- 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
- The experimental results are quite convincing.
- The paper is weel-written.
Critical weaknesses
- 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.
- 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.
- All experiments were conducted using zimage, with structural text sourced from Qwen; at least one additional MMDIT result should be included.
Questions for authors
- What is the difference between NitroFusionv2 and NitroFusionv1?
- What methodological gains does the new job bring to the community
6 comments
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,
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.