acceptodds
Under review as a conference paper at ICLR 2027

FLUX Learns To Write

10 views

Abstract

Multimodal Diffusion Transformers (MM-DiTs) jointly attend to image and text tokens, introducing image-to-text feedback and implicit visually grounded semantics, making them promising for unified understanding and generation. However, existing approaches largely expose this capability only through pixel-level grounding and editing, while their reliance on full-sequence text denoising limits coherent long-range reasoning over jointly grounded visual and textual information. Moreover, weak cross-modal coupling limits visual feedback to textual reasoning, leading to semantic misalignment. We propose UniFlux, which adapts a pretrained MM-DiT for multimodal understanding using block-wise discrete flow matching, for long and coherent multimodal reasoning while preserving its generation prior. We further introduce X-GRPO, a cross-modal reinforcement-learning objective that couples separately normalized text- and image-side advantages through positive-gated multiplicative credit, assigning positive rollout credit only when both reasoning quality and image-regeneration fidelity exceed their respective group baselines. UniFlux achieves state-of-the-art performance across eleven understanding and generation benchmarks

Primary area:
generative models
Keywords:
Unified Model, Image and Text Synthesis, Discrete Flow matching, discrete diffusion

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3QdnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces UniFlux, which aims to extend the pre-trained MM-DiT image generation model into a unified model capable of multimodal understanding, reasoning, and image generation. The authors first modify the text stream of MM-DiT using block-wise discrete flow matching, enabling the model to generate lengthy reasoning and responses step-by-step (block by block) while retaining its original image generation capabilities. In the second stage, the authors propose X-GRPO, a method that combines text-side response quality with image-side regeneration quality for reinforcement learning, allowing visual generation feedback to provide a backward constraint on text-based reasoning. Experiments conducted using FLUX.2-Klein 9B and FLUX.2-Dev 32B demonstrate strong performance across multiple benchmarks, including visual understanding, multi-image reasoning, OCR, and image generation.

Critical strengths
  1. The paper presents a compelling idea by extending the existing MM-DiT text stream for visual understanding and language reasoning, without introducing a separate language decoder.

  2. Block-wise discrete flow matching and X-GRPO complement each other well. The former supports causal long-form reasoning, while the latter uses image-generation feedback to better align understanding and generation.

  3. The experiments are extensive, covering visual understanding, multi-image reasoning, OCR, and image generation. The ablations also provide clear evidence for the contributions of the main components.

Critical weaknesses
  1. Figure 1 contains some of the most important information in the paper, but it is noticeably blurry and difficult to read. Given its central role in presenting the overall framework, a high-resolution or vector version would substantially improve readability.

  2. The methodological novelty appears somewhat incremental. Several key ingredients, including block-wise diffusion, GRPO-style optimization, and joint understanding-generation training, have been explored in prior work. The main contribution therefore seems to come from adapting and combining existing techniques within the MM-DiT framework rather than introducing a fundamentally new modeling principle.

  3. Table 1 mainly compares UniFlux with existing unified multimodal models and a few earlier understanding-only models. Although direct comparison with stronger recent multimodal LLMs may not be fully fair because of architectural and training differences, including such results would still provide a useful reference point and help readers better understand the remaining performance gap and potential room for improvement.

Questions for authors

Could the authors clarify the main methodological novelty of UniFlux compared with prior block-wise diffusion and GRPO-based multimodal training?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 3Qdn (1/2)

We thank the reviewer for appreciating the extension of the MM-DiT text stream without a separate language decoder, the complementary roles of block-wise DFM and X-GRPO, and the breadth of our evaluations and supporting ablations.

W1. Figure Readability.

We thank the reviewer for noticing this. We are working to revise the figure in the main paper with vector image and try to zoom the images to improve readability.

W2/Q1. Methodological Novelty.

Our technical contribution lies in adapting a pretrained image generator for explicit multimodal reasoning and coupling its understanding and generation through a specific credit-assignment mechanism. Concretely:

Turning the MM-DiT’s conditioning stream into a language-generation pathway. Our architectural contribution is enabling a pretrained image-generation backbone to produce visually grounded answers through its existing text stream. We introduce trainable token embeddings and a vocabulary prediction head to decode the visually conditioned text representations already processed by the MM-DiT. LoRA adaptation and block-causal discrete flow matching enable sequential reasoning across blocks and parallel refinement within each block, while retaining the native continuous image-generation pathway. This provides language output within the existing backbone without adding a separate autoregressive response decoder.

Generation feedback for understanding through X-GRPO. The main methodological novelty is X-GRPO’s assignment of joint credit to an understanding response and its associated image generations. Each sampled response, together with the source image and query, produces an image-generation instruction. The generated images provide complementary feedback for optimizing the originating text response. Our specific contribution is the cross-modal weighting: w{ij}={⁡sg}([{^A}Ti]+[{^A}I{ij}]+).w_\{ij\} = \operatorname\{sg\}\left([\hat\{A\}_T^i]_+[\hat\{A\}_I^\{ij\}]_+\right). Text and image rewards are normalized separately, and their coupled weight guides both policy objectives. Positive credit requires both advantages to be positive: a high image reward cannot compensate for a nonpositive text advantage, or vice versa. This specifies how the two tasks influence each other during optimization.

The ablations support these specific design choices. Replacing block-wise generation with full-sequence denoising reduces MMMU from 68.82 to 65.37. Positive-gated text-only RL achieves 66.54 MMMU / 0.85 GenEval, while X-GRPO achieves 68.82 / 0.91. Alternative coupling rules also perform worse: additive advantages achieve 64.03 / 0.86, and a signed product achieves 63.12 / 0.82. These comparisons support the value of both the causal text-generation design and the particular cross-modal objective.

Compared with prior block-wise diffusion: Block-wise diffusion establishes sequential generation across blocks and parallel refinement within blocks. UniFlux adapts this mechanism to the pretrained MM-DiT’s visually conditioned text stream using trainable answer embeddings, a vocabulary head, and block-causal discrete flow matching. This allows the backbone to generate reasoning and answers conditioned on the image and completed response blocks while retaining its continuous image-generation pathway. The contribution here is the architecture and training formulation that support these two generation processes within the pretrained backbone.

Compared with prior GRPO-based multimodal training: Existing approaches include applying task-specific rewards across reasoning and generation tasks, as in MMaDA, and using generated-image rewards to optimize textual planning and image generation, as in T2I-R1. More precisely, in T2I-R1, the model produces a textual plan and then an image. The generated image is scored, and that score is used to train both the textual plan and image generation. The plan is therefore rewarded indirectly through the resulting image. X-GRPO introduces a different rule for connecting the two tasks:

  • Each sampled understanding response is linked to associated image generations.

  • The answer and images are scored separately and compared with their respective sampling groups.

  • Their positive advantages are multiplied, and the resulting weight trains both pathways.

Thus, an answer–image pair receives positive credit only when both the answer and image score above their respective group averages. A highly rewarded image cannot compensate for a below-average answer, and a highly rewarded answer cannot compensate for a below-average image. This is the specific methodological distinction: both evaluations jointly determine the reinforcement assigned to each linked response–image pair.

AuthorsOpenReview ↗

Response to Reviewer 3Qdn (2/2)

W3. Stronger Understanding Baselines.

We have expanded Table 1 to include BLIP-3-4B, InternVL3-8B, and MiniCPM-o-4.5 (Instruct) as additional visual-understanding references. The comparison covers all six understanding benchmarks in Table 1, providing broader context beyond unified understanding-and-generation models.

ModelMS-COCO CIDEr ↑VQAv2 Acc ↑MME-P ↑GQA Acc ↑MMMU Acc ↑MathVista Acc ↑
BLIP-3-4B50.1142.391310.0045.6741.1039.60
InternVL3-8B70.6966.53174851.2262.7071.60
MiniCPM-o-4.5 (Instruct)69.3461.471399.0059.5267.6080.10
UniFlux-9B73.3376.781472.0073.1768.8280.43

Among the listed configurations, UniFlux-9B achieves the highest score on all six benchmarks. Compared with MiniCPM-o-4.5 in Instruct mode, it achieves gains of 1.22 percentage points on MMMU and 0.33 percentage points on MathVista, while also supporting image generation. We present these comparisons as performance references, with model configurations explicitly identified, recognizing the architectural and training differences highlighted by the reviewer.

We thank the reviewer for the constructive feedback and encouraging assessment. The clearer distinctions from prior methods and expanded comparisons with recent VLMs strengthen the presentation and clarify UniFlux’s architectural and cross-modal optimization contributions. We hope these revisions reinforce the reviewer’s support and merit a stronger recommendation.

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

This paper proposes a training method to further transform the pre-trained MM-DiT image generation model into a unified model, endowing it with image generation, text generation, and visual reasoning capabilities simultaneously. On the text side, the method employs block-wise discrete flow matching; after training, a new cross-modal X-GRPO is used to jointly optimize text understanding and image generation.

MM-DiT → Unified Model: Experiments based on FLUX.2 9B/32B demonstrate that the pre-trained image-generation model MM-DiT can acquire text generation, visual understanding, and reasoning capabilities through further training, while retaining its text-to-image (T2I) capability.

Block-wise DFM: Removing block-wise generation and replacing it with full-sequence denoising resulted in a performance drop, indicating that the block-wise generation design is effective.

X-GRPO: Outperforms methods without reinforcement learning, text-only, image-only, and other reward coupling approaches

Critical strengths

This paper proposes a practical approach for transforming the strongly pre-trained MM-DiT image generator into a unified understanding and generation model. The approach includes a well-motivated block-level text generation mechanism and an innovative cross-modal reinforcement learning objective, supported by a comprehensive set of internal ablation experiments.

Critical weaknesses

Both of these points constitute significant shortcomings and are not issues that can be resolved simply by conducting a few additional experiments.

1. The core representation claim has not been causally decoupled from large-scale VLM supervision. The paper attempts to demonstrate that the text stream from the pre-trained MM-DiT already possesses explicitly expressible visual semantic understanding capabilities. However, the model used for VLM evaluation has undergone large-scale Stage-I VLM supervision, with LoRA, token embeddings, and vocab heads all trained; the training data includes image→description, instruction response, reasoning trace, VQA/OCR/reasoning, and other content. Therefore, the results can at most demonstrate that pre-trained FLUX can serve as an initialization for further training into a strong VLM; they cannot demonstrate that the original pre-trained FLUX text stream itself possesses these understanding and reasoning capabilities. At a minimum, a control experiment is needed in which the pre-trained FLUX text stream probe is frozen, or a comparison is made between pre-training and random initialization/training from scratch using the same data and the same training volume. This critical control experiment is currently lacking.

2. The cross-architecture generalizability of X-GRPO has not been established. The paper proposes X-GRPO as a new cross-modal reinforcement learning objective, but the experiments were conducted only within the same model family—UniFlux/FLUX.2 9B and 32B—and were not validated on different architectures such as BAGEL, D-DiT, or the LLM+Diffusion dual-tower architecture. Therefore, the results only demonstrate that X-GRPO is effective for UniFlux; they do not indicate that it possesses cross-architecture generalization. These two points alone constitute serious flaws. In particular, the first point directly concerns the paper’s core motivation: whether the approach is uncovering FLUX’s inherent understanding capabilities or retraining those capabilities using large-scale VLM data. The current experiments fail to distinguish between the two.

Questions for authors
  1. Key Claims Regarding the Pre-trained MM-DiT The paper claims that the text stream used to pre-train MM-DiT already contains implicit visual semantics and visual understanding capabilities. However, prior to the final VLM evaluation, the model underwent large-scale supervision with visual instructions and reasoning, and was trained using LoRA, token embeddings, and a vocabulary head. How do the authors distinguish whether the final visual understanding capabilities stem from the pre-trained MM-DiT itself or were newly acquired during the Stage I VLM supervision?

  2. Regarding the Scope of X-GRPO The paper describes X-GRPO as a cross-modal reinforcement learning objective for jointly aligning multimodal understanding and generation, but the current experiments are validated only on the UniFlux/FLUX.2 series. Does X-GRPO rely on specific properties of the shared MM-DiT architecture? Do the authors have evidence showing that the method is transferable to other unified multimodal architectures, such as autoregressive text + diffusion image models or dual-path models? If not, should the scope of the method’s claims be further narrowed?

16 comments
AuthorsOpenReview ↗

Response to Reviewer NAvo (1/3)

We thank the reviewer for recognizing the practical value of transforming a pretrained MM-DiT into a unified model, the well-motivated block-wise generation mechanism, the innovative cross-modal reinforcement learning objective, and the comprehensive ablation studies.

W1/Q1. Pre-trained Representations.

We thank the reviewer for this important suggestion. To examine the contribution of pretrained FLUX representations separately from subsequent adaptation, we conducted the following controls using the same training data and compute budget. The external text encoder, VAE, and evaluation settings are kept identical across the controls.

Model / ExperimentCOCO CIDEr ↑VQAv2 ↑MME ↑GQA ↑MMMU ↑MathVista ↑GenEval ↑DPG-Bench ↑
UniFlux-9B (Stage I + II)73.3376.78147273.1768.8280.430.9190.08
UniFlux-9B (Stage I only)66.2171.29136265.1162.0366.290.7982.29
Pretrained FLUX, frozen backbone + trained embedding/head54.1255.54121142.2431.0435.020.7778.77
Random MM-DiT, frozen backbone + trained embedding/head14.449.8917810.119.446.670.1512.39
Random MM-DiT, Stage-I LoRA training16.3812.1723513.4312.158.290.2117.72
Random MM-DiT, full Stage-I adaptation58.3959.42127049.7237.4940.150.6870.14

The matched training budget applies to the Stage-I controls. The final UniFlux model additionally undergoes Stage II and is included as a performance reference.

Experimental configurations.

UniFlux-9B (Stage I + II): The complete model after supervised Stage-I adaptation and Stage-II X-GRPO. This row reports the final understanding and generation performance.

UniFlux-9B (Stage I only): Starts from pretrained FLUX and trains LoRA adapters, answer-token embeddings, and the vocabulary head. This provides the baseline before reinforcement learning.

Pretrained FLUX with a frozen backbone: Keeps all pretrained MM-DiT weights frozen, disables LoRA, and trains only the answer-token embeddings and vocabulary head. This tests whether useful information can be accessed through the text stream without modifying the pretrained backbone.

Random MM-DiT with a frozen backbone: Replaces the pretrained MM-DiT weights with random initialization, freezes them, and trains the same embeddings and vocabulary head. Comparison with the preceding configuration measures the benefit of pretrained backbone weights under the same language-interface training.

Random MM-DiT with Stage-I LoRA training: Applies the standard Stage-I LoRA adaptation recipe to a randomly initialized backbone. Comparison with pretrained Stage-I UniFlux tests the contribution of initialization under the same adaptation strategy.

Random MM-DiT with full Stage-I adaptation: Trains all randomly initialized MM-DiT weights, together with the embeddings and vocabulary head, without LoRA. This allows the random backbone itself to learn under the matched Stage-I data and compute budget. It does not include Stage II.

The frozen-backbone comparison directly addresses whether useful information is present before backbone adaptation. The frozen pretrained backbone achieves 31.04 on MMMU and 35.02 on MathVista, compared with 9.44 and 6.67 for the frozen random backbone. These gains of 21.60 and 28.35 percentage points support the presence of useful pretrained information that is accessible through the text stream using a trained language interface.

Under Stage-I LoRA adaptation, pretrained initialization achieves 62.03 on MMMU and 66.29 on MathVista, compared with 12.15 and 8.29 from random initialization. Allowing the random backbone to update all its parameters improves performance to 37.49 and 40.15, but it remains below pretrained Stage-I UniFlux under the same data and compute budget. This supports the practical value of pretrained initialization; the full-adaptation comparison involves different trainable parameter sets.

These results also show that Stage-I supervision contributes substantially. With the pretrained FLUX backbone frozen, training only the answer-token embeddings and vocabulary head achieves 31.04 on MMMU and 35.02 on MathVista. Adding LoRA adaptation of the backbone during Stage I increases these scores to 62.03 and 66.29, respectively. These results support two conclusions: useful information can be accessed from pretrained FLUX through a trained language interface, and adapting the backbone further improves its ability to use that information for understanding and reasoning.

We have added these experiments, together with their training and evaluation details, to the Sec 5.1 and clarify the scope of the representation claim.

AuthorsOpenReview ↗

Response to Reviewer NAvo (2/3)

W2/Q2. Cross-Architecture Generalization.

We thank the reviewer for highlighting the need to evaluate X-GRPO beyond the FLUX family. We would like to first clarify that the paper distinguishes implicit visual representations in pretrained FLUX from explicit language reasoning enabled through adaptation.

The Introduction describes “an implicit form of visual understanding,” while Section 3.2 explicitly states that these representations “serve only as internal conditioning states” and “are not parameterized to predict language tokens.” Our claim is that these pretrained representations provide a useful foundation that Stage I adapts for understanding and reasoning.

X-GRPO addresses a specific challenge in turning a pretrained image generator into a unified model: enabling language output does not, by itself, ensure that textual understanding and image generation reinforce each other. Stage I makes the MM-DiT’s visually conditioned text stream capable of producing answers.

Stage II then uses the model’s generation pathway to provide additional feedback on those answers, alongside their direct text-side evaluation. Concretely, each sampled answer, together with the source image and question, is converted into an image-generation instruction.

The answer and its associated generated images are evaluated separately. X-GRPO couples their normalized advantages so that a response–image pair receives positive credit only when both evaluations exceed their respective group baselines. This connects the two capabilities through the training objective: the reinforcement assigned to an answer depends partly on its associated image generations, and the reinforcement assigned to an image depends partly on the originating answer.

We conducted extensive additional experiments comparing four unified multimodal models: BAGEL, Show-o2-7B, OneFlow-8B, and LLaDA2.0-Uni. These cover Mixture-of-Transformer Experts, autoregressive text with continuous image flow, edit-flow text generation, and discrete-diffusion-based designs. For each model, we compare three configurations: Initial, the checkpoint before our additional RL training; Joint GRPO with independent advantages, which updates both modalities using their respective modality-specific advantages; and X-GRPO, which couples the normalized text and image advantages through the proposed positive-gated multiplicative weighting.

As shown below, X-GRPO improves over joint GRPO with independent advantages on all five evaluated benchmarks for each model. These results extend the evidence beyond the FLUX family and support transferability across the tested architectures.

ModelConfigurationMMMU ↑MathVista ↑MME ↑GenEval ↑DPG-Bench ↑
BAGELInitial58.2273.101312.000.8787.74
BAGELJoint GRPO, independent advantages60.0373.971357.000.9088.23
BAGELX-GRPO62.1777.491435.000.9290.95
Show-o2-7BInitial48.9062.831620.500.7686.14
Show-o2-7BJoint GRPO, independent advantages50.1162.981629.000.7787.15
Show-o2-7BX-GRPO54.1169.371666.000.8089.28
OneFlow-8BInitial33.1060.681542.500.8580.40
OneFlow-8BJoint GRPO, independent advantages33.9861.771550.000.8682.22
OneFlow-8BX-GRPO37.7264.471578.000.8985.29
LLaDA2.0-UniInitial50.1068.101335.000.8987.76
LLaDA2.0-UniJoint GRPO, independent advantages51.2969.021348.000.8988.93
LLaDA2.0-UniX-GRPO54.4473.391400.000.9291.19

X-GRPO outperforms independent-advantage Joint GRPO on all five benchmarks for each of the four models. The improvements range from 2.14–4.00 percentage points on MMMU, 2.70–6.39 percentage points on MathVista, 28–78 points on MME, 0.02–0.03 on GenEval, and 2.13–3.07 points on DPG-Bench.

These results extend the empirical validation from different FLUX model sizes to additional model families with different text and image generation mechanisms. They support the transferability of X-GRPO across the evaluated architectures, while keeping the claim scoped to these tested settings.

AuthorsOpenReview ↗

Response to Reviewer NAvo (3/3)

Regarding the representation concern, our separate initialization controls distinguish the contribution of pretraining from subsequent adaptation. With the backbone frozen, pretrained FLUX achieves 31.04/35.02 on MMMU/MathVista, compared with 9.44/6.67 for a frozen randomly initialized backbone. Stage-I adaptation further increases performance to 62.03/66.29. These results support contributions from both transferable pretrained representations and supervised adaptation, without implying that unmodified FLUX already expresses the final model’s reasoning capabilities.

We have added these cross-architecture experiments, the initialization controls, and their detailed training and evaluation settings to the Appendix H.5

We appreciate the constructive feedback and believe the resulting additions materially strengthen the paper. We hope these further analyses and clarifications fully address the reviewer’s concerns and support a more favorable assessment of our work.

Reviewer_f1j3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper turns FLUX.2, a pretrained MM-DiT image generator, into a unified model. It adds a vocabulary head to the text stream and trains it on about 85M examples with block-wise discrete flow matching. A second stage, X-GRPO, turns each sampled response into an image-generation instruction, renders images from it, and gives each response-image pair the advantage [A^T]+⋅[A^I]+[\hat A_T]_ + \cdot [\hat A_I]_+ (Eq. 2), with rewards from VLM judges (Gemini-3.5 Pro by default), HPSv3 and SigLIP. The 9B model reports 68.82 on MMMU and 80.43 on MathVista, and GenEval goes from 0.81 to 0.91.

Critical strengths
  1. Using the text stream of an MM-DiT as a language decoder is a neat idea, and I haven't seen it done this directly before. Block-wise DFM is a good fit for it: it gives variable length and causal context while keeping parallel refinement within a block. The block-size sweep in Figure 8(b) and the "w/o Block Diffusion" row (65.37 against 68.82) give the design some direct support.

  2. The paper is open about its recipe. Tables 5 and 6, the judge prompts in Figures 12 and 13, and Algorithms 1 and 2 make it a lot easier to see what was actually trained. Several of my concerns below only exist because these details are written down.

  3. The generation gains over the base model are big enough: GenEval position goes from 0.48 to 0.76 (Table 12).

Critical weaknesses
  1. I still can't accept the figures in the understanding table at face value since some of them are obviously incorrect. In the Qwen-VL row, 78.20 is listed under GQA and 57.50 under MMMU. A 2023 7B model scoring 78.20 on GQA would outperform all of the unified baselines in the table by about 14 points. It seems as though the VQAv2 and GQA scores for Qwen-VL-Chat (78.2 and 57.5 as given in the Qwen-VL paper) have simply been shifted one column to the right. If that is the case, then the "19.69% improvement in MMMU over Qwen-VL" mentioned in Section 4 is based on the GQA figure. Although the abstract says that state-of-the-art performance has been achieved on eleven benchmarks, on MME the 9B model (1472) scores below both Qwen-VL (1487.50) and InternVL-2.0 (1648.10), and InternVL-2.0 is highlighted in the same table. Table 10 shows a MMMU score of 66.72 for the final-layer text features, which is the default configuration, whereas in all other cases the figure is 68.82. Table 3 states that FUDOKI achieves 0.83 on GenEval, but the averages of its categories in Table 12 are 0.88.

  2. Yet even when the numbers are correct, I have no idea where the improvements come from. The training dataset consists of about 85 million examples, comprising 15 million from LLaVA-OneVision-1.5 Stage 2, 10 million from Mammoth-VL and 2 million GPT-4V chain-of-thought traces. Stage II is evaluated by Gemini-3.5 Pro, which has access to the reference answer. Since none of the unified baseline models have this data or carry out the evaluation, Table 1 compares the different recipes rather than the architectures. The Params column also includes only the DiT. FLUX.2 uses a large frozen language model to encode the query (assuming my understanding of the releases is correct, this is a roughly 24B Mistral VLM for the dev version and a Qwen3 model for klein). The paper never mentions this model or includes the number of its parameters.

The judge is important; for example, replacing Gemini-3.5 Pro with GPT-5.1 Mini causes MMMU to drop to 66.78 (see Table 10). This is almost identical to the figure obtained with text-only GRPO and the default judge (66.54, Table 4). It is also necessary to check for contamination. MathVista reuses questions from existing datasets and I think that ChartQA, AI2D and ScienceQA are included among those. All three of these datasets are part of the training data (as shown in Tables 8 and 9).

  1. X-GRPO does not carry out what it claims to do. To begin with, [A^T]+[A^I]+[\hat A_T]_ + [\hat A_I]_+ is never negative. When the advantage is nonnegative, the clipped surrogate simplifies to the advantage multiplied by min⁡(ρ,1+ϵ)\min(\rho, 1+\epsilon), which causes the gradients for below-average responses to be exactly zero. The update procedure therefore involves reward-weighted fine-tuning only on those responses that exceed both group averages. Although this might be adequate, it is not the same algorithm as GRPO. The paper fails to mention that it eliminates all negative signal, and Proposition F.2 only states that the product of two nonnegative numbers is nonnegative.

Secondly, the image reward is almost independent of whether the response is grounded. Since Gψ(q,I,A(i))G_ \psi(q, I, A^{(i)}) looks at the source image, it is able to describe the picture correctly regardless of what the response says. Of the three image rewards (HPSv3 and SigLIP when assessed against the instruction), two never examine the source image. The VLM prompt shown in Figure 13 evaluates adherence to the prompt and regards the reference image as optional. The variant that actually aligns with the cycle-consistency scenario, Gψ(q,A(i))G_\psi(q, A^{(i)}) without the image, performs worse (66.88 rather than 68.82).

Thirdly, the entry for "Understanding-only JT\mathcal{J}_ T" probably makes use of signed advantages. When it is compared with X-GRPO, both the coupling and the positive gate are affected. If instead a text-only run were carried out using [A^T]+[\hat A_T]_+, then the two effects would be separated. Moreover, the text-only reinforcement learning increases GenEval from 0.79 to 0.85, just as does the image-only RL. I cannot explain this, and it causes me to question how much of the improvement in GenEval shown in Table 4 is actually due to noise.

  1. The idea of "preserving the generation from the previous stage" is not entirely valid since Stage I alone reduces GenEval from 0.81 to 0.79, and Stage II then tunes the entire model (as shown in Table 5). The paper does not specify which backbone model was used with DanceGRPO and Flow-GRPO; since Flow-GRPO's own paper reports a score of 0.95 on GenEval, the result of 0.89 needs to be explained. Furthermore, Figures 22 and 23 compare against the Flux-2 4B rather than the 9B base model, and in Figure 7 MammothModa2 actually gets both the color binding and the "ICLR" text correct.

  2. The costs are never put in comparison with those of the baseline models. At the default setting of 300 reasoning tokens, it takes 47 seconds per question (as shown in Figure 15), with three forward passes carried out at each step for guidance, and there is no wall-clock time comparison made with the AR models that the introduction criticises for their high cost.

Questions for authors
  1. What text encoder does each UniFlux variant use and how many parameters does it bring? Can you also provide Table 1 with an open judge from Table 10 (e.g. Qwen3.6 35B) and a decontamination check of the training mix against MMMU and MathVista?

  2. Is there a run of text-only RL with the same positive gate [A^T]+[\hat A_T]_+? If X-GRPO is still beating it by more than the noise then I'd accept the cross-modal coupling as a genuine contribution.

  3. What is UU, and what did Stage II cost in judge calls and GPU hours?

9 comments
AuthorsOpenReview ↗

Response to reviewer f1j3 (1/4)

We thank the reviewer for recognizing the originality of using the MM-DiT text stream for language decoding, the suitability and ablation support of block-wise DFM, the transparency of our experimental recipe, and the substantial generation improvements.

W1/Q1. Understanding Table.

We thank the reviewer for identifying these typographical errors and apologize for the confusion. The Qwen-VL-Chat scores were misaligned across columns: 78.20 corresponds to VQAv2 and 57.50 to GQA, rather than GQA and MMMU, respectively. In Table 10, the MMMU score for the default final-layer text features should be 68.82, rather than 66.72. In Table 3, FUDOKI’s GenEval score should be 0.88, consistent with the average of its category scores in Table 12. We have corrected both entries in the revised manuscript. We have revised these tables to correct these values. InternVL-2.0 is correctly highlighted for achieving the highest listed MME score. These corrections do not change the main benchmark results reported for UniFlux. They correct the misreported final-layer ablation entry and the baseline comparisons. We have also removed the previously stated 19.69% MMMU improvement over Qwen-VL-Chat, which was calculated from the misplaced GQA score. The corrected comparison is provided below.

ModelVQAv2 ↑GQA ↑MME ↑MMMU ↑
Qwen-VL-Chat78.2057.501487.5035.90
UniFlux (9B)76.7873.171472.0068.82
UniFlux (32B)82.1879.091544.0074.05

We have also revised the abstract and introduction to clarify the scope of our comparison: “Across eleven understanding and generation benchmarks, UniFlux achieves strong performance, with leading results on multiple benchmarks among the evaluated unified models.”

W2 (a). Sources of Improvement and Parameter Accounting.

We thank the reviewer for the opportunity to clarify the sources of improvement and the model’s parameter accounting. Table 1 compares UniFlux’s overall capabilities with existing unified models using their respective training recipes. To identify the contributions of our design, we complement this benchmark comparison with controlled ablations of the pretrained initialization, Stage-I adaptation, and Stage-II objective. These experiments provide direct evidence about where the improvements arise. Pretraining and supervised adaptation. We perform experiments to distinguish the contributions of pretrained representations, supervised adaptation, and cross-modal reinforcement learning. With the MM-DiT backbone frozen and only the answer-token embeddings and vocabulary head trained, pretrained FLUX achieves 31.04 MMMU / 35.02 MathVista, compared with 9.44 / 6.67 for a randomly initialized frozen backbone. Both controls use the same training data, external text encoder, and VAE. These results show that pretrained MM-DiT representations provide useful information and that supervised adaptation contributes substantially to the final understanding capabilities.

Cross-modal optimization. Starting from the Stage-I checkpoint, positive-gated text-only RL achieves 66.54 MMMU / 0.85 GenEval, while X-GRPO achieves 68.82 / 0.91. Thus, cross-modal optimization provides gains of 2.28 percentage points on MMMU and 0.06 on GenEval beyond text-only RL. The comparison with signed text-only advantages and repeated GenEval evaluations is provided in our response on Positive Gating and Evaluation Noise.

ConfigurationMMMU ↑GenEval ↑
Stage I only62.030.79
Stage II: text-only RL with positive-gated advantages66.540.85
Stage II: X-GRPO68.820.91

Parameter accounting. The reported 9B and 32B values denote the MM-DiT backbone sizes. UniFlux-9B uses Qwen3-8B (approximately 8.2B parameters) as its text encoder, while UniFlux-32B uses Mistral-Small-3.2-24B-Instruct-2506 (approximately 24B parameters). These encoders are frozen, and their query-conditioning representations are precomputed and cached during Stage I. With cached query features available at inference, the encoders are not executed during decoding. Language responses are generated by the adapted MM-DiT text stream and vocabulary head. The parameter counts in Table 1 therefore refer to the MM-DiT backbones, with the auxiliary frozen encoder sizes specified here separately.

AuthorsOpenReview ↗

Response to reviewer f1j3 (2/4)

W2 (b). Reward Judge and Data Overlap.

We agree that reward-judge selection affects downstream performance. The text-only RL and X-GRPO comparison above uses the same Gemini-3.5 Pro judge, controlling for judge identity, where X-GRPO achieves 68.82 on MMMU, compared with 66.54 for text-only GRPO, an improvement of 2.28 percentage points. Comparing X-GRPO trained with GPT-5.1 Mini against text-only RL trained with Gemini-3.5 Pro changes both the method and the judge, so it does not isolate either effect. The reference answer is supplied to the judge for training-time reward computation. UniFlux generates its responses from the image and query and does not receive reference answers at evaluation.

To assess dependence on a proprietary judge, we additionally trained Stage II with Qwen3.6-VL-35B and evaluated all six understanding benchmarks:

ModelMS-COCO CIDErVQAv2 AccMME AccGQA AccMMMU AccMathVista
Gemini 3.5 Pro73.3376.78147273.1768.8280.43
Qwen3.6VL 35B72.0576.15146173.4467.5280.55

Using the open judge yields comparable performance across these benchmarks: MMMU decreases by 1.30 percentage points, while GQA and MathVista improve by 0.27 and 0.12 percentage points, respectively. These results demonstrate that UniFlux also achieves strong performance with an open reward judge.

Data overlap. Our training data are sourced from BAGEL, and we checked the training data against the evaluation examples but found no overlapping examples in the agent-assisted checks performed. The comparison was performed at the individual-example level, accounting for shared source datasets such as ChartQA, AI2D, and ScienceQA. No overlapping examples were identified through our screenings.

DatasetTypes of images
ChartQAReal-world vertical and horizontal bar charts, line graphs, and pie charts, containing numerical values, axis labels, legends, and category labels.
AI2DSchool-level scientific diagrams, such as food webs, life cycles, human anatomy, and moon phases.
ScienceQAA mixture of natural photographs and educational diagrams, including illustrations of physical phenomena, biological organisms, and scientific concepts. Its subject coverage extends beyond physics and chemistry to biology, geography, and other natural, social, and language-science topics.

W3 (a). X-GRPO Objective.

The reviewer is correct that a pair with zero coupled weight contributes no direct update through the reward surrogate. This is an intentional consequence of our credit-assignment rule (lines 258-261, Eq.2): positive reinforcement requires both text and image advantages to be positive. The absence of negative reward-surrogate updates is therefore intentional for X-GRPO’s cross-modal credit assignment. X-GRPO extends the GRPO-style objective with this specific coupling rule: positive reinforcement requires both modalities to exceed their respective group baselines. KL regularization remains active, and updates from other rollouts can still affect the probabilities of zero-weight responses.

The design is also supported empirically: Table 4 reports 63.12 MMMU / 0.82 GenEval for signed-product coupling, versus 68.82 / 0.91 for our positive-gated coupling. The proposition specifies the credit-assignment behavior, while the ablation evaluates its practical benefit.

Proposition F.2 characterizes both pair-level and response-level credit assignment. Beyond nonnegativity, it establishes the necessary and sufficient conditions for strictly positive credit, monotonicity with respect to each advantage, and the corresponding condition after aggregating across generated images. Specifically, an aggregated text response receives positive credit if and only if its text advantage is positive and at least one associated image has a positive advantage.

W3 (b). Image Reward.

We directly tested whether image feedback remains sensitive to response correctness when the instruction generator can access the source image. For 100 image–question pairs, we held the source image and question fixed and compared a correct response with a minimally altered incorrect response. We performed this intervention both with and without source-image access for the instruction generator.

In all conditions, the image-side VLM judge received the original image, the generated image, and the generation instruction. Although Figure 13 describes a general template with an optional reference image, the reference is always supplied in our X-GRPO experiments. HPSv3 and SigLIP provide complementary preference and instruction–image alignment signals; the VLM reward assesses fidelity to the source image. Instruction and generated-image factual accuracy are measured against the source image, while HPSv3 and SigLIP score each generated image against its corresponding generation instruction.

AuthorsOpenReview ↗

Response to reviewer f1j3 (3/4)

Source image supplied to instruction generatorResponseInstruction factual accuracy (%) ↑Generated-image factual accuracy (%) ↑HPSv3 ↑SigLIP ↑VLM reward ↑
YesCorrect89928.920.920.77
YesMinimally incorrect45343.150.210.19
NoCorrect87908.030.890.71
NoMinimally incorrect39312.970.150.18

Even with source-image access, changing the response reduces instruction factual accuracy from 89% to 45%, generated-image factual accuracy from 92% to 34%, and VLM reward from 0.77 to 0.19. Thus, in this controlled test, access to the original image does not make instruction construction or image feedback independent of response correctness. The source image provides additional visual context that may be omitted from the response. Removing this context reduces downstream MMMU performance from 68.82 to 66.88. Furthermore, full X-GRPO achieves 68.82 MMMU versus 66.54 for positive-gated text-only RL, an improvement of 2.28 percentage points under the same judge. Together, the ablation and response-perturbation results support retaining source-image context while showing that the feedback remains sensitive to the sampled response.

We have added this analysis in the Appendix E.

W3/Q3 (c). Positive Gating and Evaluation.

The understanding-only baseline uses positive-gated text advantages, [A^T]+[\widehat{A}_T]_+. We also evaluated vanilla text-only GRPO with signed advantages, as shown below.

ConfigurationMMMU ↑GenEval ↑
Text-only GRPO with signed advantages (vanilla)64.020.83
Text-only GRPO with positive-gated advantages66.540.85
X-GRPO68.820.91

Positive gating improves the text-only GRPO baseline, and X-GRPO provides a further 2.28-percentage-point improvement on MMMU and 0.06 improvement on GenEval over positive-gated text-only GRPO. Thus, the full method’s advantage over the understanding-only baseline cannot be attributed simply to introducing the positive gate.

This Experiment is added in Table 4 of revised manuscript.

Text-only optimization can also affect image generation because the two pathways share trainable MM-DiT parameters. “Text-only” specifies the reward objective; it does not freeze the shared parameters used for image generation. Consequently, improvements in generation under text-only training are consistent with transfer through the shared backbone. To assess evaluation noise, we repeated GenEval evaluation five times for each fixed trained checkpoint:

ConfigurationGenEval, mean ± standard deviation
Stage I0.80 ± 0.0190
Text-only GRPO with positive-gated advantages0.84 ± 0.0220
Image-only GRPO0.86 ± 0.0170
DanceGRPO0.82 ± 0.0422
FlowGRPO0.88 ± 0.0271
X-GRPO0.92 ± 0.0252

X-GRPO achieves a mean GenEval of 0.92 ± 0.025, compared with 0.84 ± 0.022 for positive-gated text-only GRPO and 0.86 ± 0.017 for image-only GRPO. These gaps of 0.08 and 0.06 are roughly three times the observed standard deviations and persist after averaging over five repeated evaluations. The difference between text-only and image-only GRPO (0.02) is within evaluation noise, whereas X-GRPO’s improvement over both is not. These results provide quantitative evidence against attributing X-GRPO’s GenEval improvement to evaluation noise.

W4. Generation preservation.

By preserving generation capability, we mean retaining the pretrained image-generation pathway while extending the model to support understanding and reasoning. This does not require keeping its parameters frozen and it refers to the full model after all stages. Stage I changes GenEval from 0.81 to 0.79, while Stage II jointly optimizes the shared model and increases it to 0.91. Thus, the complete method improves generation performance over the pretrained backbone while adding multimodal understanding. We will clarify the distinction between retaining generation capability and maintaining an identical score throughout training.

Backbones and baseline results. Our Flow-GRPO and DanceGRPO results were obtained by running these methods on FLUX.2 [klein] 9B using our training data. The reported 0.89 is therefore the result of our Flow-GRPO implementation under the same configuration. The original Flow-GRPO paper’s 0.95 result uses SD3.5-Medium under its own training configuration. These results correspond to different experimental settings.

Figures 22 and 23. The “FLUX-2 4B” labels were typographical errors. These examples were generated using FLUX.2 [klein] 9B, consistent with the backbone used for UniFlux, Flow-GRPO, and DanceGRPO. The labels have been corrected. These figures provide additional qualitative comparisons at the same model scale.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.