UltraFlux: Robust 4K Text-to-Image Diffusion

• UltraFlux is a data–model co-design regime that integrates the MultiAspect-4K-1M dataset with targeted architectural innovations to achieve state-of-the-art 4K text-to-image generation.
• It employs resonance-based 2D RoPE with YaRN and SNR-aware Huber wavelet training to mitigate positional drift, aliasing, and gradient imbalances across diverse aspect ratios.
• Non-adversarial VAE post-training coupled with a stage-wise aesthetic curriculum significantly enhances fidelity metrics such as FID, PSNR, and SSIM for ultra-high-resolution outputs.

UltraFlux is a data–model co-design regime for native 4K text-to-image diffusion transformers (DiTs) with robust generalization across diverse aspect ratios (ARs). Developed to address tightly coupled failures in extending diffusion transformers to 4K—including positional encoding drift, VAE compression loss, and optimization instabilities—UltraFlux integrates a multi-aspect 4K dataset (MultiAspect-4K-1M) with targeted architectural innovations in positional encoding, post-training VAE reconstruction, frequency-aware optimization, and aesthetic curriculum learning. This approach yields a single-step, detail-preserving generative model that consistently achieves state-of-the-art fidelity and aesthetic quality under resolution- and AR-aware evaluation, outperforming open-source baselines and rivaling proprietary models (Ye et al., 22 Nov 2025).

1. MultiAspect-4K-1M Dataset

UltraFlux is trained on MultiAspect-4K-1M, a 1,007,230-image corpus explicitly curated for 4K generation with broad AR coverage. The collection starts from approximately 6 million images (resolution ≥3840×2160), filtered for semantic and technical quality using Q-Align (VLM quality), ArtiMuse (aesthetic score), flatness and entropy metrics, and open-vocabulary detection (YOLOE). Person-centric images are upsampled to balance content using a dedicated augmentation path. Final AR distribution is nearly uniform across buckets (square, landscape, portrait), enabling bucketed training at resolutions such as 4096×4096, 5120×2880, 2048×4096, and 5952×2496.

Each image includes English captions generated with Gemini-2.5-Flash and translated to Chinese using Hunyuan-MT-7B; average length is 125 tokens. VLM/IQA metadata encompasses Q-Align, ArtiMuse, flatness (Sobel variance), entropy, and subject tags. This metadata supports AR-aware batching, stratified sampling, aesthetic subset selection, and facilitates downstream fine-grained analysis (Ye et al., 22 Nov 2025).

2. Resonance 2D RoPE with YaRN for Positional Encoding

Standard 2D rotary positional encoding (RoPE) in Flux-based DiTs assigns axis-wise rotary frequencies $\omega_k^{(a)}=(b\cdot\eta)^{-2k/d_a}$ for axes $a \in \{\mathrm{H}, \mathrm{W}\}$ and phases $\phi_k^{(a)}(p_a) = p_a\,\omega_k^{(a)}$. At 4K and with non-square ARs, this leads to spectrum drift and phase aliasing.

Resonance projection first snaps the original cycles $r_k^{(a)}=L_a\,\omega_k^{(a)}/2\pi$ to the nearest integer $\hat r_k^{(a)}= \max(1, \mathrm{round}(r_k^{(a)}))$, and sets projected frequency $\hat\omega_k^{(a)}=2\pi\,\hat r_k^{(a)}/L_a$, enforcing integer-cycle "standing waves" and zero phase-closure error over the window $[0, L_a]$. YaRN-style extrapolation ensures AR-aware, band-wise ramped scaling for inference length $L_a'$ via $$\omega_{k,\mathrm{yarn}}^{(a)}=(1-\gamma(\hat r_k^{(a)}))(\hat\omega_k^{(a)}s_a) + \gamma(\hat r_k^{(a)})\hat\omega_k^{(a)}$$ with $s_a=L_a'/L_a$. This procedure, combining training-window and frequency awareness with AR scaling, mitigates aliasing and ghosting at high resolutions and all ARs (Ye et al., 22 Nov 2025).

3. VAE Compression: Non-Adversarial Post-Training

UltraFlux uses the Flux F16 VAE, trading off computation for lower high-frequency fidelity relative to F8. To recover detail, a non-adversarial post-training scheme fine-tunes the F16 decoder (encoder frozen) on high-detail crops (top 50% by flatness) using:

$$\mathcal{L} = \lambda_{\mathrm{wav}} L_{\mathrm{wav}} + \lambda_{\mathrm{perc}} L_{\mathrm{perc}} + \lambda_{L_2} \Vert \hat x - x \Vert_2^2$$

with $$(\lambda_{\mathrm{wav}}, \lambda_{\mathrm{perc}}, \lambda_{L_2}) \approx (0.2, 0.1, 1)$$, AdamW at $1\mathrm{e}{-5}$, and a batch size of 384 for $\sim$4k steps. $L_{\mathrm{wav}}$ is high-frequency wavelet subband $L_2$ loss, $L_{\mathrm{perc}}$ is VGG-based perceptual loss. No adversarial (GAN) term is present for stability. On Aesthetic-4K@4096, this reduces rFID from 2.201$\to$0.547, PSNR from 26.90$\to$30.70, SSIM from 0.784$\to$0.852, and LPIPS from 0.168$\to$0.102 (Ye et al., 22 Nov 2025).

4. SNR-Aware Huber Wavelet Training Objective

UltraFlux replaces standard L2 flow-matching losses with a time- and frequency-reweighted pseudo-Huber wavelet penalty. For diffusion step $t\in(0,1)$, the model predicts velocity $v_\theta$ over

$$z_t = (1-t)z + t\epsilon,\quad \hat{z}_t = z_t - t\,v_\theta(z_t, t)$$

Wavelet coefficients on the denoised output and target, $R_\theta = W(\hat{z}_t) - W(z)$ with $W$ a one-level orthonormal DWT over (LL, LH, HL, HH), are combined per pixel using a pseudo-Huber penalty $\rho_c(r)=c^2(\sqrt{1+(r/c)^2}-1)$ with scale $c(t)$ increasing with SNR. The loss is

$$\mathcal{L}(\theta) = \mathbb{E}_{z, \epsilon, t} \left[ \omega(t)\,\ell_{\mathrm{Huber}}(R_\theta; c(t)) \right]$$

where $\omega(t)$ rebalances gradient flow using SNR-aware weighting: $\mathrm{SNR}(t) = ((1-t)/t)^2$, $$\omega(t) = (t/(1-t))[\min\{\mathrm{SNR}(t),\gamma\}]^\beta$$. This formulation addresses over-smoothing of high-frequency components, rebalances gradients across time, and decouples frequency scales (Ye et al., 22 Nov 2025).

5. Stage-wise Aesthetic Curriculum Learning

A two-stage curriculum links data resolution, noise regime, and supervision, optimizing for both prior-driven structure and high-aesthetic modes. Stage 1 trains on full MultiAspect-4K-1M with uniform timestep sampling ($t\in[0,1]$) over $\sim$30k steps for a general 4K prior. Stage 2 restricts both data (top 5% by ArtiMuse) and noise level (high-noise steps $t$ covering first $\sim$460/$1000$) for an additional $\sim$2k steps, sculpting the prior towards ultra-high-aesthetic outputs under structure-dominated sampling. This sequence targets difficult generation regimes for concentrated improvement (Ye et al., 22 Nov 2025).

6. Performance Benchmarks and Ablation Studies

UltraFlux is evaluated on Aesthetic-Eval@4096 and diverse 4K ARs (2:1, 1:2, 16:9, 2.39:1) using FID, HPSv3, PickScore, ArtiMuse, CLIP Score, Q-Align, and MUSIQ. On 4096×4096, UltraFlux achieves FID 143.11 (vs Sana 144.17, Diffusion-4K 152.43), HPS 11.47, ArtiMuse 68.36 (vs Sana 63.72), Q-Align 4.85 (vs 4.89), and MUSIQ 46.13 (vs 45.08). Across non-square 4K settings, UltraFlux outperforms open-source baselines (e.g., Sana, ScaleCrafter, FouriScale) in FID, HPS, and ArtiMuse.

When paired with a prompt refiner (GPT-4O), UltraFlux matches or slightly exceeds proprietary Seedream 4.0 in Q-Align and MUSIQ under identical prompts, e.g., Q-Align (UltraFlux 4.93, Seedream 4.71), MUSIQ (UltraFlux 45.93, Seedream 30.21), with a narrower FID gap (147.06 vs 132.87).

Human evaluations (Gemini-2.5) indicate UltraFlux is preferred in 70–82% of visual appeal and 60–89% of prompt alignment comparisons to open models.

Ablation studies show additive gains from each architectural and optimization component: SNR-Aware Huber Wavelet yields ∼2.6 FID improvement, SACL an additional ∼1.5, and Resonance 2D RoPE with YaRN ∼0.4, with each step improving HPS and ArtiMuse monotonicity (Ye et al., 22 Nov 2025).

---

In aggregate, UltraFlux demonstrates that coordinated data–model innovations—AR-diverse curation, frequency- and SNR-aware training objectives, standing-wave positional encodings, and curriculum learning—enable stable, high-fidelity 4K text-to-image generation. The regime attenuates the decomposition of significant error sources (positional drift, compression loss, gradient imbalance) and provides a reproducible foundation for further advances in ultra-high-resolution generative modeling (Ye et al., 22 Nov 2025).