PaperlayerView exact revision on arXiv

Source-linked AI summary

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 Ge

arXiv:2610.19882v1cs.CV

TL;DR

Native-4K editing models have emerged, but their training recipes remain undisclosed, while single end-to-end editors have difficulty with precise adjustments, specialized tasks, and vague instructions. BluePixel pairs a native-4K model and a tool-using agent harness, with a training recipe that adapts a 1K-pretrained editor. On VINS-4KEval++, it ranks second among seven models and second in non-edited-region consistency.

  • Problem

    Recent open-source editors have moved to 4K while their training recipes remain undisclosed, and single-model editing falls short on precise adjustments, specialized tasks, and vague instructions.

    Source: §1 · p. 1

  • Method

    BluePixel combines a native-4K editor adapted from a 1K-pretrained model with an agent harness that routes requests among tools and uses reward-guided feedback.

    Source: §1 · p. 3

  • Results

    BluePixel ranks second only to GPT Image 2 in overall score among seven models, and second in non-edited-region consistency.

    Source: §6 · p. 19

  • Takeaways & Limitations

    The results suggest a path beyond the single-model paradigm that currently dominates image-editing research.

    Source: §6 · p. 19

    The reward model’s training recipe is deferred to a companion paper.

    Source: §4.3 · p. 11

Abstract

from arXiv · show

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.

1 Introduction

Native-4K editing has advanced, but the training recipes of recent open-source models remain undisclosed, and cascaded upscaling can introduce artifacts. BluePixel addresses both efficient native-4K training and limitations of single-model editing with a model-and-agent system.

1 Introduction

  • Cascading a 1K editor with super-resolution can introduce ghosting, texture repetition, and loss of fine-grained control at the upsampling boundary.VINS-120K introduced post-training a 1K-pretrained editor directly into a native-4K model.Source: §1 · p. 1
  • A single end-to-end editor struggles with precise parametric adjustments, specialized tasks such as text rendering, and vague user instructions.The introduction gives “make this photo look more professional” as an example of an underspecified request.Source: §1 · p. 3
  • BluePixel turns a 1K-pretrained editor into a native-4K editor without cascaded super-resolution, using progressive VAE adaptation, continual learning, and reinforcement learning.The paper reports completing the 4K adaptation in four days on 32 Kunlunxin P800 XPUs.Source: §1 · p. 3
  • BluePixel Agent decomposes requests into specialist and general-purpose operations, then uses reward-based feedback to retry or select candidate images.The system also preserves image provenance for multi-turn refinement.Source: §1 · p. 3

2 Related Work

Prior work advances instruction-driven editing and agentic tool use, while BluePixel extends editing to native 4K and links tool orchestration to reward-guided candidate feedback. Its training approach addresses the resource demands of high-resolution editing.

Instruction-driven image editing

  • Instruction-editing research has improved control through methods including test-time optimization, larger training data, posterior sampling, and semantic intent understanding.The paper positions BluePixel as retaining the image-and-instruction interface while lifting editing to native 4K.Source: §2 · p. 3

Training at native 4K

  • BluePixel replaces a standard f=8 VAE with DCAE-F64 (f=64), cutting 4K token count by 16× to the level of a 1K edit.The paper identifies instruction fidelity and non-edited-region consistency as additional demands of editing relative to generation.Source: §2 · p. 3

Agentic image editing

  • BluePixel Agent delegates text, tone, character effects, and general edits to different tools within a structured editing pipeline.Its loop couples tool routing with persistent image provenance and reward-based feedback.Source: §2 · pp. 3–4
  • BluePixel Agent uses a reward model to judge every candidate and drive local retries and final selection, rather than providing only a post hoc quality report.It combines this feedback loop with turn-level routing, persistent image provenance, and specialized editing tools.Source: §2 · pp. 3–4

3 BluePixel

BluePixel targets practical native-4K editing and greater controllability than a single model can provide. It pairs a 1K-pretrained editor adapted for native 4K and four-step inference with an agent harness for routing, tool composition, and feedback.

3 BluePixel

  • BluePixel addresses efficient native-4K editing and controllability through structured tool use and feedback.The paper frames these as two challenges that motivate the system.Source: §3 · p. 4
  • Its training recipe adapts a 1K-pretrained editor into a native-4K editor with four-step inference.Source: §3 · p. 4
  • The agentic harness combines routing, composable editing tools, reward-guided retries, and candidate selection.Source: §3 · p. 4

3.1 Efficient Native 4K Training

BluePixel turns a pretrained low-resolution editor into a native-4K editor through three training phases: efficient initialization, capability refinement, and few-step acceleration. The recipe combines progressive VAE adaptation, further training for quality and specialized skills, and source-anchored distillation.

3.1.1 Phase 1: Preparation — Lightweight 4K Initialization

  • Phase 1 replaces the native VAE with DCAE-F64, reducing 4K token count by 16× to the count of a 1K edit and matching its training memory footprint.Progressive adaptation moves from f=8 to f=32 to f=64 because direct adaptation to f=64 is infeasible; the phase takes about four days on 32 Kunlunxin P800 XPUs with about 200K samples.Source: §3.1.1 · pp. 4–5, §3.1 · p. 4

3.1.2 Phase 2: Enhancement — Quality and Capability Refinement

  • Phase 2 continues training at 1024×1024 before native 4096×4096 fine-tuning, then rebalances data against benchmark failures and merges checkpoints.Native 4K fine-tuning addresses the quality ceiling at 1K, including persistently degraded face rendering; the Phase 1 checkpoint is a useful but limited initialization.Source: §3.1.2 · pp. 5–6, Figure 2 · p. 5
  • ACE-MOPD learns specialized editing skills in groups, projecting later-group gradients away from earlier groups’ LoRA subspaces and consolidating each group into the base model with multi-teacher distillation.The approach targets scarce, numerous vertical-editing tasks while keeping the adapter count bounded.Source: §3.1.2 · p. 6
  • Reinforcement learning first improves aesthetics with PickScore on text-to-image tasks but degrades editing capability, so BluePixReward is then used on image-to-image editing tasks to improve editing performance.Prompts with low reward variance are filtered to improve training efficiency and effectiveness.Source: §3.1.2 · p. 6
  • The vertical-skill adaptation set contains 14 tasks, but teeth retouching and clothing-wrinkle removal have no held-out test set.Source: Table 1 · p. 7

3.1.3 Phase 3: Acceleration — Few-step Distillation

  • Source-anchored target projection (SATP) anchors non-edited regions to the source latent while retaining the teacher’s estimate in edited regions.The teacher prediction separates edited from non-edited positions; the student is trained with standard distillation and a fidelity term on the anchored target.Source: §3.1.3 · p. 7
  • The 4-step student preserves task success rate and background consistency relative to the 50-step teacher, with comparable editing quality and a 4K edit time of 8.7 s on one Kunlunxin P800 XPU.The reported inference speedup is approximately 4×.Source: §3.1.3 · p. 7

3.2 Agentic Editing Harness

BluePixel Agent wraps the native-4K editor in a harness that routes requests to composable skills, evaluates candidate edits, and uses feedback to guide retries and final selection.

3.2 Agentic Editing Harness

  • For compositional edits, BluePixel Agent combines replaceable skills with closed-loop verification, triggering automatic re-editing when an edit fails assessment.The paper says a single model’s capability ceiling on such tasks cannot be resolved by additional training data alone.Source: §3.2 · p. 8

3.2.1 System Architecture

  • The harness separates request interpretation, skill execution, and feedback-driven candidate selection while retaining conversational constraints and image provenance across turns.For underspecified factual references, the intent analyst can use web search to inform the editing instruction.Source: §3.2.1 · p. 8

3.2.2 Skill-Based Editing

  • The skill repertoire combines a primary diffusion editor for open-ended changes, specialist tools for scoped operations, and auxiliary capabilities that guide editing without modifying images.For factual references beyond the source image, the analyst can gather information for instructions, as illustrated by character effects and dialogue and a zodiac-year toy.Source: §3.2.2 · pp. 8–9, §3.2.1 · p. 8, Figure 3 · p. 9

3.2.3 Closed-Loop Editing

  • The harness assesses each candidate against the source image and instruction, uses diagnostic feedback for bounded retries, and selects the final result against the complete request.BluePixReward assesses instruction following, non-edited-region consistency, and visual quality.Source: §3.2.3 · p. 9

3.2.2 Skill-Based Editing

  • For filter requests, BluePixel Agent uses LUT-based color grading, while the compared diffusion-model baselines produce plausible but aesthetically inconsistent results.The figure compares the input image, BluePixel Agent, GPT Image 2, Nano Banana Pro, and Seedream 5.0 Pro.Source: §3.2.1 · p. 8, Figure 4 · p. 10

4 Data Construction and Quality Assessment

BluePixel constructs 1K and 4K editing datasets using quality filters and an edit-evaluation reward model, with 4K examples drawn chiefly from mined video frames and screened upscaled images.

4.1 Overview

  • The training pipeline combines public datasets and in-house pairs, filters for quality and edit accuracy, and organizes 1K and 4K data into eight task types.The two gates assess image quality and score source-edit-instruction triples for instruction following, non-edited-region consistency, visual quality, and overall quality.Source: §4.1 · p. 10

4.2 Image Quality Filtering

  • A Qwen3-VL-8B VLM scores aesthetics, visual clarity, saturation, information density, and watermark presence, replacing a stack of hand-tuned image-quality thresholds.The model outputs scores and reasoned analysis, then makes its own retain-or-discard judgment.Source: §4.2 · p. 11

4.3 Edit Evaluation Reward Model

  • BluePixReward distills expert editing models into a Qwen3.5-9B judge that scores source image, edited image, and instruction triples across four quality dimensions.Its training recipe is deferred to a companion paper.Source: §4.3 · pp. 11–12
  • Pearson correlations with expert ratings on a 150-pair validation set are r = 0.82 (IF), r = 0.78 (NC), r = 0.76 (VQ), and r = 0.84 (Overall).BluePixReward supplies a unified quality signal for training-data filtering, DiffusionNFT reward, and VINS-4KEval++ benchmarking.Source: §4.3 · p. 12

4.4 4K Data Construction

  • Most 4K pairs come from mined video-frame edits; a second stream upscales sharp-screened 1K sources, and all pairs pass the same quality and edit-evaluation filters.MANIQA screens candidate source images for sharpness before upscaling.Source: §4.4 · p. 12

5 Experiments

BluePixel is evaluated on standard-resolution, native-4K, and continual-learning benchmarks, with qualitative comparisons, component analyses, and quantitative results. Its results include second-place VINS-4KEval++ performance, gains across CIE-Bench capability tasks, and competitive ImgEdit-Bench scores.

5.1.1 Qualitative Results

  • BluePixel edits the requested attributes, expressions, and text in VINS-4KEval++ examples while preserving untargeted scene details that some compared models alter or miss.On CIE-Bench examples, it also realizes causal-relation and refocus edits and transfers a garment while preserving the original framing, unlike Seedream 5.0.Source: §5.1.1 · pp. 12–14, Figure 6 · p. 13, Figure 7 · p. 14

5.1.2 Model Analysis

  • The verification-and-retry harness corrects approximate backbone outputs in three examples: expression, text rendering, and mask repair.The backbone overshoots the requested expression, garbles requested text, and leaves a mask crack visible; the agent returns outputs matching those requests.Source: §5.1.2 · p. 14, Figure 8 · p. 15
  • Continual learning improves all five CIE-Bench capability tasks, with the largest gains on lighting (+4.69) and virtual try-on (+4.45).CIE total rises from 6.24 in the base model to 8.87 after continual learning, dips to 8.52 after aesthetic RL, and recovers to 8.57 after reward RL; general-editing quality on VINS-4KEval++ follows the opposite trajectory. Templates use about 15 to 200 samples, and the results show retained few-shot capability over a long sequence of tasks.Source: §5.1.2 · pp. 14–15, Table 3 · p. 16, Figure 9 · p. 16

5.1.3 Quantitative Results

  • On ImgEdit-Bench, BluePixel scores 4.32 Overall, ranks first on Hybrid (4.59) and Action (4.93), ties for first on Add (4.93), and ranks second on Adjust (4.76).The remaining gap to the strongest methods concentrates on Extract and Style.Source: §5.1.3 · p. 16
  • BluePixel ranks second among seven models on VINS-4KEval++ with Overall 8.62 and success rate 63.2%, and ranks second in non-edited-region consistency (NC, 9.06).Its instruction-following (IF, 8.74) and visual quality (VQ, 8.90) scores fall below four external models; its 0.02-point Overall margin over Nano Banana Pro is within seed noise, and it uses prompt rewriting while external models produce one output per case.Source: §5.1.3 · pp. 17–18

6 Conclusion

BluePixel combines a native-4K editor with an agent framework and a disclosed training recipe. It ranks second on VINS-4KEval++ overall and in non-edited-region consistency.

6 Conclusion

  • The training recipe uses progressive VAE adaptation, f: 8→32→64, group-wise ACE-LoRA continual learning with MOPD consolidation, BluePixReward reinforcement learning, and SATP distillation to compress a 50-step teacher to 4 steps.The 4-step model completes a 4K edit in 8.7 s.Source: §6 · p. 18
  • BluePixel ranks second only to GPT Image 2 in overall score among seven models on VINS-4KEval++, and second in non-edited-region consistency.Table 8 reports scores on 2,513 cases shared by all models; BluePixel’s score is an eight-seed mean with prompt rewriting, while each external model has one output per case.Source: Table 8 · p. 19, §6 · p. 19
  • BluePixel Agent decomposes requests across specialist and general tools, grounds underspecified instructions through retrieval and prompt rewriting, and verifies candidates with bounded recovery.A unified data pipeline uses a VLM-based quality gate and a four-dimensional reward model from training-data filtering to final evaluation.Source: §6 · p. 19

The BluePixel Team

The paper credits contributors across research, algorithm design, data construction, evaluation, training, and deployment infrastructure.

The BluePixel Team

  • Shanyan Guan led the project and designed its framework, core algorithms, data construction, and evaluation pipeline; Shiyu Li developed the BluePixel Agent harness.Xin Li contributed to data construction, training-instruction refinement, and instruction recommendation, and developed the reward model for editing evaluation.Source: The BluePixel Team · p. 24
  • Yuehao Liu developed the continual-learning algorithm, Yunzhe Xu the distillation pipeline, and Ruixue Shen the data-construction pipeline and DPO-based post-training exploration.Zhizhou Chen validated the initial training algorithm.Source: The BluePixel Team · p. 24
  • Engineering contributors developed and maintained training and deployment infrastructure, while collaborators provided research guidance and support.The acknowledgments name Chao Ma and Ying Tai as long-term collaborators, and list Chao Ma, Ying Tai, Zhongang Qi, Wei Li, and Yanhao Ge as providing guidance and support.Source: The BluePixel Team · p. 24

A Hairstyle Sub-templates

The hairstyle continual-learning experiments use 26 sub-templates for the Change Hairstyle task.

A Hairstyle Sub-templates

Related papers

Loading 2610.19882v1…