acceptodds
Under review as a conference paper at ICLR 2027

Bringing Adversarial Loss to the Text Stream of Multimodal Diffusion Transformers

14 views

Abstract

Adversarial diffusion distillation has conventionally relied on image features for discrimination. We revisit this choice for multimodal diffusion transformers (MMDiTs), whose text stream becomes image-dependent through joint attention. We introduce Text-Stream Adversarial Distillation (TSAD), which exploits intermediate text-token states as a new adversarial representation. Because the text stream has already read the image while remaining tied to the text condition, changing that condition changes what information it extracts from the image. Building on this property, we introduce Aspect-Cued Discrimination (ACD), which conditions the discriminator on different aspects of the sample, such as typography, objects, or style, to steer which properties are emphasized during adversarial training. Building on the expert-discriminator design of NitroFusion, we integrate TSAD and ACD in NitroFusion2. NitroFusion2 brings few-step distillation from the conventional 4–8-step regime down to just 1–2 steps at comparable quality. At one step, it produces sharp rendered text and rich fine-grained details, making one-step generation practical rather than merely a stress test for distillation. At two steps, it remains competitive across the evaluated benchmarks. These results establish the MMDiT text stream as a powerful and steerable representation for adversarial diffusion distillation.

Primary area:
generative models
Keywords:
diffusion models, text-to-image generation, diffusion distillation, adversarial distillation, few-step generation, one-step generation, multimodal diffusion transformer, MMDiT, text-stream discrimination

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_wfMTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This study introduces the idea of adversarial distillation in this article, using intermediate text tokens to guide adversarial representations to resist interference and assist diffusion models. Furthermore, the ACD module can further improve the generation quality. The system engineering demo in this article is well done.

Critical strengths
  1. The experimental results are quite convincing.
  2. The paper is weel-written.
Critical weaknesses
  1. The primary claim of this paper is that the text stream input discriminator outperforms the image-based approach; however, the current experimental results do not sufficiently support this conclusion. The results in Table 3 exhibit two issues: first, only 1,000 steps were trained, which differs from the final full training configuration; second, the three experimental setups are not aligned, making it difficult to pinpoint the exact source of the performance improvement—for instance, the fact that the "Text-Image" setup likely involves two heads and requires alignment of parameters is not clearly documented. Table 4 is even more problematic, as it essentially lacks controlled variables.
  2. The discriminator can access unique detail information from the reference image, whereas the generator only receives the original prompt, which may lead to a misalignment of the training objective: even when generated samples satisfy the original prompt, they may still be classified as fake samples by the discriminator due to inconsistencies in details unrelated to the reference image. It is necessary to clarify whether the training objective is to optimize generation quality or to enforce the faithful reproduction of reference image details not specified in the original prompt, and to supplement with ablation studies to differentiate the effects of various cue types on both generation quality and diversity.
  3. All experiments were conducted using zimage, with structural text sourced from Qwen; at least one additional MMDIT result should be included.
Questions for authors
  1. What is the difference between NitroFusionv2 and NitroFusionv1?
  2. What methodological gains does the new job bring to the community
6 comments
AuthorsOpenReview ↗

Response to Reviewer wfMT (1/2)

We appreciate the kind words on the experimental results and the writing. The two questions go to the heart of the paper, so we begin with them. Each of the three weaknesses led to a new experiment in the general response.

Q1. What is the difference between NitroFusion2 and NitroFusion?

NitroFusion focuses on the design of the expert discriminator. It introduces dynamic expert ensembles to stabilise adversarial distillation, and it relies on conventional image features. NitroFusion2 addresses a different question, namely what the discriminator reads and what it is asked to judge. Its two contributions, TSAD and ACD, are general. By design, they can be applied to any adversarial training framework built on an MMDiT and are not limited to the NitroFusion-style design.

The main contribution is TSAD, a new discriminative representation. We identify the image-dependent text stream of MMDiTs as an effective adversarial representation. Any method that attaches discriminator heads to an MMDiT backbone can read the text stream in the same way.

The second contribution is aspect-cued discrimination (ACD). Building on TSAD, ACD conditions the discriminator on a specific aspect of the sample, such as typography, object, or style, which makes the adversarial supervision more interpretable and steerable. This idea is also independent of the expert design. In NitroFusion2 we combine it with the NitroFusion experts, which had no predefined semantic role, so that each expert group receives a clear semantic focus.

Q2. What methodological gains does the new work bring to the community?

We see two gains that reach beyond this particular model.

Text has traditionally been viewed as a conditioning signal for image generation, while training objectives have relied on image-stream features. Our work shows that the image-dependent text stream of an MMDiT can itself provide effective adversarial supervision to the image generator, which gives adversarial training a new source of supervision. We hope this also encourages the exploration of text-stream representations in other training objectives, such as REPA-style feature matching and representation alignment.

Adversarial training has traditionally aimed to improve image fidelity, but which visual properties the discriminator emphasises usually remains implicit. ACD offers a way to make this supervision explicit and steerable, by directing the discriminator toward specific semantic aspects. This makes it possible to study how different aspects contribute to generation quality.

AuthorsOpenReview ↗

Response to Reviewer wfMT (2/2)

W1. Controls in Table 3 and Table 4.

This careful reading of the ablations was very helpful, and we address its three points in turn.

On the alignment, the three settings of Table 3 differed in head count and in discriminator input. Table R3 in the general response lists both for each setting, and Table R4 repeats the comparison with the original prompt p as the discriminator input and 6 heads in every setting. Under these matched conditions, the text-only setting scores higher than the image-only setting on LongText-EN (25.54 against 18.54), LongText-CN (31.80 against 19.06), DPG-Bench, and OneIG, while GenEval is lower. We will replace Table 3 with Table R4 in the revision.

On the training length, we understand the concern that a short schedule may not reflect the final ranking. We chose 1,000 steps so that we could compare many settings under the same budget. By that point the student has already moved far from its starting point (one-step LongText-EN rises from 11.72 to 24.47), and all settings are compared at the same stage of training.

On Table 4, we apologise that the setup was unclear. All of its rows use the joint text- and image-stream discriminator with the same schedule. The four single-cue rows (Typography, Style, Object, and Full caption) each have 3 text-stream and 3 image-stream heads, so they differ only in the cue, and the cue clearly changes the result. The typography cue, for example, is far ahead in text rendering. This ablation alone cannot show that our setting is the best one. What it does show is that the cue steers what the student learns, and it invites further thinking on how to combine the cues for an even better balance.

W2. Training objective and the reference-derived cues.

This is an important point. Our intention is to improve generation quality at the level of the distribution, and not to reproduce individual reference images. The discriminator does not compare a student sample with a reference image directly. It scores one noised image at a time under a cue, and the same cue conditions both the reference pass and the student pass.

At the same time, this concern points to a real effect. The cue is written from the reference image, so it can contain details that the original prompt does not. The adversarial signal encourages the student to produce such details, which is the intended effect, since it leads to richer detail and better alignment. However, the student only receives the original prompt, so it tends toward the details that are most typical for that prompt, and this can reduce diversity. Following this suggestion, we ran the ablation in Table R2. It shows this trade-off. The original prompt preserves diversity, and the full caption reduces it. We are grateful for this suggestion, as it gave us a clearer picture of where the diversity loss comes from.

W3. Results on an additional MMDiT.

We agree that a second backbone makes the paper more convincing, and we appreciate the suggestion. We distilled FLUX.2 [klein] 4B using the text-stream adversarial loss only and no image-stream heads (Table R1 in the general response). At one step, all five metrics are at or above the base model, most clearly GenEval (58.53 to 65.08), and four of five at two steps. This indicates that the text stream can provide a useful adversarial signal beyond Z-Image-Turbo.

Joint attention makes the text tokens image-dependent in every MMDiT, so text-stream discrimination can be applied to any of them. These tokens are also tied to the words of the conditioning text, so the adversarial signal carries explicit semantics, and changing that text steers what the discriminator emphasises. This makes adversarial training steerable in an explicit and interpretable way, and we see it as the core benefit of text-stream discrimination.

We hope this reply addresses the concerns raised, and we would be very happy to discuss any further questions.

Thank you for the author's sincere response. Most of the concerns have been resolved, however, you only respond W1 halfway. This can indicate that 1000 steps are not "untrained", but it cannot prove that the method ranking remains unchanged after complete training. What I am questioning is the difference between the short program and the final configuration,

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

This paper introduces Text-Stream Adversarial Distillation (TSAD), which uses intermediate text-token representations from a frozen multimodal diffusion transformer for adversarial supervision. Through joint attention, these representations incorporate image information while remaining associated with the conditioning text. The paper further proposes Aspect-Cued Discrimination (ACD), which uses image-grounded descriptions of properties such as typography, objects, and style to steer the discriminator’s emphasis. Integrated with image-stream discrimination and the expert-head design of NitroFusion, these components form NitroFusion2. Experiments demonstrate improved one-step text rendering, compositional generation, and overall image quality, together with competitive two-step performance.

Critical strengths
  1. Clear motivation and an interesting approach. The paper identifies image-dependent text-stream representations in MMDiTs as a useful source of adversarial supervision. This idea is well motivated by the interaction between image and text tokens through joint attention. Aspect-Cued Discrimination further provides an intuitive way to emphasize specific properties, including typography, objects, and style.

  2. Strong empirical results and informative ablations. NitroFusion2 outperforms Z-Image-Turbo and TwinFlow on all five reported one-step benchmark metrics, covering English and Chinese text rendering, compositional generation, and overall image quality. The experiments also compare image-stream, text-stream, and joint discrimination, while the cue ablations demonstrate aspect-specific improvements and the benefits of combining multiple cues.

Critical weaknesses
  1. Insufficient analysis of the quality–diversity trade-off. Table 8 shows that NitroFusion2 achieves a one-step Diversity score of 8.50, compared with 12.68 for Z-Image-Turbo and 14.97 for TwinFlow, and this deficit persists at two steps. Although the authors acknowledge reduced diversity, the paper does not examine how this trade-off varies with training progress, adversarial loss weights, or cue configurations. Further analysis would help clarify which components contribute to this limitation and whether it can be mitigated.

  2. Limited validation across model backbones. Both the student and the frozen discriminator backbone are initialized from Z-Image-Turbo, concentrating the methodological validation within a single model family. The additional comparisons with Qwen-Image and Krea evaluate competing generators without applying TSAD or ACD to those backbones. Consequently, it remains unclear whether the observed advantage of text-stream over image-stream discrimination generalizes across different MMDiT architectures. A controlled experiment on another backbone would strengthen the broader applicability claims.

Questions for authors
  1. Have you investigated whether adjusting the adversarial loss weights or cue-sampling strategy can recover diversity while preserving the quality improvements?

  2. Have you evaluated TSAD or ACD on another MMDiT backbone? What evidence suggests that the advantage of text-stream over image-stream discrimination extends beyond Z-Image-Turbo?

3 comments
AuthorsOpenReview ↗

Response to Reviewer pGNa

We are grateful for the encouraging assessment. It means a lot to us that the motivation was found clear, the approach interesting, and the results and ablations informative. Both suggestions led to new experiments (Tables R1 and R2 in the general response).

Q1. Can adjusting the adversarial loss weights or the cue-sampling strategy recover diversity while preserving the quality improvements?

This is a valuable suggestion, and we are glad to add an analysis of the reduced diversity.

The text that conditions the discriminator has a clear effect on diversity. Table R2 in the general response compares two ways of conditioning the discriminator. When the discriminator is conditioned on the original prompt, diversity is preserved (12.77, against 12.68 for the base model Z-Image-Turbo). The full caption lowers it to 10.09, because it describes many details of one reference image at once, and this pushes the student toward the details that are most typical for a prompt. The conditioning text is therefore one handle on the balance between quality and diversity. This ablation was also suggested by Reviewer wfMT.

For the loss weights, we determined them empirically in our early experiments. The flow-matching loss works only as a regulariser in the initial stage, where the student is still far from the data distribution, and it is discarded in the second stage, which uses the adversarial loss alone (Appendix A).

Among the factors we examined, the conditioning text is the one with a clear effect on diversity (Table R2). We find this suggestion inspiring. How to preserve diversity during adversarial training is an important question in its own right, and we will study it in future work.

Q2. Has TSAD or ACD been evaluated on another MMDiT backbone? What evidence suggests that the advantage of the text stream extends beyond Z-Image-Turbo?

Following this suggestion, we ran an experiment on FLUX.2 [klein] 4B, with adversarial training on the text stream alone (Table R1 in the general response). All five metrics improve at one step, and four of the five at two steps, which suggests that the text-stream signal is also useful outside Z-Image-Turbo.

A text-stream discriminator reads the sample through the words of its prompt, so the same discriminator can be asked to judge typography under one cue and objects under another. This ability to direct the critic is what text-stream adversarial training adds.

We hope this reply addresses the concerns raised, and we would be very happy to discuss any further questions.

Thank you for the response and the additional experiments. My concerns have been addressed, and I believe this is a strong piece of work. I will maintain my positive rating.

One small suggestion: including the relevant numerical results directly in the response to each reviewer would make the rebuttal easier to follow and reduce the need to switch back and forth with the general response.

AuthorsOpenReview ↗

Thanks to Reviewer pGNa

We are very grateful for the kind words and for the continued support. It is encouraging to hear that the concerns have been addressed, and the two suggestions made the paper stronger.

Reviewer_YDGfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes NitroFusion2, which introduces TSAD to enable the discriminator to leverage a text stream that evolves through interactions with images, and ACD to provide additional conditioning that guides the model to focus on specific image attributes. These components allow the model to generate images of reasonably good visual quality in fewer steps than existing methods.

Critical strengths

The paper leverages text-stream representations that incorporate image information through MMDiT’s joint attention for adversarial discrimination. Comparisons with conventional image-stream-based discrimination empirically demonstrate that text-stream features can provide useful discriminative signals. The results show that performance varies depending on the aspect cue, with particularly strong performance in typography being a notable finding. Overall, the method achieves strong performance with fewer parameters and sampling steps, outperforming competing methods under comparable settings.

Critical weaknesses

The proposed TSAD extends the discriminator used in existing adversarial diffusion distillation methods to also leverage features extracted from MMDiT’s text stream. Its technical distinction from prior methods therefore appears limited. In particular, the incorporation of image information into the text stream through MMDiT’s joint attention is an inherent property of the existing architecture. Simply using these representations as input features for the discriminator does not, in my view, establish sufficient methodological novelty. Similarly, ACD appears to extend the expert discriminator architecture used in the original NitroFusion through aspect-specific text conditioning and head selection. Although performance improvements from aspect cues are demonstrated, the technical contribution beyond a straightforward conditioning-based extension of the existing method remains unclear.

Questions for authors
  1. Table 4 appears to report results for only a subset of the aspect cues. Are results available for the remaining cues, and do they support the same conclusions?

  2. Since the proposed method extends NitroFusion, how does it compare directly with the original NitroFusion under matched experimental settings?

  3. What are the results for TSAD-only and ACD-only variants? Separate ablations would help establish each component’s individual contribution and whether their combination provides additional benefits.

  4. Figure 2 does not clearly illustrate the TSAD component. Could you clarify where text-stream features are extracted and how they are used for adversarial discrimination?

5 comments
AuthorsOpenReview ↗

Response to Reviewer YDGf (1/2)

We thank the reviewer for the thoughtful review, and for recognising the text stream as a useful discriminative signal and the strong performance with fewer parameters and steps. The questions helped us state our contribution more clearly.

W. Technical contribution of TSAD and ACD.

The insight we hope to share is that, in an MMDiT, the text stream is no longer only a read-only input condition. After joint attention, it becomes a discriminative representation in its own right, one that carries rich information about the image and can be controlled through the text. This was not possible in earlier architectures. In a cross-attention UNet, the text representation is computed independently of the image, so a discriminator attached to the text states gives the student no adversarial gradient (Section 3.2). Joint attention is what makes the text stream usable as an adversarial representation. TSAD turns this representation into adversarial supervision, and ACD uses the text given to the discriminator to decide what it judges. Both are general by design and can be used by any adversarial training framework built on an MMDiT. We also see this empirically. Text-stream heads alone already provide an effective adversarial signal on Z-Image-Turbo (Table R4 in the general response), and the same holds on a second MMDiT, FLUX.2 [klein] 4B, without any image-stream adversarial training (Table R1).

We hope the paper encourages further study of text representations as a supervision signal, for example how to balance the cues adaptively, and whether the same idea helps in broader multimodal generation and control tasks.

AuthorsOpenReview ↗

Response to Reviewer YDGf (2/2)

Q1. Table 4 reports only a subset of the aspect cues. Are results available for the remaining cues?

We admit that Table 4 covers only a subset of the cues. It reports typography, style, and object, since existing benchmarks measure these aspects most directly, while background, action, and layout are harder to identify through existing metrics.

The attention maps in Figure 3 show that a different text input to the discriminator changes what the text stream attends to in the image, which is the mechanism every cue relies on. Table 4 then shows empirically that the discriminator, and with it the adversarial training, is steerable through the cue. This opens an interesting question of how to best construct a cue for a desired aspect. We believe the answer is connected to how image and text interact inside an MMDiT, which has long been underexplored.

Q2. How does the method compare directly with the original NitroFusion under matched settings?

We use our image-only setting as the NitroFusion-style baseline, with image-stream features, the expert head pool, and the discriminator conditioned on the same original prompt as the student. Table 3 reports it, but its rows differ in head count and in the discriminator input, which Reviewer wfMT also pointed out, so we ran a new experiment for a fair comparison (Table R4).

With the same prompt and head count, text-stream adversarial training scores higher than the NitroFusion setting on LongText-EN, LongText-CN, DPG-Bench, and OneIG, while GenEval is lower. Combining the two streams further improves text rendering and OneIG, with lower GenEval.

Q3. What are the results for TSAD-only and ACD-only variants?

The TSAD-only variant is the "Text" row of Table R4. It uses text-stream heads only, with the original prompt as the discriminator input and no aspect cue. Compared with the image-only setting, it improves text rendering, DPG-Bench, and OneIG, and is behind only on GenEval.

To isolate ACD in the ablations, Table R5 below extends the single-cue rows of Table 4 with the original prompt p, which uses no reference-derived cues. Only the conditioning text changes across the rows. Compared with the original prompt, the typography cue is higher on text rendering, the style cue on style, and the object cue on the object metrics. Together, these rows show that the cue given to the discriminator decides which aspect the student improves on, so ACD offers a direct way to control what adversarial training emphasises.

Table R5. Single cues against the original prompt (1,000 steps, 1 NFE).

Text renderingStyleObjectAlignment
CueLT-ENLT-CNOneIG-StyleGenEval-SingleDPG-EntityOneIG-Align.DPG-Global
Original prompt p (w/o ACD)28.5534.3932.9399.0691.6276.1079.17
Typography45.0653.0132.6699.3892.5278.4879.33
Style16.0919.2833.7598.4492.3676.6279.60
Object20.8225.1932.67100.0092.9479.1478.72

Q4. Figure 2 does not clearly illustrate TSAD. Where are the text-stream features extracted, and how are they used?

The text-stream features are the attention outputs of the text tokens, UTU_T (defined after Eq. (3)), taken from four joint-attention layers of the discriminator backbone (Section 3.2; the layers are listed in Table 6). Lightweight discriminator heads, which follow the design of APT (Appendix A), take these features as input and are trained with the hinge loss of Eq. (4) to predict real or fake.

We apologise that Figure 2 and its caption do not show this clearly enough, and we appreciate this being pointed out. We have updated the caption as follows.

Figure 2 (updated caption). Architecture of Text-Stream Adversarial Distillation (TSAD) and Aspect-Cued Discrimination (ACD). The noised reference image or student sample passes through the frozen MMDiT discriminator backbone. TSAD takes the attention outputs of the text tokens at selected joint-attention layers and feeds them to lightweight text-stream heads [33] (APT, Lin et al., ICML 2025), used alongside the image-stream heads, that predict real or fake. ACD builds image-grounded aspect fields, then samples one cue per training step. The cue conditions the MMDiT discriminator backbone and selects the corresponding text-stream head, steering discrimination toward that aspect.

We hope this reply addresses the concerns raised, and we would be very happy to discuss any further questions.

AuthorsOpenReview ↗

Follow-up to Reviewer YDGf

We would like to follow up, as the discussion period is coming to a close.

Our response covers the technical contribution of TSAD and ACD, the aspect cues in Table 4, the comparison with the NitroFusion setting (Table R4), the separate ablations of the two components, and the updated caption of Figure 2.

We would be grateful to hear whether the concerns have been addressed, and we are happy to clarify anything that remains unclear in the time left.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all three reviewers for their time and their careful, constructive feedback. We are encouraged that the reviewers recognised the core idea of the paper, Text-Stream Adversarial Distillation (TSAD), in which the image-dependent text stream of an MMDiT provides a useful signal for adversarial supervision (pGNa, YDGf). The reviewers described the experimental results as convincing (wfMT) and strong (pGNa, YDGf). We also appreciate the comments on Aspect-Cued Discrimination (ACD) as an intuitive way to emphasise specific properties (pGNa), on the notable typography result (YDGf), on the clear motivation and informative ablations (pGNa), on the writing (wfMT), and on the strong performance with fewer parameters and sampling steps (YDGf). The reviews also raised three shared concerns. They ask whether the method works on another MMDiT, how the reference-derived cues affect the balance between quality and diversity, and whether the comparison between the text stream and the image stream is properly controlled.

We ran additional experiments on each of these and present them in the next comment; our replies to each reviewer refer back to it. Tables with the "R" prefix are new; tables without it refer to the submission.

AuthorsOpenReview ↗

Additional Experiments and Planned Revisions

Generalisation to a second MMDiT

To test whether text-stream discrimination extends beyond Z-Image-Turbo, we distil FLUX.2 [klein] 4B. Adversarial training is applied to the text stream only, with all seven cues and no image-stream adversarial loss. At one step, all five metrics improve over the base model; at two steps, four of the five improve, while OneIG decreases slightly. This indicates that text-stream adversarial training alone can improve a second MMDiT.

This is a preliminary experiment. It runs for 3,000 steps and reuses the dataset generated by Z-Image-Turbo, so the reference distribution does not match the student's model family, and the improvement is not as significant as that in Table 1. We find it encouraging that text-stream adversarial training already helps under these conditions.

Table R1. Text-stream discrimination on FLUX.2 [klein] 4B (TSAD and ACD on the text stream only, with no image-stream heads).

SettingNFELongText-ENLongText-CNGenEvalDPG-BenchOneIG
FLUX.2 [klein] 4B14.690.4958.5371.3527.96
+ TSAD + ACD16.080.7065.0871.5428.62
FLUX.2 [klein] 4B241.945.0079.4583.0140.93
+ TSAD + ACD243.085.5780.5883.6440.15

Discriminator conditioning and the quality–diversity trade-off

The discriminator is conditioned on the detailed full caption c derived from the reference image, while the student only receives the original prompt p. To study how this affects diversity, we vary this conditioning text.

In Table R2, with the original prompt, diversity is preserved (12.77, against 12.68 for the base model). With the full caption, diversity drops, while GenEval, DPG-Bench, and alignment improve. Our intuition is that the full caption describes details not in the original prompt, so the adversarial signal pushes the student toward the most typical details. This narrows the variation among samples, but the student also learns richer detail and follows the prompt more closely.

The original prompt gives the highest OneIG-Text. The reference images of this ablation contain rendered text, and a large portion of their original prompts describes that text, so conditioning on the prompt concentrates on typography. This is consistent with Table 4, where the typography cue gives the highest text-rendering scores, and it shows the same mechanism at work, in which the text given to the discriminator changes the behaviour of the student.

Table R2. Effect of the discriminator conditioning on quality and diversity (1,000 steps, 1 NFE).

Discriminator conditioningGenEvalDPG-BenchOneIGOneIG-TextOneIG-Align.OneIG-Diversity
Z-Image-Turbo (base)57.0877.2332.4329.7965.5712.68
Original prompt p62.2483.5340.7556.0876.1012.77
Full caption c63.9284.5037.0642.7276.9610.09

Controlled comparison of the two streams

The image stream and each cue carry a pool of three discriminator heads (Appendix A), so the three settings of Table 3 differ in their number of heads (Table R3). The image-only setting is conditioned on the original prompt p, as is standard in adversarial distillation.

Table R3. Discriminator input and head count of the settings in Table 3.

SettingDiscriminator inputHeads
ImageOriginal prompt p3
TextAll cues21
Text + imageAll cues24

For a fair comparison, Table R4 fixes the discriminator input to the original prompt p and the number of heads to 6, so that both are matched across the streams. The text-only setting scores higher than the image-only setting on LongText-EN, LongText-CN, DPG-Bench, and OneIG, while GenEval is lower. Combining the two streams gives the best text rendering and OneIG, with lower GenEval. On Z-Image-Turbo, text-stream adversarial training therefore yields a student with better text rendering (LongText-CN 31.80 against 19.06), which is one of the most challenging tasks in image generation.

Table R4. Comparison of the two streams under the original prompt (1,000 steps, 1 NFE).

SettingHeadsLongText-ENLongText-CNGenEvalDPG-BenchOneIG
Image618.5419.0666.4082.7635.88
Text625.5431.8064.8384.2139.67
Text + image3 + 328.5534.3962.2483.5340.75

Planned Revisions

We will replace Table 3 with Table R4 and update the related discussion and statements throughout the paper accordingly. We will add Table R2 and its discussion to the supplementary material. The FLUX.2 [klein] experiment in Table R1 is still preliminary, and we will complete it and include it in the supplementary material as well.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.