NitroFusion2: Bringing Adversarial Loss to the Text Stream of Multimodal Diffusion Transformers
Organizations: SketchX, CVSSP, University of Surrey · Turing Elite Research Labs
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.
Figures & tables
| Text rendering | Overall | ||||||
| Method | Params | NFE | LongText-EN | LongText-CN | GenEval | DPG-Bench | OneIG |
| NitroFusion2 (ours) | 6B | 1 | 43.34 | 47.78 | 75.86 | 86.48 | 47.63 |
| Z-Image-Turbo [ 77 ] | 6B | 1 | 11.72 | 12.70 | 57.08 | 77.23 | 32.43 |
| TwinFlow [ 8 ] | 6B | 1 | 29.18 | 36.85 | 66.34 | 85.17 | 42.20 |
| NitroFusion2 (ours) | 6B | 2 | 76.03 | 81.95 | 74.46 | 86.21 | 51.45 |
| Z-Image-Turbo [ 77 ] | 6B | 2 | 72.89 | 78.48 | 76.53 | 85.78 | 50.94 |
| Setting | LongText-EN | LongText-CN | GenEval | DPG-Bench | OneIG |
|---|---|---|---|---|---|
| Text | 22.31 | 24.37 | 66.71 | 84.39 | 38.45 |
| Image | 17.86 | 19.26 | 66.00 | 82.87 | 35.79 |
| Text + image | 24.47 | 26.63 | 64.39 | 84.86 | 40.02 |
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Cue | Text the discriminator is conditioned on |
|---|---|
| Full caption | Every field concatenated: the complete specification of the image. |
| Typography | The rendered text, transcribed line by line with each line’s position and lettering. Populated only when the image contains visible text. |
| Object | The subject: what is depicted, with the counts and identities of the entities preserved. |
| Style | The palette, artistic medium and rendering treatment, describing how the image is rendered rather than what it shows. |
| Background | The setting the subject sits in: surroundings, surfaces, weather and ambient detail, excluding the subject itself. |
| Action | What the subject is doing, or an explicit statement that the scene is static. |
| Cue | Top four, best first | Selected |
|---|---|---|
| Full caption | ||
| Typography | ||
| Object | ||
| Style | ||
| Background | ||
| Action |
| Text-stream taps | LongText-EN | LongText-CN | GenEval | DPG-Bench | OneIG |
|---|---|---|---|---|---|
| Selected ( Table 6 ) | 24.47 | 26.63 | 64.39 | 84.86 | 40.02 |
| Top four | 21.26 | 25.17 | 64.91 | 83.40 | 36.53 |
| Evenly spaced, | 22.56 | 24.67 | 67.62 | 84.74 | 38.71 |
| Method | Params | NFE | Overall | Alignment | Text | Diversity | Style | Reasoning |
|---|---|---|---|---|---|---|---|---|
| NitroFusion2 | 6B | 1 | 47.63 | 85.13 | 76.28 | 8.50 | 43.00 | 25.24 |
| Z-Image-Turbo | 6B | 1 | 32.43 | 65.57 | 29.79 | 12.68 | 31.37 | 22.75 |
| TwinFlow | 6B | 1 | 42.20 | 81.68 | 54.13 | 14.97 | 36.10 | 24.12 |
| NitroFusion2 | 6B | 2 | 51.45 | 85.41 | 91.34 | 9.63 | 44.20 | 26.68 |
| Z-Image-Turbo | 6B | 2 | 50.94 | 83.92 | 92.85 | 13.85 | 36.25 | 27.86 |
| TwinFlow | 6B | 2 | 51.38 | 85.26 | 90.61 | 15.49 | 38.03 | 27.52 |
| Method | Params | NFE | Overall | Entity | Attribute | Relation | Global | Other |
|---|---|---|---|---|---|---|---|---|
| NitroFusion2 | 6B | 1 | 86.48 | 92.36 | 89.62 | 94.74 | 84.19 | 88.00 |
| Z-Image-Turbo | 6B | 1 | 77.23 | 84.85 | 83.08 | 91.03 | 76.90 | 82.80 |
| TwinFlow | 6B | 1 | 85.17 | 90.99 | 88.56 | 94.43 | 83.89 | 87.20 |
| NitroFusion2 | 6B | 2 | 86.21 | 92.47 | 88.94 | 93.81 | 83.28 | 86.00 |
| Z-Image-Turbo | 6B | 2 | 85.78 | 91.73 | 88.78 | 94.16 | 84.19 | 89.60 |
| TwinFlow | 6B | 2 | 85.98 | 91.83 | 89.02 | 94.16 | 81.46 | 87.60 |
| Text rendering | Overall | ||||||
|---|---|---|---|---|---|---|---|
| Method | Params | NFE | LongText-EN | LongText-CN | GenEval | DPG-Bench | OneIG |
| NitroFusion2 (ours) | 6B | 1 | 43.34 | 47.78 | 75.86 | 86.48 | 47.63 |
| Qwen-Image-turbo-2steps | 20B | 1 | 40.44 | 51.50 | 69.34 | 86.06 | 40.03 |
| Krea2-turbo-2steps | 12B | 1 | 13.86 | 3.67 | 51.24 | 74.62 | 30.47 |
| TwinFlow | 6B | 1 | 29.18 | 36.85 | 66.34 | 85.17 | 42.20 |
| NitroFusion2 (ours) | 6B | 2 | 76.03 | 81.95 | 74.46 | 86.21 | 51.45 |
Diffusion Model Distillation
Diffusion model distillation transfers a large model’s generative behavior to faster or smaller models that need fewer sampling steps.
See all 213 papers