Prompt engineering is commonly used to encourage text-to-image models to generate more detailed images. However, it does not provide reliable, continuous, or disentangled control. We analyze feature tokens produced with and without detail-enhancing prompts across six diffusion transformers and observe a common pattern: these prompts tend to increase the entropy of each token's energy distribution while largely preserving its ordering. That is, flattening the distribution increases the level of detail. Motivated by this counter-intuitive finding, we introduce a simple training-free approach that continuously adjusts the level of detail according to a scalar. Specifically, it shifts the squared channel magnitudes of each token toward a higher entropy distribution with a time-aware scheduler to concentrate the intervention at intermediate sampling steps. Across extensive experiments on different types of diffusion transformers, we show that our approach provides continuous control over detail metrics while largely preserving image quality and semantic structure. It requires only a few lines of code and adds negligible inference time overhead.
Figures & tables
Figure 1: Qualitative examples of our method across two representative diffusion transformers. Our method controls detail richness through a user-specified scalar α as a lightweight plugin.
Figure 2: Overview of the analysis experiment described in section 3 . Given the same image latent, we compare DiT features obtained using an original prompt and its detail-guided version. For corresponding token pairs, detail instructions produce higher entropy over squared channel magnitudes.
Model
H(p)
Added text
H(p^)
Aall
Aprompt
Spearman
FLUX.2 Klein 9B Base
4.463
content
4.460
46.0%
45.6%
0.956
detail
4.530
78.0%
100.0%
0.965
FLUX.2 Klein 9B
4.307
content
4.292
39.7%
37.6%
0.799
detail
4.433
83.7%
100.0%
0.878
SD3.5 Large
3.065
content
3.070
59.2%
66.6%
0.929
detail
3.080
71.4%
98.2%
0.966
Table 1: For each model, we first report the baseline mean entropy ( H(p) ) of the token density defined in eq. 1 . Under each type of added text condition, we report the mean entropy ( H(p^) ), together with pairwise statistics computed between spatially corresponding tokens of baseline: the entropy increasing agreement ratios ( eqs. 2 and 3 ) and the mean Spearman correlation. Bold values indicate the patterns observed under detail instructions as discussed in section 3 .
Figure 3: Illustration of the T ♭ pseudo-code corresponding to eqs. 1 , 4 , 5 and 6 . T ♭ is injected into the residual branch of each DiT block.
Figure 4: Qualitative effects of applying T ♭ at a single sampling step with the same positive weights using FLUX.2 Klein 9B. At early steps, our method tends to enhance low-frequency details, such as adding objects; but it may also alter the global semantic structure if applying it too early. At later steps, T ♭ introduces more localized, high-frequency details, such as subtle freckles on faces and coarse fabric textures. That motivates our time-aware scheduler design in section 4 .
Input α
Image Quality
Detail Statistics
Structure Difference
CLIP ↑
HPSv3 ↑
Edge
Sharpness
f95
Wavelet
DINO
LPIPS 128
Qwen Image
−0.15
0.2771
11.212
0.0170
0.0048
0.0336
0.0196
0.1146
0.1517
−0.1
0.2774
11.474
0.0171
0.0051
0.0357
0.0209
0.0922
0.1251
−0.05
0.2773
11.749
0.0173
0.0055
0.0381
0.0227
0.0645
0.0846
0 (baseline)
0.2769
12.011
0.0182
0.0063
0.0435
0.0268
0.0000
0.0000
Table 2: T ♭ results across three diffusion transformer models and three metric groups over a range of α values. ↗α indicates a monotonic increase with respect to the given α s. T ♭ provides continuous detail control while maintaining image quality and causing only small structural changes.
Figure 5: A qualitative example of T ♭ from the experiment in table 2 across three models. Although the prompt already contains detail instructions, T ♭ further enhances visual details, demonstrating that it is orthogonal to prompt engineering. Further examples are provided in section B.8 .
Method
Latency over Baseline ↓
Image Quality
Detail Statistics
Structure Difference
CLIP ↑
HPSv3 ↑
Edge ↑
Sharpness ↑
f95↑
Wavelet ↑
DINO ↓
LPIPS 128↓
Baseline
1.00 ×
0.2750
11.0413
0.0347
0.0268
0.0682
0.0407
0.0000
0.0000
SAG
1.56 ×
0.2740
11.0230
0.0389
0.0241
0.0776
0.0491
0.0690
0.1286
PAG
1.53 ×
0.2743
10.9532
0.0363
0.0315
0.0761
0.0464
0.0663
0.1012
SEG
1.53 ×
0.2749
10.5784
0.0372
0.0293
0.0713
0.0436
0.1081
0.1763
TPG
1.53 ×
0.2738
10.5830
0.0336
0.0218
0.0606
0.0385
0.0752
0.1303
Table 3: Comparison with training-free detail enhancement methods on FLUX.2 Klein 9B Base. T ♭ achieves the strongest detail statistics and quality with smallest structural changes. It adds negligible inference time compared to the baseline.
Figure 6: Comparison with the best-performing competing methods in table 3 . T ♭ better enhances local details over the baseline. Zoom in to better view the differences. More examples in section B.9 .
Figure 7: Percentage change in spectral energy of different frequency bands when T ♭ is applied at a single sampling step using the same weight with FLUX.2 Klein 9B.
Figure 11
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Magnitude
H(p)
H(p^)
Aall
Aprompt
FLUX.2 Klein 9B Base
square
4.463
4.530
78.0%
100.0%
absolute
7.824
7.827
68.2%
100.0%
FLUX.2 Klein 9B
square
4.307
4.433
83.7%
100.0%
absolute
7.825
7.830
68.9%
100.0%
SD3.5 Large
square
3.065
3.080
71.4%
98.2%
absolute
7.150
7.149
47.6%
37.4%
Appendix
Table 4: Comparison of token densities constructed from squared and absolute channel magnitudes. For each model, H(p) and H(p^) denote the mean entropy without and with detail instructions, respectively. We also report the token-level and prompt-level entropy-increase agreement ratios, Aall and Aprompt .
Figure 10: Block-wise entropy-increase agreement ratios computed from squared channel magnitudes. Each point is averaged over timesteps and image tokens.
Figure 11: Time-wise entropy-increase agreement ratios computed from squared channel magnitudes. Each point is averaged over blocks and image tokens.
Model
H(p)
Added text
H(p^)
Aall
Aprompt
Spearman
FLUX.2 Klein 9B Base
4.510
content
4.507
48.7%
45.6%
0.967
detail
4.561
71.0%
100.0%
0.969
FLUX.2 Klein 9B
4.355
content
4.328
38.8%
18.4%
0.846
detail
4.442
70.8%
100.0%
0.902
SD3.5 Large
3.115
content
3.118
58.8%
64.0%
0.950
detail
3.131
69.8%
98.8%
0.971
Appendix
Table 5: Same comparison as table 1 , computed at 1024×1024 resolution instead of 512×512 . The same pattern holds: detail instructions increase mean entropy and agreement ratios more than content instructions across models, confirming the conclusion is stable across resolutions.
Detail instruction
Top eight shares among 64 bins
1
2
3
4
5
6
7
8
0.6438
0.0416
0.0317
0.0262
0.0226
0.0198
0.0177
0.0159
✓
0.6310
0.0430
0.0328
0.0272
0.0234
0.0205
0.0183
0.0165
Appendix
Table 6: Top eight shares of the averaged 64-bin squared channel-magnitude distribution for the distilled FLUX.2 Klein 9B model, without and with a detail instruction.
Input α
Image Quality
Detail Statistics ↗α
Structure Difference
CLIP ↑
HPSv3 ↑
Edge ↑
Sharpness ↑
f95↑
Wavelet ↑
DINO
LPIPS 128
SD3.5 Large
−0.03
0.2722
11.2652
0.0317
0.0188
0.0377
0.0219
0.0933
0.0922
−0.02
0.2724
11.5866
0.0324
0.0193
0.0395
0.0232
0.0760
0.0767
−0.01
0.2726
11.8132
0.0331
0.0195
0.0416
0.0248
0.0600
0.0627
0 (baseline)
0.2726
12.0507
0.0340
0.0194
0.0441
0.0268
0.0000
0.0000
Appendix
Table 7: T ♭ results for the SD3.5 and Z-Image families over a range of α values.
Figure 12: Additional qualitative examples following the α ranges in table 2 .
Figure 13: Additional qualitative examples following the α ranges in table 2 . Failure example: the overall color tone in the last row (FLUX.2 Klein 9B Base) also shifts slightly.
Figure 14: Additional qualitative examples following the α ranges in table 2 . Failure examples: Qwen-Image becomes more cartoonish at larger α , and FLUX.2 Klein 9B Base shows some local sharp color artifacts.
Figure 15: Additional qualitative examples following the α ranges in table 2 . When the image is already highly complex with abundant detail, applying our method further can look somewhat inharmonious.
Figure 16: Additional qualitative examples following the α ranges in table 2 . Failure example: in the last row (FLUX.2 Klein 9B Base), the arm pose changes across α , showing that structure preservation is not always guaranteed.
Figure 17: Additional comparison with other competing methods in table 3 on FLUX.2 Klein 9B Base. Zoom in to better view the differences.
Figure 18: Additional comparison with other competing methods in table 3 on FLUX.2 Klein 9B Base. Zoom in to better view the differences.
Figure 19: Additional comparison with other competing methods in table 3 on FLUX.2 Klein 9B Base. Zoom in to better view the differences.
Figure 20: Additional comparison with other competing methods in table 3 on FLUX.2 Klein 9B Base. Zoom in to better view the differences.
Figure 21: Additional comparison with other competing methods in table 3 on FLUX.2 Klein 9B Base. Zoom in to better view the differences.
Text-to-image diffusion transformers (DiTs) are powerful generators, yet direct prompting provides limited control interface for style intensity and can fail to suppress unwanted concepts. To enable these controls, we introduce AcFlow, an inference-time controller that transports intermediate layer image-token activations through a learned concept-conditioned velocity field while keeping the base DiT frozen. A textual concept description specifies the desired intervention, while the integration horizon provides a continuous control parameter. The field produces token-varying, activation-dependent updates. With parameters shared across concepts within each task family, one field covers over 15,000 style descriptions or over 1,000 suppression concepts, and generalizes to concepts unseen during training without per-concept fitting. On style control, AcFlow achieves the best style--content trade-off among the evaluated baselines in the high-style-alignment regime. At a fixed operating point, AcFlow attains style--content alignment of 0.5365/0.2860, compared with 0.4397/0.2684 for the baseline with the highest style alignment. On concept suppression, AcFlow reduces the fraction of images showing the concept from 95.3%/82.1% to 41.6%/40.5% on held-in/held-out concepts, including cases where deleting them from the prompt fails to remove them. Our analyses support the learned velocity field as an adaptive control mechanism, with update directions varying across tokens and depending on their activation states. Our code is available at https://github.com/Nove1yst/AcFlow.
Junran Wang, Zehao Jin, Tianyu Luan +3
Georgia Institute of technology · South China University of Technology
Pixel-space generative Transformers typically use one token per image patch, so capturing finer detail requires more tokens and makes global attention increasingly expensive. We observe that finer detail does not necessarily require more global communication: low-frequency structure benefits from global exchange, while high-frequency detail can be modeled locally. Based on this insight, we introduce SpecSlot, a pixel-space rectified-flow Transformer that decouples output detail from the length of the global attention sequence. SpecSlot expands each spatial anchor into multiple slots, each assigned a disjoint band of DCT coefficients; this separates anchor density from spectral subdivision. Only the lowest-frequency slot from each anchor participates in global attention, while all slots interact locally. This allows the model to increase spectral detail without enlarging the global sequence. A spatial head further adds a local correction to the spectral prediction, while the generative state, target, and loss remain in pixel space. On ImageNet-256, SpecSlot achieves a comparable final FID to a FLOPs-matched JiT control while converging faster. On ImageNet-512, it improves FID by 2.84 over the control with 3.9% fewer FLOPs, and comes within only 0.69 FID of JiT-B/16 at 38.2% lower cost. On text-to-image generation, SpecSlot surpasses JiT-B/8 on DPG-Bench by 1.06 at 40.6% lower forward cost.
Jiongfan Yang, Weichen Fan, Haiwen Diao +2
Peking University · SenseTime · Nanyang Technological University
Generating images beyond the training resolution of Diffusion Transformers (DiTs) remains challenging, as the number of tokens grows rapidly with resolution, full attention becomes expensive, and outputs often suffer from structural distortions and degraded detail. We identify uniform full attention as the common source of both problems. Attending to all tokens at every step causes quadratic cost and exposes the model to more tokens and relative distances beyond the training range, pushing the model out of distribution. We observe that attention's effective receptive field narrows during denoising, from global in early steps that form structure to local in later steps that refine details. We propose AReA, a training-free attention mechanism whose receptive field evolves from coarse to fine, following the denoising progression. In early steps, AReA applies cross-scale coarse attention over downsampled keys and values, capturing global context, while in later steps it applies local attention within progressively shrinking windows. Without additional parameters or training, experiments show that AReA consistently outperforms state-of-the-art training-free baselines across target resolutions and different models, while also being faster and more efficient.