🔇
4K

BLUEPIXEL
TOWARD NATIVE 4K AGENTIC IMAGE EDITING

A native 4K editing model + an agent framework — complete training recipe disclosed.▌
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 — vivo
▶ DEMO WAITLIST ·SOON 📕 PAPER CODE ·SOON 🤗 MODEL MODELSCOPE
★ NATIVE 4K IN / 4K OUT  ✦  4-STEP DISTILLATION  ✦  8.7 S PER EDIT ON A SINGLE P800 XPU  ✦  AGENTIC VERIFY-AND-RETRY  ✦  ACE-MOPD CONTINUAL LEARNING  ✦  RANK 2ND ON VINS-4KEval++  ✦     ★ NATIVE 4K IN / 4K OUT  ✦  4-STEP DISTILLATION  ✦  8.7 S PER EDIT ON A SINGLE P800 XPU  ✦  AGENTIC VERIFY-AND-RETRY  ✦  ACE-MOPD CONTINUAL LEARNING  ✦  RANK 2ND ON VINS-4KEval++  ✦    
4K×4K
native 4K in, native 4K out — no cascade SR
8.7s
per 4K edit — 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)

THE QUEST MAP 🗺

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

PHASE 1 ⚙ INITIALIZE

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

PHASE 2 ⚙ ENHANCE

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

PHASE 3 ⚙ ACCELERATE

Background-anchored distillation compresses 50 steps to 4 while anchoring non-edited regions — keeping 4K fidelity at production latency.

▼ then an agent takes over ▼

BLUEPIXEL AGENT 🤖

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 ⚙

Rewire a 1K editor for native 4K: DCAE-F64 + progressive 8→32→64 adaptation. ~4 days, 32 Kunlunxin XPUs, 200K samples. Phase 1 already lands the edit — the later stages make it faithful and crisp. Drag the gear:
Phase 1 output
final output
⚙
PHASE 1 FINAL · ALL STAGES
same instruction, same input — Phase 1 output vs the finished model.

STAGE 2/4 ABSORB THE LONG TAIL 🧠

ACE-MOPD continual learning adds a long tail of vertical skills group by group — CIE total 6.24 → 8.87 while general editing moves less than ±0.03. The template bank, as visualization strips from the paper:
hairstyle templates
hairstyles — 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 — 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 background-anchored distillation cuts 50 denoising steps to 4 — anchoring non-edited regions so 4K fidelity survives.
50 → 4
denoising steps
~4×
inference speedup
8.7s
per native-4K edit, single P800 XPU (bf16)
✓
success rate & background consistency preserved

STAGE 4/4 DELEGATE TO THE AGENT 🤖

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

BONUS BEFORE / AFTER — DRAG THE BENCHMARK 🎚

The paper’s three headline examples, panel by panel. Drag the wand — and compare the whole field while you’re at it.
input image
BluePixel output
🪄
AFTER · BLUEPIXEL BEFORE · INPUT
COMPARE THE FIELD ▸
The same job takes 8.7 s at native 4K — registration opens soon.

THE SCOREBOARD 🏆

BenchmarkWhat it showsHeadline
VINS-4KEval++
3,305 native-4K cases		main 4K comparison, 7 models [8]Overall 8.62 — 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 → 8.87 with ACE-MOPD; long-sequence few-shot learning
Speedproduction latency50 steps → 4 steps; a 4K edit in 8.7 s (single P800 XPU, bf16)
Per-seed stability — 8-seed mean Overall 8.624, span 0.053:
CIE total across the recipe (SFT → +ACE-MOPD → +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, a native-4K editing model with an agent framework; the complete training recipe is disclosed in this paper. To make 4K training practical, BluePixel turns a 1K-pretrained editor into a native-4K one through a 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. Background-anchored distillation 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. For controllability, BluePixel Agent decomposes user instructions into specialist and general operations, verifies every result with the same reward model that guided its training, and retries within bounded attempts. On VINS-4KEval++ (3,305 native-4K editing cases), BluePixel ranks second among seven models, with non-edited-region consistency second only to GPT Image 2.

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},  % TODO: fill in arXiv id when available
  year   = {2026}
}