πŸ”‡
4K

BLUEPIXEL
TOWARD NATIVE 4K AGENTIC IMAGE EDITING

A native 4K editing model + an agent framework β€” complete training recipe disclosed.β–Œ
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 Ge β€” vivo
β˜… NATIVE 4K IN / 4K OUT  βœ¦  4-STEP DISTILLATION  βœ¦  8.7 S PER EDIT ON A SINGLE P800 XPU  βœ¦  AGENTIC VERIFY-AND-RETRY  βœ¦  ACE-MOPD CONTINUAL LEARNING  βœ¦  RANK 2ND ON VINS-4KEval++  βœ¦     β˜… NATIVE 4K IN / 4K OUT  βœ¦  4-STEP DISTILLATION  βœ¦  8.7 S PER EDIT ON A SINGLE P800 XPU  βœ¦  AGENTIC VERIFY-AND-RETRY  βœ¦  ACE-MOPD CONTINUAL LEARNING  βœ¦  RANK 2ND ON VINS-4KEval++  βœ¦    
4KΓ—4K
native 4K in, native 4K out β€” no cascade SR
8.7s
per 4K edit β€” 4 steps, bf16, single Kunlunxin P800 XPU
2nd/7
overall on VINS-4KEval++ (3,305 cases); NC 2nd only to GPT Image 2
1st
Hybrid & Action on ImgEdit-Bench (tied 1st on Add)

SHOWCASE EDITING RESULTS & MODEL COMPARISON 🎚

One instruction, the same input, seven editing models. View the complete results, then select a model to compare details with BluePixel.

EDITING INSTRUCTIONZoom in on the man and horse, change the white shirt to a dark brown tunic, and swap the saddle pad for red leather.

INPUT IMAGE Original
Original image for the compound editing example

OTHER MODELS Β· SAME INSTRUCTION

Select a result to compare with BluePixel ↓

THE QUEST MAP πŸ—Ί

One 1K editor, three training stages, one agent β€” every number and picture below comes straight from the paper. Scroll the journey.

PHASE 1 βš™ INITIALIZE

Swap the f=8 VAE for DCAE-F64[1]; progressive 8β†’32β†’64 projection distillation. A native-4K editor in ~4 days on 32 Kunlunxin P800 XPUs.

PHASE 2 βš™ ENHANCE

Progressive SFT β†’ benchmark-driven rebalancing β†’ ACE-MOPD[2,4] continual learning (gradient-orthogonal LoRA[3,5]) β†’ two-stage RL with DiffusionNFT[6] (PickScore[7], then BluePixReward).

PHASE 3 βš™ ACCELERATE

Source-anchored target projection (SATP) compresses 50 steps to 4 while anchoring non-edited regions β€” keeping 4K fidelity at production latency.

β–Ό then an agent takes over β–Ό

BLUEPIXEL AGENT πŸ€–

Routes each request across specialist + general tools, grounds vague instructions, verifies every candidate with BluePixReward, and retries within bounded attempts.

STAGE 1/4 INITIALIZE 4K βš™

Progressive VAE adaptation (8β†’32β†’64) aligns the new DCAE latents with the pretrained editor. Input projection distillation, interface calibration and light backbone adaptation produce an initial 4K model in ~4 days on 32 Kunlunxin XPUs with ~200K samples.

STAGE 2/4 ABSORB THE LONG TAIL 🧠

ACE-MOPD continual learning adds a long tail of vertical skills group by group β€” CIE total 6.24 β†’ 8.87 while general editing moves less than Β±0.03. The template bank, as visualization strips from the paper:
hairstyle templates
hairstyles β€” 26 templates in the bank, each learned from a few dozen samples
expression templates pose templates global effect templates
expression, pose, and global photo effects β€” each strip: the source image, then the template outputs. Every group is distilled back into the base model (MOPD) before the next group starts.

STAGE 3/4 SHARPEN & GO FAST ⚑

Two-stage RL refines aesthetics and instruction fidelity. Then SATP cuts 50 denoising steps to 4 β€” anchoring non-edited regions so 4K fidelity survives.
50 β†’ 4
denoising steps
~4Γ—
inference speedup
8.7s
per native-4K edit, single P800 XPU (bf16)
βœ“
success rate & background consistency preserved

STAGE 4/4 DELEGATE TO THE AGENT πŸ€–

The harness decomposes requests, grounds vague instructions (even web facts like β€œthis year’s zodiac”), verifies every candidate with BluePixReward, and retries within bounded attempts.
BACKBONE
AGENT
verify-and-retry: the backbone garbles β€œDecaf” β€” the agent lands it
BACKBONE
AGENT
verify-and-retry: the backbone overshoots to a stiff grin β€” the agent returns serene
BACKBONE
AGENT
verify-and-retry: the backbone leaves the crack β€” the agent restores the metal
INPUT
AGENT
grounding: β€œthis year’s zodiac” resolved by web search
INPUT
AGENT
specialist tools: effects + dialogue beyond the photo
INPUT
AGENT Β· LUT
specialist route: LUT color grading for a calibrated look

THE SCOREBOARD πŸ†

BenchmarkWhat it showsHeadline
VINS-4KEval++
3,305 native-4K cases
main 4K comparison, 7 models [8]Overall 8.62 β€” 2nd of 7; NC 9.06, 2nd only to GPT Image 2
ImgEdit-Bench[9]nine editing categories1st on Hybrid (4.59) & Action (4.93), tied 1st on Add
CIE-Bench[2]continual learning of vertical skillstotal 6.24 β†’ 8.87 with ACE-MOPD; long-sequence few-shot learning
Speedproduction latency50 steps β†’ 4 steps; a 4K edit in 8.7 s (single P800 XPU, bf16)
Per-seed stability β€” 8-seed mean Overall 8.624, span 0.053:
CIE total across the recipe (SFT β†’ +ACE-MOPD β†’ +RL):

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.

REFERENCES

  1. [1] Junyu Chen et al. Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models. arXiv:2410.10733, 2024.
  2. [2] Yuehao Liu et al. ACE-LoRA: Adaptive Orthogonal Decoupling for Continual Image Editing (introduces CIE-Bench). arXiv:2605.14948, 2026.
  3. [3] Edward J. Hu et al. LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022.
  4. [4] Quanhao Li et al. DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Models. arXiv:2605.15055, 2026.
  5. [5] Yuehao Liu et al. Octopus: History-Free Gradient Orthogonalization for Continual Learning. CVPR 2026.
  6. [6] Kaiwen Zheng et al. DiffusionNFT: Online Diffusion Reinforcement with Forward Process. arXiv:2509.16117, 2025.
  7. [7] Yuval Kirstain et al. Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation. NeurIPS 2023.
  8. [8] Zhizhou Chen et al. VINS-120K: Ultra High-Resolution Image Editing with a Large-Scale Dataset (VINS-4KEval benchmark). CVPR 2026.
  9. [9] Yang Ye et al. ImgEdit: A Unified Image Editing Dataset and Benchmark. arXiv:2505.20275, 2025.

BIBTEX

@article{bluepixel2026,
  title  = {BluePixel: Toward Native 4K Agentic Image Editing},
  author = {Guan, Shanyan and Li, Shiyu and Li, Xin and Chen, Zhizhou and
            Xu, Yunzhe and Liu, Yuehao and Shen, Ruixue and Zhang, Shiju and
            Jin, Min and Liu, Qiang and Liu, Junbo and Shi, Penggang and Sun, Sheng and
            Wang, Xiao and Zhang, Yong and Qi, Zhongang and Li, Wei and
            Ma, Chao and Tai, Ying and Ge, Yanhao},
  journal= {arXiv preprint arXiv:2610.19882},
  url    = {https://arxiv.org/abs/2610.19882},
  year   = {2026}
}