scaling comparisons, and a human study — the coverage is complete; points are deducted for the absence of automatic identity/text-alignment metrics (no CLIP-I / DINO / CLIP-T), a human study with only six annotators, and no quantitative control for the video-target domain gap.
• Writing Quality: ⭐⭐⭐⭐ The four-stage pipeline is described clearly, the verifier's reliability gets its own LAION-threshold comparison, and every key design is backed by numbers; against this, the text repeats itself noticeably (the "To address this gap / challenges, we introduce OpenSubject…" passage appears twice in the introduction) and some tables are already misaligned in the rendered version.
• Value: ⭐⭐⭐⭐⭐ With 2.5M samples, 4.35M images, coverage of both generation and manipulation, and released code, it is a practical piece of infrastructure for follow-up work in the subject-driven line, and OSBench's geometric-mean protocol is likely to be adopted directly.