메뉴
Decompose complex image editing instructions into simpler sub-tasks with automatically generated control guidance. Handles multi-object edits, preserves identity of surrounding regions, and eliminates manual mask creation.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/ADu2021/skillXiv --skill complex-image-editing명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | complex-image-editing |
| title | Beyond Simple Edits: X-Planner for Complex Instruction-Based Image Editing |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2507.05259 |
| keywords | ["Image Editing","Instruction Following","Multimodal Planning","Mask Generation","Object Localization"] |
| description | Decompose complex image editing instructions into simpler sub-tasks with automatically generated control guidance. Handles multi-object edits, preserves identity of surrounding regions, and eliminates manual mask creation. |
Editing images based on complex instructions requires more than direct pixel manipulation. When a user says "make the building taller and the sky more dramatic," the system must understand that these are two separate edits targeting different objects, generate precise boundaries for each, and apply appropriate transformations without bleeding into adjacent regions. X-Planner solves this by decomposing complex instructions into manageable sub-tasks, automatically generating the masks and control signals that guide editing models.
The core challenge is that complex instructions are indirectly specified and often target multiple objects. Current approaches either require users to manually provide masks or fail when identity preservation matters—editing one object corrupts its surroundings.
X-Planner operates as a three-stage pipeline that separates planning from execution:
By treating masking as a learned task conditioned on edit type, the system generates tighter masks for texture edits and dilated masks for shape changes—each adapted to the specific editing goal.
Start by analyzing a complex instruction and decomposing it into sub-tasks:
from xplanner.decomposer import InstructionDecomposer
from xplanner.masker import MaskGenerator
decomposer = InstructionDecomposer(model="gpt-4-vision")
# Complex instruction that targets multiple objects implicitly
instruction =
sub_tasks = decomposer.decompose(
instruction=instruction,
image=image
)
For each sub-task, generate a specialized mask conditioned on the edit type:
masker = MaskGenerator()
for sub_task in sub_tasks:
edit_type = sub_task["type"]
# Generate mask adapted to edit type
mask = masker.generate_mask(
image=image,
target_description=sub_task["text"],
edit_type=edit_type,
# Different masks for different edits:
# - "texture" or "color_change": tight mask (exact object)
# - "shape" or "size": dilated mask (include context)
# - "deletion": precise boundary
# - "global": full image mask
)
# Validate mask covers the target
assert masker.validate_coverage(mask, sub_task["target"])
sub_task["mask"] = mask
For insertion tasks, predict bounding boxes since existing detectors can't hallucinate objects not in the original image:
from xplanner.spatial import BoundingBoxPredictor
predictor = BoundingBoxPredictor()
insertion_tasks = [t for t in sub_tasks if t["type"] == "insertion"]
for task in insertion_tasks:
# Predict where new object should appear
bbox = predictor.predict(
image=image,
instruction=task["text"],
context_objects=get_visible_objects(image)
)
# Bbox provides spatial guidance to editing model
task["bbox"] = bbox
Apply the sub-tasks iteratively using a compatible editing model:
from xplanner.executor import ImageEditor
editor = ImageEditor(backend="ultarEdit") # or InstructPix2Pix
result_image = image.copy()
# Apply sub-tasks sequentially
for i, sub_task in enumerate(sub_tasks):
# Get mask and optional spatial guidance
mask = sub_task["mask"]
bbox = sub_task.get("bbox", None)
# Edit using specified mask and guidance
result_image = editor.edit(
image=result_image,
instruction=sub_task["text"],
mask=mask,
spatial_guidance=bbox,
preserve_identity=True # Keep regions outside mask unchanged
)
# Validate edit quality
assert editor.validate_quality(result_image, result_image_prev)
return result_image
Use this approach for:
Avoid X-Planner for:
| Edit Type | Mask Strategy | Example |
|---|---|---|
| Color change | Tight mask (exact object boundary) | "Make the car blue" |
| Shape change | Dilated mask (object + buffer) | "Make the building taller" |
| Style transfer | Full region mask | "Make the road surface glossy" |
| Deletion | Precise boundary | "Remove the traffic cone" |
| Insertion | Bounding box guidance | "Add a tree near the building" |
| Global edit | Full image mask | "Brighten the entire scene" |
| Parameter | Typical Range | Guidance |
|---|---|---|
| Mask dilation | 0-30 pixels | Larger for shape edits, smaller for color |
| Confidence threshold | 0.5-0.9 | Higher = more selective masks |
| Iteration count | 1-5 steps | More iterations for complex edits, but slower |
| Model backbone | GPT-4V, Claude | Larger models decompose better |
"Beyond Simple Edits: X-Planner for Complex Instruction-Based Image Editing" - arXiv:2507.05259