Although text-to-image generative models produce impressive results, they struggle to generate densely detailed, high-resolution (HR) images. Current literature addresses this issue with a low-to-high-resolution approach. First, a low-resolution (LR) image is generated. Then, an upsampled version is generated using the LR image as an additional cue. In this paper, we present Joint Latent Trajectories (JoLT). To generate an image, JoLT uses two streams that jointly denoise LR and HR latent images at each sampling step. The LR latent controls the overall layout, while the HR latent controls the details. We interconnect both branches to jointly integrate their information. We extensively validate our method, demonstrating its advantages over competing baselines. The resulting images are not only richly detailed but also visually pleasing, opening new avenues for artistic creation.
Figures & tables
Figure 1: Controllable high-resolution generation with JoLT. The master prompt c conditions both the FLUX.2 Klein-9B base generation (left) and our result (right), while the detail prompt cˉ and complexity strength α control the added local detail. Dashed boxes identify matched groups of figures near the image center, enlarged in the overlaid crops. The base image is displayed at half the width and height of our result; the bands summarize their qualitative capabilities.
Figure 2: Overview of JoLT. At t=0 , a shared Gaussian draw initializes the HR canvas xˉ0 and is randomly subsampled to initialize the LR latent x0 . Two denoising trajectories then run jointly over the same 4 steps of a distilled flow-matching model. The LR stream denoises the whole scene from the master prompt; the HR stream denoises overlapping windows and adds local detail. At every step, the LR stream’s clean estimate x^1 is decoded, upsampled, and re-encoded into a reference latent rˉt whose aligned crops Fi(rˉt) condition the HR stream. The updated HR canvas is randomly subsampled and blended into the LR update with weight α before the next step.
Complexity
Text alignment
Aesthetics & quality
Latency
Method
IC9600 ( ↑ )
DCT ( ↑ )
JPEG ( ↑ )
LMM corr. ( ↑ )
LMM percept. ( ↑ )
QIB Quality ( ↑ )
Time (s) ( ↓ )
SEGA
0.615±0.070
483±236
0.096±0.022
0.371±0.038
0.423±0.160
47.2±10.4
75±9
ResDiT
0.714±0.073
490±206
0.095±0.019
0.360±0.043
0.448±0.152
47.3±10.0
86±22
DemoFusion
0.689±0.068
613±219
0.104±0.017
0.356±0.044
0.411±0.167
50.2±9.2
125±10
Ours
0.753±0.055
920±140
0.131±0.011
0.367±0.047
0.462±0.164
50.9±9.4
127±15
Table 1: Quantitative comparison on 200 prompts. Our full method leads all three displayed complexity metrics. Compared with the strongest baseline, the gains are 5% on IC9600, 50% on DCT energy, and 26% on JPEG ratio. It also has the highest means on LMM4LMM perception and QIB Quality. Its LMM4LMM correspondence is within 0.004 of the leading baseline. Entries are mean ± sample standard deviation. Boldface and underlining mark the best and second-best means only when their paired difference is significant after metric-family correction ( padj<.05 , two-sided Wilcoxon). Latency reports seconds per image on one H100 80 GB
Baseline
Preference [95% CI]
Order agree.
pHolm
Outcome
DemoFusion
.458[.398,.520]
79.5%
.204
Same range
SEGA
.748[.693,.800]
83.5%
<.001
JoLT preferred
ResDiT
.940[.910,.965]
93.0%
<.001
JoLT preferred
Table 2: Global–local coherence comparison on 200 paired prompts per baseline. Order agreement is the percentage of pairs for which the decision is unchanged after reversing the A/B presentation order. Intervals are 95% percentile-bootstrap CIs and pHolm corrects exact sign tests. More details are provided in .
Figure 3: Qualitative comparison with high-resolution generation baselines. Our result is shown in the leftmost panel. All methods use the same backbone and output resolution. Each lower-right inset magnifies the square center crop spanning one eighth of the source image width and height. Among the compared outputs, ours presents the densest and most varied fine-grained detail in both the full-frame view and the magnified inset.
Baseline
More detailed / complex (95% CI)
Better prompt match (95% CI)
More artistically appealing (95% CI)
SEGA
77.3%±5.1%
69.0%±5.4%
68.0%±5.5%
DemoFusion
76.3%±5.1%
48.5%±5.6%
46.3%±5.7%
ResDiT
76.0%±5.1%
68.3%±5.5%
71.0%±5.4%
Table 3: Two-alternative forced-choice user study. Entries report the percentage of trials preferring our method as the mean ± the 95% confidence-interval half-width. A light-green background marks a preference significantly above the 50% chance level (95% confidence interval entirely above 50%).
Figure 4: Complexity response to HR → LR conditioning strength α ( n=25 matched prompts, tiled denoising at 40962 px final image). Both IC9600 and DCT energy metrics increase most strongly above α=0.5 , confirming that HR → LR conditioning provides controllable complexity.
Figure 5: Qualitative ablation of HR → LR conditioning strength α . Across both examples, increasing α produces a clear increase in the number and richness of visible details, especially in the magnified regions shown in the lower-right insets.
Figure 6: Component ablation. Left: the four configurations and their native output resolutions; colors match the bars. Right: mean IC9600, LMM4LMM correspondence, and QIB Quality over 200 prompts; error bars show one standard deviation. Since one window is limited to 20482 px, 40962 px generation requires MD.
Figure 7: Qualitative detail-prompt ablation. We hold the master prompt and seed fixed and vary the detail prompt shown below each image. The master prompt is “A flooded hotel atrium contains a functioning print shop with reed boats, mechanical clocks, stained-glass fish, and citrus trees […].” All results are generated at 40962 px.
Ultra-high-resolution text-to-image generation is increasingly vital for applications requiring fine-grained textures and global structural fidelity, yet state-of-the-art text-to-image diffusion models such as FLUX and SD3 remain confined to sub 2MP (< 1K×2K) resolutions due to the quadratic complexity of attention mechanisms and the scarcity of high-quality high-resolution training data. We present UltraImageGen, a novel framework that introduces hierarchical local attention with low-resolution global guidance, enabling efficient, scalable, and semantically coherent image synthesis at ultra-high resolutions. Specifically, high-resolution latents are divided into hardware aligned fixed-size local windows to reduce attention complexity from quadratic to near-linear, while a low-resolution latent equipped with scaled positional embeddings injects global semantics as an anchor. A lightweight LoRA adaptation bridges global and local pathways during denoising, ensuring consistency across structure and detail. To maximize inference efficiency and achieve scalable ultra-high-resolution generation, we repermute token sequence in window-first order, so that the GPU-friendly dense local blocks in attention calculation equals to the fixed-size local window in 2D regardless of resolution. Together ourwork reliably scales the pretrained model to resolutions higher than 8K with more than 10× speed up and significantly lower memory usage. Extensive experiments demonstrate that ourwork achieves superior quality while maintaining computational efficiency, establishing a practical paradigm for advancing ultra-high-resolution image generation.
Training-free text-to-high-resolution image generation has recently attracted growing research attention. However, existing studies on this task primarily focus on adapting off-the-shelf U-Net-based diffusion models to high resolutions, with limited progress on adapting off-the-shelf Diffusion Transformer (DiT) models despite their strong text-to-image generation capabilities at limited resolutions. In this work, we find two key challenges particularly hindering the application of off-the-shelf DiT models for high-resolution image synthesis in a training-free manner, namely, spatial disorder and long generation time. To address these challenges, we propose a novel method tailored to adapt off-the-shelf DiT models for high-resolution image synthesis. Extensive experiments show the efficacy of our method. Our code is available at: https://github.com/zylwithxy/HRDiT.
Yu Xue, Haoxuan Qu, Zhuoling Li +5
Lancaster University, Lancaster, UK · South China University of Technology, Guangzhou, China · NVIDIA AI Tech Centre, Singapore
Text-to-image diffusion models have two major drawbacks that severely limit their practical utility: (1) standard models lack an intrinsic mechanism for continuous, concept-specific guidance (e.g., for precisely controlling how aesthetically pleasing an image looks), and (2) they lack reliability for tasks requiring high local coherence (e.g., generating text or human hands). To tackle these issues, we introduce a novel notion of concept-wise mutual information and find large, concept-dependent differences between individual layers, demonstrating that the generation of specific structures is localized in distinct parts of the network. We exploit this insight by reinforcing the impact of concept-relevant layers in Concept Guidance (CoG), a precise, target-specific guidance method that works for models out-of-the-box without additional training, external models, gradients, or prompt engineering. CoG first quantifies each layer's concept-specific impact and then guides denoising using a weighted combination of predictions generated with concept-relevant layers skipped. We demonstrate performance increases across various targets and popular models like PixArt-alpha, SD3, SD3.5, and FLUX.1-dev. Code is available at https://github.com/CompVis/concept_guidance
Nikolai Röhrich, Isabell Hans, Felix Krause +1
LMU Munich, Germany · Konrad Zuse School of Excellence in Reliable AI (relAI), Germany · Munich Center for Machine Learning (MCML), Germany