Verifiable Visual Rewards Transfer from Synthetic Scenes to Natural Prompts

Shuyue Stella Li; Xiaochuang Han; Yulia Tsvetkov; Luke Zettlemoyer

cs.AISep 28, 2026

Verifiable Visual Rewards Transfer from Synthetic Scenes to Natural Prompts

Authors: Shuyue Stella Li, Xiaochuang Han, Yulia Tsvetkov, Luke Zettlemoyer

Organizations: University of Washington

Abstract

Precise instruction following in image generation, such as satisfying object counts and spatial relations, remains an open challenge at least in part because it is learned using unreliable reward models such as object detectors and vision-language models. We introduce Verifiable Visual Rewards (VVR), the first framework for programmatically verifiable image rewards, and show that training on it generalizes to natural prompts. Each VVR task is a scene of geometric objects and relations among them, from which we derive both the prompt and a deterministic verifier, so tasks can be generated in any number and at any chosen complexity. We release VVRBench, with 10,000 tasks over 32 constraint types, and VVRBench-Challenge, with 720 more complex tasks; the strongest model we evaluate---GPT-Image-2.5---solves 21.4% of VVRBench-Challenge. Using VVR scores as rewards for reinforcement learning (RLVVR) raises the accuracy of Stable Diffusion 3.5 Medium on VVRBench from 2.8% to 28.3% and demonstrates consistent easy-to-hard generalization. These gains extend to out-of-domain benchmarks, and mixing VVR into existing objectives further improves overall performance and human preference, motivating the adoption of VVR into standard image generation post-training recipes.

Figures & tables

Figure 1: Example VVRBench tasks from different complexity ranges ( $C_{1}$ – $C_{5}$ ) and Challenge ( $C^{*}$ ). Each panel shows the prompt, its complexity, the number of object instances, the number of constraints, the active constraint families, and a reference image that satisfies every constraint.

Panel of Table 1 — Table 1: Constraint library. VVR groups 46 constraint types into five families. Appendix A.1 lists all types and their supported values, and Appendix C.2 defines their verifiers.

Dataset	Use	Size	Complexity
VVRBench	evaluation	10,000	3–48
VVRBench -Fast	evaluation	820	3–44, 20 each
VVRBench -Challenge	evaluation	720	45–80, 20 each
VVR-Easy	training	100,000	$\leq$ 20
VVR-Matched	training	100,000	$\sim$ VVRBench

Table 2: Training and evaluation datasets produced by the VVR task generator. Appendix B.5 provides details on target distribution.

Model	Accuracy (%) $\uparrow$	$C_{1}$	$C_{2}$	$C_{3}$	$C_{4}$	$C_{5}$
GPT-Image-2	86.86 ±0.68	97.65 ±0.74	98.18 ±0.70	91.00 ±1.34	81.83 ±1.75	65.72 ±2.10
GPT-Image-1-mini	26.40 ±0.87	74.06 ±1.93	36.78 ±2.18	12.47 ±1.53	4.70 ±1.02	2.39 ±0.76
FLUX.2-dev	19.15 ±0.78	49.86 ±2.15	24.26 ±1.96	12.42 ±1.52	5.91 ±1.12	2.24 ±0.74
HunyuanImage-2.1	18.79 ±0.78	45.53 ±2.15	19.64 ±1.83	14.84 ±1.63	8.61 ±1.31	4.29 ±0.98
Qwen-Image-2512	5.79 ±0.47	19.16 ±1.75	5.87 ±1.14	2.11 ±0.73	1.10 ±0.56	0.15 ±0.29
HiDream-I1-Full	4.22 ±0.41	16.47 ±1.65	3.06 ±0.87	0.80 ±0.50	0.20 ±0.31	0.00 ±0.19

Table 3: Accuracy (%) on the 10,000 VVRBench tasks, overall and by complexity range.

C_{1}

C_{5}

split the tasks by structural complexity

C(s)

into five ranges of about 2,000 tasks each: 3–16, 16–21, 21–26, 26–31, and 31–48. Accuracy generally falls with complexity. GPT-Image-2 drops from 97.65% in

C_{1}

to 65.72% in

C_{5}

, and no open weight model exceeds 20% overall.

Figure 3: Pass rates of individual constraint for API models on VVRBench -Challenge, by family (left) and for the four constraint types with the lowest and the four with the highest average pass rates among those with at least 100 checks per model (right).

Training reward	VVR	GenEval	GenEval2	OCR	PickScore	HPSv2.1	HPSv3	CLIPScore	Aesthetic	ImageReward	UnifiedReward
Pretrained	0.028	0.616	0.237	0.476	0.841	0.300	7.689	0.956	5.517	0.929	0.636
$+$ VVR-Easy	0.283	0.729	0.268	0.587	0.849	0.294	8.275	0.979	5.483	1.114	0.641
$\Delta$	+0.255	+0.113	+0.031	+0.111	+0.008	$-0.006$	+0.586	+0.023	$-0.034$	+0.185	+0.005
GenEval2	0.039	0.688	0.454	0.501	0.848	0.297	8.289	0.974	5.523	1.110	0.635
$+$ VVR-Easy	0.218	0.718	0.478	0.532	0.848	0.302	8.511	0.976	5.533	1.155	0.637
$\Delta$	+0.180	+0.030	+0.025	+0.030	+0.0004	+0.005	+0.222	+0.002	+0.009	+0.045	+0.0019

Table 5: RLVVR with VVR-Easy and reward mixtures transfers to most benchmarks and metrics.

VVR win rate vs. baseline	VVR	GenEval2	GenEval	OCR	DrawBench	Outside VVR
VVR-Easy vs. pretrained	93.3	78.8	62.9	69.6	75.0	71.6
	[86.7, 98.3]	[67.5, 88.8]	[50.4, 75.0]	[58.3, 80.4]	[65.4, 84.2]	[65.9, 77.1]
GenEval2 $+$ VVR-Easy vs. GenEval2	89.2	62.9	57.5	56.7	57.5	58.6
	[81.7, 95.4]	[50.8, 74.6]	[45.0, 69.6]	[44.6, 68.8]	[45.4, 69.2]	[52.6, 64.6]

Table 6: Human preference win-rate (%) for the RLVVR-trained model over its baseline.

Figure 6: Example generations from the RLVVR and baseline models from the same prompts and initial seed. On the VVR prompt, the verifier accepts both RLVVR outputs and rejects both baselines. On the DrawBench prompt, VVR-Easy renders the vase as a flat shape without shading, consistent with its lower aesthetic and HPSv2.1 scores; adding GenEval2 to VVR-Easy recovers shading and improves aesthetic and HPSv2.1.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Family	Exact constraint types & their supported values	Contribution to $C(s)$
Grounding	color_attribute : red, orange, yellow, green, cyan, blue, purple, pink	$L(n_{i})$ for the referenced group
	shape_attribute : circle, square, triangle	$L(n_{i})$ for the referenced group
	color_shape_binding	No additional term; the bound group’s color and shape terms already account for it
Cardinality	exact_count : 1–10	$L(n_{i})$ for the referenced group
	same_count ; more_than_count ; fewer_than_count	$L(N_{r})$
	times_as_many : factor $k\in\{2,3,4,5\}$	$L(N_{r})+\log_{2}k$ , $k\in\{2,3,4,5\}$

Appendix

Table 7: The exact 46 constraint types, their supported values, and their contributions to structural complexity. † marks types that appear only in VVRBench -Challenge. Types with the same cost rule share a row, and types without listed values take only object groups as arguments. The background constraint supports white, black, light gray, dark gray, beige, pale pink, and pale cyan. The background and forbidden-content constraints in

\mathcal{B}

and

\mathcal{F}

apply to every task and do not contribute to

C(s)

Figure 7: Reference image of the example scene.

Dataset	Size	Candidates	Target distribution
VVRBench	10,000	Single strata and compositions of two to six strata at five scene-size settings	Complexity 3–48, capped at each integer complexity
VVRBench -Fast	820	VVRBench	20 tasks at each attainable integer complexity from 3 to 44
VVRBench -Challenge	720	Scenes seeded by each of the 46 constraint types, with up to six added relation or layout constraints	20 tasks at each integer complexity from 45 to 80, and at least 20 tasks per non-grounding constraint type
VVR-Easy	100,000	One constraint type from one stratum	Complexity at most 20, equal quotas over the nine strata
VVR-Matched	100,000	The VVRBench and VVRBench -Challenge generators	The strata and complexity distribution of VVRBench

Appendix

Table 8: Candidate pool and target distribution of each dataset.

Figure 8: Object extraction on three VVRBench -Fast outputs of API models. (a) The generated image. (b) The cleaned mask of one requested color over the grayed-out image. (c) The connected components of all requested colors, each labeled with the shape that receives its highest score. The bottom image is a textured crayon drawing under uneven light: the orange mask still covers the whole triangle, and the lit background adds small orange fragments (dashed boxes) that are too small to count as objects. The verifier accepts all three images. Prompts: (top) “Place the red triangle inside the yellow square. Add a purple circle and a cyan circle as well. Keep the background plain white, with no additional colored objects.” (middle) “Show a cyan circle, a red square, a green triangle, an orange circle, and two yellow squares. Use a plain pale pink background and no other colored objects.” (bottom) “The image should contain a pink circle, a green square, and an orange triangle. Set the objects against a plain white background; do not add other colored objects.”

Figure 9: Object extraction on distorted generations of open-weight models. Panels as in Figure 8 , except that (c) shows only the components of the highlighted color. (Top, ambiguous color) The circle requested as purple shades from purple into pink; the purple mask covers only its upper part, whose highest shape score is triangle 0.62. (Middle, irregular contour) The green circle grows a tail that reaches into the orange square; its circle score drops to 0.69, compared with 0.89 to 0.99 for the undistorted circles in the same image. (Bottom, blurred boundaries) The purple mask covers the whole blurred square, which scores 1.00 as a square; the image still fails because the prompt asks for two cyan triangles and two purple squares and the image shows one of each. The verifier rejects all three images.

Open-weight model	Steps	Guidance	API model	Settings
FLUX.2-dev	50	4.0	GPT-Image-2.5-Sunburst (2026-09-08)	medium quality, $1024\times 1024$
HunyuanImage-2.1	50	3.5	GPT-Image-2 (2026-04-21)	medium quality, $1024\times 1024$
Qwen-Image-2512	50	4.0	GPT-Image-1-mini	medium quality, $1024\times 1024$
HiDream-I1-Full	50	5.0	Gemini-3-Pro-Image	1K
FLUX.1-dev	28	3.5	Gemini-3.1-Flash-Image	1K
FLUX.1-schnell	4	0.0	Gemini-3.1-Flash-Lite-Image	1K

Appendix

Table 9: Generation settings. Guidance is the classifier-free guidance scale.

Model	Accuracy (%) $\uparrow$	3 to 10	11 to 18	19 to 26	27 to 35	36 to 44
GPT-Image-2.5-Sunburst	84.51 ±2.64	100.00 ±2.67	98.75 ±3.19	96.25 ±4.19	77.22 ±6.66	56.67 ±7.30
GPT-Image-2	82.20 ±2.77	98.57 ±3.63	98.75 ±3.19	95.62 ±4.38	74.44 ±6.84	50.56 ±7.24
Gemini-3.1-Flash-Image	47.20 ±3.42	85.71 ±6.75	60.00 ±7.74	51.25 ±7.68	30.56 ±7.08	18.89 ±6.35
Gemini-2.5-Flash-Image	45.61 ±3.42	76.43 ±7.68	65.62 ±7.65	51.25 ±7.68	31.11 ±7.10	13.33 ±5.74
Gemini-3.1-Flash-Lite-Image	42.07 ±3.41	85.71 ±6.75	56.25 ±7.74	40.00 ±7.74	25.00 ±6.80	14.44 ±5.88
Gemini-3-Pro-Image	37.20 ±3.36	62.14 ±8.26	52.50 ±7.71	38.75 ±7.73	26.67 ±6.90	13.33 ±5.74

Appendix

Table 10: Accuracy (%) of API models on VVRBench -Fast, an 820 task subset of VVRBench with 20 tasks at each attainable integer complexity from 3 to 44. The GPT-Image-2 models exceed 80% overall but fall to 51% to 57% in

C_{5}

. Gemini models reach 37% to 47%. Subscripts are 95% confidence margins.

Panel of Table 11 — Table 11: Pass rates (%) of API models for every constraint type on VVRBench -Challenge. $n$ is the number of checks of each type per model, and Avg is the unweighted mean over the seven models. Responses without an image count as failures. Within each family, types are sorted by Avg. Cell shading is proportional to the pass rate.

Model	Overall	Grounding	Cardinality	Spatial	Size	Topology
GPT-Image-2	86.9	86.8 $(+0.5)$	80.0 $(+8.0)$	85.4 $(-4.8)$	87.6 $(+10.0)$	85.6 $(+7.7)$
GPT-Image-1-mini	26.4	26.6 $(+0.5)$	14.2 $(+1.8)$	20.7 $(-9.7)$	19.2 $(+2.5)$	10.7 $(-3.3)$
FLUX.2-dev	19.1	19.0 $(-0.5)$	8.5 $(-4.0)$	16.3 $(+0.3)$	11.7 $(-1.4)$	9.1 $(-1.2)$
HunyuanImage-2.1	18.8	19.0 $(+0.6)$	10.3 $(+0.0)$	16.5 $(-0.4)$	6.5 $(-10.7)$	11.5 $(-0.7)$
Qwen-Image-2512	5.8	6.7 $(+1.0)$	2.2 $(-0.7)$	4.4 $(+0.2)$	2.4 $(-1.3)$	3.3 $(+1.1)$
HiDream-I1-Full	4.2	5.3 $(+0.1)$	2.0 $(+0.1)$	2.6 $(-1.8)$	1.6 $(-0.5)$	1.5 $(+0.0)$

Appendix

Table 12: VVRBench accuracy (%) on prompts that contain each constraint family. In parentheses is the difference from prompts without that family at matched complexity. Models show distinct weaknesses: Spatial for GPT-Image-1-mini (

-9.7

), Size for HunyuanImage 2.1 (

-10.7

), and Cardinality for FLUX.2 dev (

-4.0

Panel of Figure 10 — Figure 10: Failed API model outputs on VVRBench -Fast. Prompts: (a) “Place an orange circle in the top area. Set the objects against a plain pale pink background; do not add other colored objects.” (b) “Place a yellow square in the top area. Set the objects against a plain pale pink background; do not add other colored objects.” (c) “Arrange three purple squares in these cells of a 3-by-3 grid: top center, top right, and middle right. Set the objects against a plain black background; do not add other colored objects.”

Setting	Value
Trainable parameters	LoRA on the eight attention projections add_k , add_q , add_v , add_out , k , q , v , and out ; rank 32 and $\alpha=64$ .
Generation	$512\!\times\!512$ pixels; 25 denoising steps; classifier-free guidance 4.5; Gaussian sampling noise with level 0.7.
Rollout batch	32 prompt groups per update, 24 rollouts per prompt, and 768 generated images per update.
Flow GRPO	One inner epoch; advantages centered within each prompt group and divided by the standard deviation over the complete rollout batch; advantages clipped to $[-5,5]$ ; policy-ratio clip $10^{-4}$ ; KL coefficient 0.04. The first 24 of the 25 sampled transitions contribute to the update.
Optimization	AdamW; learning rate $3\!\times\!10^{-4}$ ; $\beta_{1}=0.9$ , $\beta_{2}=0.999$ ; $\epsilon=10^{-8}$ ; weight decay $10^{-4}$ ; maximum gradient norm 1.0; FP16 mixed precision with TF32 enabled; exponential moving average.
Training duration	3,000 optimizer updates, corresponding to 96,000 prompt groups and 2,304,000 generated images.

Appendix

Table 13: Reproducible configuration for the final SD3.5 Medium post-training experiments. Reward proportions are fractions of prompt groups in each update; every rollout is scored only by the reward attached to its prompt.

Training reward	Accuracy (%) $\uparrow$	$C_{1}$	$C_{2}$	$C_{3}$	$C_{4}$	$C_{5}$
SD3.5-M (pretrained)	2.81 ±0.34	12.01 ±1.47	1.19 ±0.59	0.35 ±0.37	0.05 ±0.23	0.00 ±0.19
GenEval2	3.87 ±0.40	15.42 ±1.61	2.39 ±0.78	0.91 ±0.52	0.05 ±0.23	0.05 ±0.23
GenEval2 + VVR-Easy	21.82 ±0.82	54.51 ±2.15	32.05 ±2.12	14.08 ±1.60	6.41 ±1.16	1.10 ±0.56
VVR-Easy	28.27 ±0.89	67.72 ±2.04	45.45 ±2.23	17.51 ±1.73	8.36 ±1.29	1.35 ±0.60
VVR-Matched	46.60 ±0.98	67.68 ±2.04	59.17 ±2.21	45.62 ±2.20	38.39 ±2.15	21.82 ±1.86
GenEval2 + VVR-Matched	33.50 ±0.93	57.93 ±2.13	47.27 ±2.23	32.34 ±2.09	20.32 ±1.82	9.22 ±1.34

Appendix

Table 14: VVRBench accuracy (%) of SD3.5 M after post-training. Training on VVR raises accuracy from 2.81% to 28.27% with Easy tasks and 46.60% with Matched tasks. Matched training gives the largest gains at high complexity. Adding VVR-Easy to GenEval2, OCR, or the five-reward objective raises accuracy by a factor of three to seven. Subscripts are 95% confidence margins.

Range	Model	Partial	All	Partial	All	Accuracy (%)
		Counts		Relations
$C_{1}$	Pretrained	0.76	0.46	0.29	0.19	12.01
	VVR-Easy	0.98	0.92	0.80	0.58	67.72
	VVR-Matched	0.98	0.93	0.80	0.57	67.68
$C_{2}$	Pretrained	0.70	0.17	0.25	0.13	1.19
	VVR-Easy	0.96	0.79	0.75	0.51	45.45

Appendix

Table 15: Partial and joint constraint satisfaction on VVRBench by complexity range. Partial scores are the mean graded count and relation scores; “all” is the fraction of tasks in which every count or every relation is satisfied. Relation columns use only tasks with at least one relation.

Training reward	VVR	GenEval	GenEval2	OCR	PickScore	HPSv2.1	HPSv3	CLIPScore	Aesthetic	ImageReward	UnifiedReward
Pretrained	0.028	0.616	0.237	0.476	0.841	0.300	7.689	0.956	5.517	0.929	0.636
GenEval2	0.039	0.688	0.454	0.501	0.848	0.297	8.289	0.974	5.523	1.110	0.635
GenEval2 + VVR-Easy	0.218	0.718	0.478	0.532	0.848	0.302	8.511	0.976	5.533	1.155	0.637
VVR-Easy	0.283	0.729	0.268	0.587	0.849	0.294	8.275	0.979	5.483	1.114	0.641
$\Delta$ add VVR-Easy	+0.180	+0.030	+0.025	+0.030	+0.0004	+0.005	+0.222	+0.002	+0.009	+0.045	+0.002
VVR-Matched	0.466	0.709	0.360	0.536	0.847	0.291	7.997	0.980	5.472	1.109	0.636

Appendix

Table 16: Complete final evaluation of the pretrained model and nine post-training conditions. PickScore uses the Pick-a-Pic v1 validation benchmark and HPSv2.1 uses the official four-domain HPDv2 benchmark. HPSv3, CLIPScore, aesthetic score, ImageReward, and UnifiedReward are macro-averaged across the shared cross-domain panel. VVR accuracy, GenEval, GenEval2, and OCR retain their native units. Every HPSv3 value, including the pretrained one, is the mean over the four cross-domain prompt sets. Images of the pretrained model use sampling seed 42; images of trained models use seed 20260912, except for PickScore and HPSv2.1, which use seed 42 for every model.

Contrast	PickScore	HPSv2.1	CLIPScore	Aesthetic	ImageReward
GenEval2 $+$ VVR-Easy $-$ GenEval2	$+0.0004$ [ $-0.0012$ , $0.0021$ ]	$+0.0047$ [ $0.0042$ , $0.0052$ ]	$+0.0023$ [ $-0.0014$ , $0.0058$ ]	$+0.009$ [ $-0.003$ , $0.022$ ]	$+0.045$ [ $0.025$ , $0.066$ ]
GenEval2 $+$ VVR-Matched $-$ GenEval2	$-0.0002$ [ $-0.0019$ , $0.0015$ ]	$+0.0039$ [ $0.0034$ , $0.0045$ ]	$-0.0011$ [ $-0.0044$ , $0.0021$ ]	$-0.007$ [ $-0.020$ , $0.006$ ]	$+0.038$ [ $0.020$ , $0.057$ ]
OCR $+$ VVR-Easy $-$ OCR	$+0.0018$ [ $-0.0002$ , $0.0037$ ]	$0.0000$ [ $-0.0007$ , $0.0006$ ]	$+0.0086$ [ $0.0054$ , $0.0118$ ]	$+0.012$ [ $-0.001$ , $0.025$ ]	$+0.092$ [ $0.071$ , $0.113$ ]
Five-reward $+$ VVR-Easy $-$ Five-reward	$-0.0024$ [ $-0.0041$ , $-0.0006$ ]	$+0.0060$ [ $0.0054$ , $0.0066$ ]	$+0.0023$ [ $-0.0009$ , $0.0055$ ]	$+0.031$ [ $0.020$ , $0.042$ ]	$+0.027$ [ $0.009$ , $0.046$ ]

Appendix

Table 17: Effect of adding VVR to an existing reward, with paired bootstrap 95% intervals over prompts (10,000 resamples; stratified by prompt set for macro-averaged metrics and by domain for HPSv2.1). The GenEval, GenEval2, OCR, HPSv3, and UnifiedReward evaluations retained only aggregate scores, so they have no intervals.

Annotator pair	Agreement, both chose (%)	Agreement, ties as a label (%)	Cohen’s $\kappa$
1 vs. 2	84.9 (298/351)	76.0 (304/400)	0.50
1 vs. 3	81.9 (276/337)	72.0 (288/400)	0.44
2 vs. 3	84.6 (312/369)	78.8 (315/400)	0.53
Mean	83.8	75.6	0.49

Appendix

Table 18: Agreement between annotator pairs on the 400 preference pairs. Labels are decoded to the preferred model before comparison. “Both chose” uses only the pairs on which neither annotator chose a tie; “ties as a label” uses all 400 pairs with tie as a third label, which is also the label set of Cohen’s

\kappa

. Fleiss’

\kappa

over the three annotators is 0.49.

Explore similar work

Visual Jev Rewards: Reference-Bound Verification for Multi-Subject Image Generation
Oct 7, 2026Baoteng Li, Wenzhuo Wu, Kongming Liang +1Multimodal Generation
School of Artificial Intelligence, Beijing University of Posts and Telecommunications1 · Beijing Key Laboratory of Multimodal Data Intelligent Perception and Governance2
VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning
Date pendingJunxiang Xu, Ruisi Wang, Fanyi Pu +49Visual Reasoning Multimodal Reasoning Benchmarks
Qwen-Image-2.0-RL Technical Report
Jun 25, 2026Yixian Xu, Kaiyuan Gao, Yuxiang Chen +25Reward Functions Offline Reinforcement Learning
Equal contribution. · Corresponding author.

Oct 7, 2026, cs.CVJ/K move · Enter open · S save

Visual Jev Rewards: Reference-Bound Verification for Multi-Subject Image Generation

Baoteng Li, Wenzhuo Wu, Kongming Liang, Zhanyu Ma

School of Artificial Intelligence, Beijing University of Posts and Telecommunications1 · Beijing Key Laboratory of Multimodal Data Intelligent Perception and Governance2

Multimodal Generation

Open paper AskPDF arXiv

Multi-subject image generation requires rewards that verify whether requested attributes, actions, and relations hold for the specified reference subjects. Subject presence alone does not establish that the correct subjects participate in a requested interaction. We present reference-bound Visual Jev rewards that turn these visual decisions into generator training signals. Each subject-related question receives a positive label only when the requested condition and the relevant reference identities hold jointly. We construct fixed questions offline, train a Qwen3.5-4B verifier with binary supervision, and directly read Yes probabilities from its language-model head. Their mean supplies a GRPO reward while retaining individual judgments for inspection. Using 200 MICo-150K training tasks and 30 updates, the framework raises a GPT-5.4 composite score from 41.78 to 52.50 on a manually selected 897-task MICo-Bench subset; direct 27B rewards yield 51.84. Each reward is tested in one GRPO run, and offline human evaluation does not establish a statistically significant advantage over direct scoring. The study provides an initial implementation and evaluation of Visual Jev as a reference-bound reward for multi-subject image generation.

Abstract

Explore similar work

Visual Jev Rewards: Reference-Bound Verification for Multi-Subject Image Generation

VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning

Qwen-Image-2.0-RL Technical Report