TIDE: Text-Informed Dynamic Extrapolation with Step-Aware Temperature Control for Diffusion Transformers

Yihua Liu; Fanjiang Ye; Bowen Lin; Rongyu Fang; Chengming Zhang

cs.CVMar 9, 2026

TIDE: Text-Informed Dynamic Extrapolation with Step-Aware Temperature Control for Diffusion Transformers

Authors: Yihua Liu, Fanjiang Ye, Bowen Lin, Rongyu Fang, Chengming Zhang

Organizations: Independent Researcher · Rice University · University of Houston

Abstract

Diffusion Transformer (DiT) faces challenges when generating images with higher resolution compared at training resolution, causing especially structural degradation due to attention dilution. Previous approaches attempt to mitigate this by sharpening attention distributions, but fail to preserve fine-grained semantic details and introduce obvious artifacts. In this work, we analyze the characteristics of DiTs and propose TIDE, a training-free text-to-image (T2I) extrapolation method that enables generation with arbitrary resolution and aspect ratio without additional sampling overhead. We identify the core factor for prompt information loss, and introduce a text anchoring mechanism to correct the imbalance between text and image tokens. To further eliminate artifacts, we design a dynamic temperature control mechanism that leverages the pattern of spectral progression in the diffusion process. Extensive evaluations demonstrate that TIDE delivers high-quality resolution extrapolation capability and integrates seamlessly with existing state-of-the-art methods.

Figures & tables

Figure 1: Collage of multi-resolution results generated by TIDE. Prompts are from DrawBench and Aesthetic-4K. Zoom in for details.

Figure 2: Overview of our proposed TIDE framework. (a) Visual comparison at $4096\times 4096$ resolution. While existing methods (left) suffer from information loss and repetitive artifacts, our approach (right) preserves prompt fidelity and generates realistic details. (b) Dynamic Temperature Control method, which dynamically adjusts the attention temperature across time-steps to eliminate high-frequency artifacts. (c) Text Anchoring mechanism, which counteracts attention dilution by reinforcing the cross-attention scores between image queries and text keys, recovering the influence of text tokens.

Figure 3: Visual comparison of subject vanishing issues at 4K resolution. We compare four extrapolation baselines at $4096\times 4096$ with the native resolution $1024\times 1024$ . While the model at native resolution produces coherent subjects, baseline methods except YaRN suffer from severe subject vanishing issue, and YaRN’s content richness has also noticeably declined.

Figure 4: Visualization of text token influence decay. We visualize the spatial influence of text prompts at the early sampling stage, revealing that text guidance influence diminishes as the target resolution scales to $4096\times 4096$ . Specifically, Direct Extrapolation, NTK-Aware Interpolation and NTK-by-Parts Interpolation exhibit a near-total loss of spatial influence, providing a mechanistic explanation for the subject vanishing issue. YaRN only partially recovers token influence in the central region, resulting in suboptimal generation quality.

Figure 5: Qualitative comparison with baseline method. We compare TIDE against baseline methods. The input prompts correspond to the rows: 2K Prompt 1: “ A Moai statue gazes upward against a starry night sky, with a colorful sunset illuminating the horizon. ”; 2K Prompt 2: “ New York Skyline with ’Diffusion’ written with fireworks on the sky. ”; 4K Prompt 1: “ A futuristic spaceship flies over a lush landscape featuring rocky cliffs and a settlement with domed buildings near a serene body of water surrounded by greenery. ”; 4K Prompt 2: “ Narrow cobblestone alley featuring colorful buildings with green and orange facades, potted plants lining the street, and a puddle reflecting the scene. ”.

Resolution	Method	DrawBench			Aesthetic-4K
Resolution	Method	CLIP $\uparrow$	IR $\uparrow$	Aesth $\uparrow$	FID $\downarrow$	KID $\downarrow$	CLIP $\uparrow$	IR $\uparrow$	Aesth $\uparrow$
$2048\times 2048$	FLUX	26.63	0.46	5.53	162.59	0.0087	27.87	1.02	6.54
	YaRN	26.76	0.66	5.76	141.93	0.0045	28.07	1.14	6.59
	Dy-YaRN	26.52	0.72	5.74	161.32	0.0068	28.09	1.10	6.62
	TIDE (Ours)	27.34	0.83	5.77	159.35	0.0066	28.60	1.21	6.68
$4096\times 4096$	FLUX	15.16	-2.19	2.92	341.79	0.1952	13.46	-2.25	2.58

Table 1: Quantitative comparison with baselines. We compare our TIDE with FLUX, YaRN, and Dy-YaRN on DrawBench and Aesthetic-4K. All methods are built on FLUX. Best results are bolded .

Method	TIDE (Ours)	Dy-YaRN
Text Alignment	4.57	3.60
Global Structure	4.33	3.21
Texture Quality	4.46	2.48

Table 2: User study results. We conduct a subjective evaluation comparing TIDE and Dy-YaRN on 20 randomly selected prompts. Participants rated images on a scale of 1 to 5 (higher is better). We report the mean score for each aspect.

Configurations	CLIP $\uparrow$	IR $\uparrow$	Aesth $\uparrow$
Baseline	23.60	-0.44	5.64
TA Only	26.60	0.66	6.54
DTC Only	23.28	-0.30	5.65
Full Method	26.76	0.86	6.46

Table 3: Component-wise ablation study. We investigate the impact of Text Anchoring and Dynamic Temperature Control on Asethetic-4K. The baseline model is Dy-YaRN. Best results are bolded .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Scale-adaptive bias. Our adaptive bias increases logarithmically with the extrapolation scale.

Method	Bias Scale	2048 $\times$ 2048			3072 $\times$ 3072			4096 $\times$ 4096
Method	Bias Scale	CLIP $\uparrow$	IR $\uparrow$	Aesth $\uparrow$	CLIP $\uparrow$	IR $\uparrow$	Aesth $\uparrow$	CLIP $\uparrow$	IR $\uparrow$	Aesth $\uparrow$
Dy-YaRN	-	29.10	1.12	6.83	26.87	0.69	6.56	23.60	-0.44	5.64
TIDE (Ours)	1.0	29.16	1.15	6.86	27.81	0.94	6.78	26.05	0.54	6.26
	2.0	28.87	1.19	6.80	28.08	1.08	6.76	26.91	0.83	6.47
	3.0	28.42	1.07	6.84	28.14	1.03	6.77	27.25	0.71	6.41
	Adaptive	29.12	1.17	6.85	28.28	1.09	6.78	26.76	0.86	6.46

Appendix

Table 4: Ablation study on Text Anchoring bias across different resolutions. We evaluate the impact of bias strategies on image generation quality at

2048^{2}

3072^{2}

, and

4096^{2}

resolutions. Dy-YaRN is adopted as the baseline. The best results for each resolution are bolded , and the second-best results are underlined .

Figure 7: Temperature evolution patterns with different

\alpha

. (a) demonstrates different patterns for component with lowest frequency of positional embeddings. (b) demonstrates different patterns for component with highest frequency of positional embeddings.

Method	$\alpha_{\text{low}}$	$\alpha_{\text{high}}$	CLIP $\uparrow$	IR $\uparrow$	Aesth $\uparrow$
Dy-YaRN	-	-	23.60	-0.44	5.64
TIDE	0.0	0.0	26.60	0.66	6.54
	0.6	0.1	27.25	0.84	6.49
		0.2	26.76	0.86	6.46
		0.5	27.09	0.84	6.52
		1.0	26.03	0.53	6.19

Appendix

Table 5: Hyperparameter sensitivity of Dynamic Temperature Control. We evaluate the impact of low-frequency (

\alpha_{\text{low}}

) and high-frequency (

\alpha_{\text{high}}

) exponents. Dy-YaRN serves as the baseline. The overall best results are bolded .

Figure 8: Qualitative comparison at different $\alpha_{\text{low}}$ and $\alpha_{\text{high}}$ . The image on the left (with $\alpha_{\text{low}}=0.0$ and $\alpha_{\text{high}}=0.0$ ) suffers from severe artifacts, while the image on the right shows blurry details. The image in the middle has the best quality.

Figure 9: Qualitative comparison for high-resolution text-to-image generation on DrawBench. Our method outperforms baselines across by delivering high visual quality.

Figure 10: Qualitative comparison for high-resolution text-to-image generation on Aesthetic-4K. Our method outperforms baselines across by delivering high visual quality.

Figure 11: Qualitative comparison of panoramic generation at $4096\times 1536$ resolution.

Figure 12: Qualitative comparison of panoramic generation at $4096\times 1536$ resolution.

T2I Diffusion Models

Text-to-image (T2I) diffusion models generate images from prompts by learning to reverse noise under text conditioning.

See all 403 papers

SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers
May 21, 2026Javad Rajabi, Kimia Shaban, Koorosh Roohi +2Rotary Positional Embeddings Diffusion Model Guidance
University of Toronto · Vector Institute
HRDiT: Training-Free High-Resolution Image Generation with Off-the-Shelf Diffusion Transformer Models
Aug 7, 2026Yu Xue, Haoxuan Qu, Zhuoling Li +5Diffusion Model Adaptation Diffusion Transformer
Lancaster University, Lancaster, UK · South China University of Technology, Guangzhou, China · NVIDIA AI Tech Centre, Singapore
RAISE: Requirement-Adaptive Evolutionary Refinement for Training-Free Text-to-Image Alignment
Feb 28, 2026Liyao Jiang, Ruichen Chen, Chao Gao +1T2I Generation Text-Image Alignment
Department of ECE, University of Alberta, Canada · Huawei Technologies, Canada

May 21, 2026J/K move · Enter open · S save

SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers

Javad Rajabi, Kimia Shaban, Koorosh Roohi, David B. Lindell+1

University of Toronto · Vector Institute

Rotary Positional Embeddings Diffusion Model Guidance Diffusion Transformer Position Extrapolation T2I Diffusion Models

Open paper AskPDF arXiv

Diffusion transformers (DiTs) have emerged as a dominant architecture for text-to-image generation, yet their performance drops when generating at resolutions beyond their training range. Existing training-free approaches mitigate this by modifying inference-time attention behavior, often through Rotary Position Embeddings (RoPE) extrapolation combined with attention scaling. However, these strategies apply a uniform and content-agnostic scaling across RoPE components with distinct frequency characteristics, inducing a trade-off between preserving global structure and recovering fine detail. We introduce SEGA, a training-free method that dynamically scales attention across RoPE components according to the latent's spatial-frequency structure at each denoising step. This adaptive scaling improves both structural coherence and fine-detail fidelity. Experiments show that SEGA consistently improves high-resolution synthesis across multiple target resolutions, outperforming state-of-the-art training-free baselines.

Abstract

T2I Diffusion Models

Explore similar work

SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers

HRDiT: Training-Free High-Resolution Image Generation with Off-the-Shelf Diffusion Transformer Models

RAISE: Requirement-Adaptive Evolutionary Refinement for Training-Free Text-to-Image Alignment