BluePixel: Toward Native 4K Agentic Image Editing

Shanyan Guan, Shiyu Li, Xin Li, Zhizhou Chen, Yunzhe Xu, Yuehao Liu, Ruixue Shen, Shiju Zhang, Min Jin, Qiang Liu, Junbo Liu, Penggang Shi, Sheng Sun, Xiao Wang, Yong Zhang, Zhongang Qi, Wei Li, Chao Ma, Ying Tai, Yanhao GearXiv 2026 · arXiv

TL;DR

BluePixel is a comprehensive system designed to make native 4K image editing both practical and controllable. By combining a novel "progressive VAE adaptation" training recipe with an agentic harness that manages specialized tools and closed-loop verification, the authors achieve high-fidelity 4K edits in under 9 seconds on a single XPU.

What Problem Does This Paper Address?

While AI image generation has reached 4K, high-resolution editing remains a significant challenge. The authors identify three primary hurdles:

  1. Token Explosion: Standard Variational Autoencoders (VAEs) produce latent sequences for 4K images that are too massive for current Diffusion Transformers (DiTs) to process efficiently.
  2. Data Scarcity: High-quality 4K image-editing pairs are rare and expensive to produce.
  3. The "Single-Model" Trap: A single end-to-end model often struggles to follow vague user instructions, perform precise parametric adjustments (like specific color grading), or match the quality of specialized expert models for tasks like text rendering or portrait retouching.

BluePixel reframes 4K editing as a system problem rather than a single-model problem.

Methods

The BluePixel framework consists of two main components: an efficient native 4K model and an agentic harness.

1. Efficient Native 4K Training

To overcome the computational cost of 4K, the authors use a three-phase recipe:

  • Phase 1: Progressive VAE Adaptation. Instead of a standard $f=8$ VAE, they use a highly compressed $f=64$ VAE (DCAE-F64). To bridge the representation gap, they adapt the model progressively ($f: 8 \to 32 \to 64$), keeping the 4K training memory footprint identical to 1K training.
  • Phase 2: Enhancement. They use ACE-MOPD (a combination of gradient-orthogonal LoRA and multi-teacher distillation) to absorb "vertical" skills—like hairstyle changes or teeth whitening—without "catastrophic forgetting" of general editing abilities.
  • Phase 3: Acceleration. They introduce Source-Anchored Target Projection (SATP) distillation. This technique uses the teacher's prediction to distinguish edited from non-edited regions, anchoring the background to the source image to ensure 4K consistency while compressing 50 denoising steps down to just 4.

2. BluePixel Agent

The agentic harness acts as an "orchestrator" that:

  • Decomposes Instructions: Breaks down complex user requests into operations for specific tools.
  • Skill Routing: Delegates tasks like color grading to precise LUT-based (Look-Up Table) tools while reserving the diffusion model for generative changes.
  • Closed-Loop Verification: Uses a VLM-based reward model (BluePixReward) to evaluate every edit. If the result fails to meet quality or instruction-following thresholds, the agent automatically retries the edit with revised instructions.

Datasets

The authors utilize a mix of public and newly curated data:

  • Training: Public datasets like UnicEdit-10M and TextEdit, supplemented by in-house 4K pairs mined from high-resolution video footage.
  • Evaluation: They introduce VINS-4KEval++, a benchmark containing 3,305 native 4K cases across nine task types, significantly expanding on previous 4K benchmarks.

Key Insights and Ideas

  • 4K is a System Problem: The most significant "aha moment" is that an agent managing a suite of tools (including a diffusion model) outperforms a single monolithic model. As seen in Figure 4, routing a "filter" request to a dedicated LUT tool produces more professional results than a diffusion model's approximation.
  • Progressive Adaptation is Mandatory: The authors found that jumping directly from $f=8$ to $f=64$ compression is infeasible; the intermediate $f=32$ step is crucial for the DiT to learn the new latent space.
  • Background Consistency at 4K: At 4K resolution, even minute pixel drifts in non-edited regions become glaringly obvious. SATP is the key innovation that allows for fast (4-step) inference without sacrificing this background stability.
Figure 4: BluePixel Agent routes specific requests (like filters) to specialized tools rather than relying solely on the diffusion model.

Key Results

  • Speed: BluePixel completes a native 4K edit in 8.7 seconds on a single Kunlunxin P800 XPU, representing a ~4x speedup over non-distilled models.
  • Benchmark Performance: On the VINS-4KEval++ benchmark, BluePixel ranks second overall among seven top-tier models (including GPT Image 2 and Seedream 5.0), specifically excelling in non-edited-region consistency.
  • Qualitative Fidelity: As shown in Figure 1 and Figure 6, the model preserves fine details—such as crisp text on a storefront or individual strands of hair—that are often lost in upsampled 1K-to-4K workflows.
Figure 1: Showcase of BluePixel's diverse 4K editing capabilities, from text and pose editing to low-level image restoration.
Figure 6: Qualitative comparison on VINS-4KEval++ showing BluePixel's superior instruction following and detail preservation compared to baselines.

Limitations and Future Work

  • Viewpoint Challenges: The model currently struggles with instructions requiring significant camera viewpoint or perspective changes, as these often conflict with the strict "non-edited-region consistency" constraints.
  • Ranking: While it leads in many categories, it still trails GPT Image 2 in some specific consistency metrics.
  • Future Directions: The authors suggest that further refining the agent's reasoning capabilities and expanding the library of "specialist" tools will be the primary path forward for professional-grade editing.

Summary updated

Abstract

Instruction-driven image editing has reached native 4K, yet how such editors are trained remains undisclosed, and pretraining one from scratch is beyond most teams. The task itself compounds the difficulty: a single end-to-end editor cannot make precise parametric adjustments, match specialized experts, or interpret the vague instructions users actually write. We argue that 4K editing is a system problem rather than a single-model problem, and present BluePixel, which couples a native-4K editing model with an agent harness; the complete training recipe is disclosed in this paper. The model makes 4K training practical: it turns a 1K-pretrained editor into a native-4K one through progressive VAE adaptation that keeps 4K training as affordable as 1K, absorbs a long tail of vertical editing skills through continual learning, and refines the aesthetic-fidelity trade-off with reinforcement learning. Distillation via source-anchored target projection (SATP) then compresses 50 denoising steps to 4, completing a 4K edit in 8.7 s on a single Kunlunxin P800 XPU while preserving task success rate and background consistency. The harness, BluePixel Agent, makes editing controllable: it decomposes user instructions into specialist and general operations, verifies every result with the same reward model that guided training, and retries within bounded attempts. On VINS-4KEval++ (3,305 native-4K cases), BluePixel ranks second among seven models, behind only GPT Image 2 in non-edited-region consistency.
BibTeX
@misc{guan2026bluepixel,
  author = {Shanyan Guan and Shiyu Li and Xin Li and Zhizhou Chen and Yunzhe Xu and Yuehao Liu and Ruixue Shen and Shiju Zhang and Min Jin and Qiang Liu and Junbo Liu and Penggang Shi and Sheng Sun and Xiao Wang and Yong Zhang and Zhongang Qi and Wei Li and Chao Ma and Ying Tai and Yanhao Ge},
  title = {{BluePixel: Toward Native 4K Agentic Image Editing}},
  year = {2026},
  eprint = {2610.19882},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.19882},
}
Figure