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.
Figures & tables
Figure 1: BluePixel editing showcase. Diverse 4K editing capabilities including text editing, pose editing, scene recomposition, anime-style expression editing, and low-level image editing. The model preserves fine details faithfully: even small text remains crisp and legible.
Figure 2: Phase 1 vs. final model. The Phase 1 checkpoint produces usable but limited results; after Phase 2 enhancement (progressive SFT, benchmark-driven fine-tuning, continual learning, and reinforcement learning), the final model achieves substantially higher editing quality and detail fidelity.
Group
Task
Sub-templates
#Templates
#Samples
Portrait Beauty
Change Hairstyle
side-part, braids, bangs, curly, …(see Table 9 )
26
420
Adjust Expression
wink, cool, closed mouth, laugh, smile, smirk, open eyes
Table 1: The 14 vertical editing tasks used for ACE-LoRA adaptation, spanning portrait beauty and three capability categories. Teeth retouching and clothing-wrinkle removal have no held-out test set.
Figure 3: Agent-assisted editing with supplementary factual context. Top: adding visual effects and dialogue associated with the depicted character. Bottom: adding a plush toy corresponding to the Chinese zodiac year specified in the instruction. Each row shows, left to right: the input image, BluePixel Agent, GPT Image 2, Nano Banana Pro, and Seedream 5.0 Pro.
Figure 4: Filter-style editing comparison. BluePixel Agent vs. GPT Image 2, Nano Banana Pro, and Seedream 5.0 Pro on a filter task requiring professional aesthetic judgment. The three compared baselines produce plausible but aesthetically inconsistent results. BluePixel Agent routes the request to LUT-based (look-up-table) color grading, a specialist editing capability in Table 2 , rather than relying solely on the diffusion model; this produces more accurate and professionally consistent filter effects.
Tier
Capability
Role
Primary editing
BluePixel 4K diffusion editor
Native-4K, open-ended edits to objects, backgrounds, and scene composition
Specialist editing
LUT-based color grading
Global color transformations for filter-style edits
Tone and lighting effects
Preset-based adjustments to illumination and visual atmosphere
Template-guided editing
Application of predefined styles or scene treatments
Mask-guided local editing
Changes restricted to a selected image region
Hairstyle editing
Hair-region edits guided by a user-provided mask
Table 2: Skill-based editing capabilities. Representative editing and supporting capabilities of BluePixel Agent, organized by functional tier.
Figure 5: The 4K set concentrates on three task types , while the 1K set is more evenly distributed: task type proportions for the 1K and 4K training datasets across the eight unified task types.
Figure 6: Main qualitative results on VINS-4KEval++. Each two-row block shows one instruction: compound attribute editing, facial-expression editing, or text restyling. Columns, left to right: the input image, BluePixel, GPT Image 2, Nano Banana Pro, Nano Banana 2.0, Seedream 5.0 Lite, Wan 2.7 Pro, and Kling.
Figure 7: Qualitative comparison on CIE-Bench [ 36 ] . BluePixel vs. Seedream 5.0 on a causal-relation edit and a refocus edit (left), and a virtual try-on (right). BluePixel renders each requested outcome — the badger’s bite, the legible refocused spine text, and the garment transfer at the original framing — whereas Seedream 5.0 shows an intact potato, unchanged spine text, and a full-body shot.
Figure 8: Effect of the agent harness. Three editing instructions (facial expression, text restyling, and mask repair) executed by the BluePixel backbone alone (middle) and by BluePixel Agent (right). The agent produces the requested expression, renders the target word correctly, and fully repairs the cracked mask, while the backbone under-modifies or corrupts these details.
Figure 11
Seed
Overall ↑
Success (%) ↑
Δ vs. Nano Banana Pro
2026
8.653
63.8
+0.052
42
8.651
64.0
+0.050
1337
8.621
63.2
+0.020
1
8.619
63.4
+0.018
3407
8.619
62.4
+0.018
777
8.615
63.0
+0.014
Table 4: Per-seed results on the 2,513 shared VINS-4KEval++ cases (BluePixReward judge, prompt rewriting), ranked by Overall. Δ is the Overall margin over Nano Banana Pro.
Task type
Overall ↑
Success (%) ↑
Add object or element
9.203
79.3
Remove object or element
8.994
72.9
Edit people/faces/character
8.931
74.0
Multi-step (compound)
8.769
62.6
Modify existing object
8.663
62.0
Alter background/texture/style
8.579
56.9
Table 5: Per-task-type results on the full VINS-4KEval++ benchmark (8-seed mean, BluePixReward judge).
Figure 10: ACE-MOPD qualitative visualization: hairstyle manipulation. The leftmost column shows the source image; subsequent columns correspond to six target hairstyles: wind-swept long hair under warm backlight, wind-swept long hair, braids, a short pixie cut, loose waves, and a high ponytail. The edits change hair structure and appearance while retaining subject identity, apparel, pose, and scene layout.
Figure 11: ACE-MOPD qualitative visualization: global image-effect rendering. The leftmost column shows the source image; subsequent columns apply chromatic aberration, directional motion blur, radial zoom blur, vignetting, fisheye distortion, a DV-camera effect, and chromatic noise, respectively. The source image’s semantic content and scene structure are retained under these global appearance transformations.
Figure 12: ACE-MOPD qualitative visualization: facial-expression manipulation. The leftmost column shows the source image; subsequent columns depict a surprised expression, a subtle smile, a wink, and a broad smile. The transformation is localized to the face while preserving identity cues, hairstyle, apparel, viewpoint, and background.
Figure 13: ACE-MOPD qualitative visualization: human pose manipulation. The leftmost column shows the source image; subsequent columns show a relaxed standing pose, a cheek-heart gesture, a head tilt, a peace sign, and a raised-arms pose. The model changes the articulated body configuration while maintaining subject identity, clothing, full-body framing, and the outdoor scene.
Method
Add
Adjust
Extract
Replace
Remove
Background
Style
Hybrid (Compose)
Action
Overall
GPT Image 2
4.85
4.82
4.07
4.92
4.75
4.88
4.96
4.47
4.85
4.73
Qwen-Image-Edit-2511
4.54
4.57
4.13
4.70
4.46
4.36
4.89
4.16
4.81
4.51
Nano Banana 2.0
4.56
4.41
4.04
4.62
4.74
4.54
4.87
4.01
4.81
4.51
Seedream 5.0 Lite
4.93
4.69
3.01
4.41
4.45
4.65
4.93
3.91
4.82
4.42
Emu3.5
4.61
4.32
3.96
4.84
4.58
4.35
4.79
3.69
4.57
4.41
Nano Banana Pro
4.44
4.62
3.42
4.60
4.63
4.32
4.97
3.64
4.69
4.37
Table 6: BluePixel ranks first on Hybrid and Action, ties for first on Add, and remains competitive overall. Main results on ImgEdit-Bench: each cell is the average score (1–5); Overall is the average across the nine task categories. Superscripts distinguish the reporting source for models with multiple reported scores: a FireRed-Image-Edit, b LongCat-Image, c HiDream-O1-Image, d Boogu-Image-0.1.
Method
Instr. Adh. ↑
Nat. ↑
Detail ↑
Nano Banana 2.0
4.68
4.79
4.74
Seedream 4.0
4.60
4.79
4.70
Seedream 5.0 Lite
4.57
4.65
4.61
FLUX.1 Kontext [Dev] + VINS [ 11 ]
4.23
4.70
4.47
FLUX.1 Kontext [Dev]
4.22
4.60
4.41
OmniGen2
4.14
4.54
4.34
Table 7: Quantitative comparison on the VINS-4KEval benchmark [ 11 ] (509 tasks, 4K). BluePixel matches the naturalness of FLUX.1 Kontext [Dev] + VINS while exceeding it on instruction adherence and detail. ImgEditJudge scores each row on instruction following (Instr. Adh.) and naturalness (Nat.); Detail is their average. BluePixel is the final model after the full training recipe. Per-column best in bold, second best underlined.
Method
Overall ↑
IF ↑
NC ↑
VQ ↑
Success (%) ↑
GPT Image 2 [ 39 ]
9.15
9.22
9.34
9.42
74.8
Nano Banana Pro [ 17 ]
8.60
8.80
8.72
9.12
59.4
Nano Banana 2.0 [ 17 ]
8.56
8.81
8.63
9.08
57.9
Wan 2.7 Pro [ 1 ]
8.56
8.72
8.77
9.04
59.5
Seedream 5.0 Lite [ 8 ]
8.44
8.95
8.53
8.81
51.8
Kling [ 28 ]
7.90
8.01
8.37
8.55
45.4
Table 8: BluePixel ranks second on VINS-4KEval++ behind GPT Image 2, with the second-best non-edited-region consistency: BluePixReward scores on the 2,513 cases shared by all models. BluePixel reports the mean of an eight-seed run with prompt rewriting (per-seed detail in Table 4 ); each external model has one output per case. Scores are on a 0–10 scale, with higher values indicating better performance. Success is the percentage of outputs judged successful by the same evaluator.
#
Template
#
Template
#
Template
1
side-part
2
braids
3
bangs
4
half-up princess
5
half-up bun
6
curly
7
glowing strands
8
flying hair
9
short wavy
10
layered perm
11
long wavy
12
latin curl
13
side bangs
14
japanese short curly
15
japanese bangs
16
straight
17
sakura
18
french bangs
Table 9: Full list of hairstyle sub-templates for the Change Hairstyle task.
Modern image editing models produce realistic results but struggle with abstract, multi step instructions (e.g., ``make this advertisement more vegetarian-friendly''). Prior agent based methods decompose such tasks but rely on handcrafted pipelines or teacher imitation, limiting flexibility and decoupling learning from actual editing outcomes. We propose an experiential framework for long-horizon image editing, where a planner generates structured atomic decompositions and an orchestrator selects tools and regions to execute each step. A vision language judge provides outcome-based rewards for instruction adherence and visual quality. The orchestrator is trained to maximize these rewards, and successful trajectories are used to refine the planner. By tightly coupling planning with reward driven execution, our approach yields more coherent and reliable edits than single-step or rule-based multistep baselines.
Anirudh Sundara Rajan, Krishna Kumar Singh, Yong Jae Lee
University of Wisconsin–Madison, WI, USA · Adobe Research, San Jose, CA, USA
With the rapid advancement of commercial multi-modal models, image editing has garnered significant attention due to its widespread applicability in daily life. Despite impressive progress, existing image editing systems, particularly closed-source or proprietary models, often struggle with complex, indirect, or multi-step user instructions. These limitations hinder their ability to perform nuanced, context-aware edits that align with human intent. In this work, we propose ImageEdit-R1, a multi-agent framework for intelligent image editing that leverages reinforcement learning to coordinate high-level decision-making across a set of specialized, pretrained vision-language and generative agents. Each agent is responsible for distinct capabilities--such as understanding user intent, identifying regions of interest, selecting appropriate editing actions, and synthesizing visual content--while reinforcement learning governs their collaboration to ensure coherent and goal-directed behavior. Unlike existing approaches that rely on monolithic models or hand-crafted pipelines, our method treats image editing as a sequential decision-making problem, enabling dynamic and context-aware editing strategies. Experimental results demonstrate that ImageEdit-R1 consistently outperforms both individual closed-source diffusion models and alternative multi-agent framework baselines across multiple image editing datasets.
Instruction guided image editing has advanced substantially with recent generative models, yet it still fails to produce reliable results across many seemingly simple cases. We observe that a large portion of these failures stem not from insufficient model capacity, but from poorly formulated editing tasks, such as those involving small targets, implicit spatial relations, or under-specified instructions. In this work, we frame image editing failures as a task formulation problem and propose an adaptive task reformulation framework that improves editing performance without modifying the underlying model. Our key idea is to transform the original image-instruction pair into a sequence of operations that are dynamically determined and executed by a MLLM agent through analysis, routing, reformulation, and feedback-driven refinement. Experiments on multiple benchmarks, including ImgEdit, PICA, and RePlan, across diverse editing backbones such as Qwen Image Edit and Nano Banana, show consistent improvements, with especially large gains on challenging cases. These results suggest that task reformulation is a critical but underexplored factor, and that substantial gains can be achieved by better matching editing tasks to the effective operating regime of existing models.
Bo Zhao, Kairui Guo, Runnan Du +6
Nanjing University China · Beijing Institute of Technology China · College of Artificial Intelligence, China University of Petroleum (Beijing) China +3