Datasets:

TaobaoTmall-AlgorithmProducts
/

CPI-benchmark

+{
+    "Material Modification": "\nYou are a data rater specializing in grading material modification edits. You will be given two images (before and after editing) and the editing instruction. Your task is to evaluate the material change on a 5-point scale from three perspectives:\n\nInstruction Compliance\n1 Target material not changed, wrong object edited, or geometry materially altered.\n2 Correct object but wrong target material; only part of surface modified; other objects also affected.\n3 Main material correct but noticeable deviation in texture/roughness/metallic properties; small collateral edits.\n4 Requested objects modified accurately; only tiny texture mismatch or very minor edge artifacts.\n5 Exactly and only the specified regions modified; texture/reflectivity/roughness precisely match the prompt; shapes fully intact.\n\nVisual Seamlessness\n1 Severe texture tiling, blurring, noise, or obvious copy-paste artifacts.\n2 Clear edge bleeding/halos; abrupt resolution shift between old and new material; inconsistent grain.\n3 Overall OK but local lighting/texture mismatch; material boundary slightly visible.\n4 Smooth transitions; micro-details preserved; artifacts only when zoomed.\n5 No detectable edit traces; new material fuses perfectly with original scene lighting and perspective.\n\nPhysical Consistency\n1 Severe violation of physics (e.g., liquid looks solid, glass lacks refraction); highlights/shadows broken.\n2 Noticeable mismatch in specular response, roughness, or subsurface scattering; background distortion.\n3 Material properties mostly plausible with minor errors in reflection or shadow interaction.\n4 Material change respects physical laws and scene illumination; details like bumps/pores well preserved.\n5 Physically excellent: material looks authentic with correct Fresnel effects, anisotropy, and light interaction.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nInstruction Compliance: A number from 1 to 5.\nVisual Seamlessness: A number from 1 to 5.\nPhysical Consistency: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Environment Change": "\nYou are a data rater specializing in grading comprehensive environment change edits. You will be given two images (before and after editing) and the editing instruction. Your task is to evaluate changes involving weather, season, time/era, lighting, or specific tonal adjustments (e.g., exposure, contrast, temperature) on a 5-point scale from three perspectives:\n\nInstruction Compliance\n1 No environmental shift; wrong scene altered; geometry distorted; or specific numeric values (exposure/temp) ignored.\n2 Correct category (e.g., weather) but wrong target (rain vs snow); partial update; unintended object changes; values off significantly.\n3 Main shift correct (e.g., day to night) but intensity/tone mismatch; numeric parameters slightly inaccurate; minor artifacts.\n4 Requested change applied accurately; slight deviation in specific numeric values (e.g., +30 instead of +36) or color grade.\n5 Exactly and only the requested shift applied; weather/season/time/light perfectly matched; all numeric tonal values precise; scene intact.\n\nVisual Seamlessness\n1 Severe banding, halos, ghosting, unnatural global casts, or broken sky-ground transitions; image unusable.\n2 Abrupt transitions between edited/unedited areas; inconsistent noise/grain; lighting direction conflicts.\n3 Overall coherent but local lighting/shadow conflicts; edges of atmospheric elements slightly visible.\n4 Smooth global integration; grain/noise consistent with new atmosphere; artifacts only when zoomed.\n5 No detectable edit traces; new weather/time/light fuses naturally with original scene geometry and depth.\n\nPhysical Consistency\n1 Impossible physics (e.g., shadows opposite sun, snow melting instantly, wrong era tech); broken light logic.\n2 Mismatched shadow direction/intensity for new time/weather; incorrect reflection colors; floating particles.\n3 Lighting mostly plausible; minor errors in ambient occlusion, fog density, or historical accuracy.\n4 Light interaction respects scene geometry; shadows/highlights align perfectly with new weather/time/era.\n5 Physically perfect: accurate atmospheric scattering, ray tracing logic, and material response to specific lighting conditions.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nInstruction Compliance: A number from 1 to 5.\nVisual Seamlessness: A number from 1 to 5.\nPhysical Consistency: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Subject Re-orientation": "\nYou are a data rater specializing in grading subject re-orientation edits (object rotation while camera remains static). You will be given two images (before and after editing) and the editing instruction. Your task is to evaluate the object transformation on a 5-point scale from three perspectives:\n\nInstruction Compliance\n1  No object rotation, or wrong rotation direction; background accidentally moved/distorted; object identity severely altered.\n2  Partial rotation but clearly incorrect angle/axis; significant unintended changes to background or object shape.\n3  Rotation broadly matches the instruction but with noticeable errors in degree, occluded parts, or object pose stability.\n4  Rotation matches well; background remains perfectly static; object identity preserved; only minor angular deviations.\n5  Exact re-orientation as requested: object rotates precisely per prompt, background is unchanged, no unrelated edits.\n\nVisual Seamlessness\n1  Heavy artifacts at object boundaries, broken geometry, severe blur/noise, or visible background warping.\n2  Obvious texture stretching, cut-out halos around the rotating object, or background misalignment.\n3  Acceptable but visible inconsistencies in edge blending or texture sharpness on the rotated surface.\n4  Clean render; minor artifacts only on close inspection; background remains crisp and untouched.\n5  Highly natural: textures, edges, and resolution remain consistent; rotation looks physically plausible without background disturbance.\n\nPhysical Consistency (Geometry, Self-Occlusion, Lighting)\n1  Severe 3D inconsistency: impossible geometry during rotation, wrong self-occlusion, lighting conflicts with fixed camera position.\n2  Noticeable depth errors; shadows/highlights do not update correctly for the new object angle relative to fixed light/camera.\n3  Mostly believable with minor self-occlusion glitches or slight lighting mismatches on newly revealed surfaces.\n4  Geometry and self-occlusions largely correct; lighting/shadows update naturally consistent with the original scene's fixed light source.\n5  Excellent 3D realism: correct self-occlusion as object turns, plausible newly revealed surfaces, perfect lighting consistency with static camera/environment.\n\nCRITICAL CONSTRAINT: The second and third scores must NOT be higher than the first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nInstruction Compliance: A number from 1 to 5.\nVisual Seamlessness: A number from 1 to 5.\nPhysical Consistency: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "PS Human": "\nYou are a senior data rater specializing in grading all types of portrait editing tasks. You will be given a reference image (source), an edited result image, and a specific editing instruction. Your scope covers everything from micro-skin repairs and makeup application to macro transformations involving body, age, gender, hairstyle, and clothing.\n\nEvaluate the result on a 5-point scale from the following three core perspectives:\n\n1. Instruction Compliance & Edit Effect\n   - Assess whether the edit accurately executed the instruction (e.g., were acne spots removed? Is the lipstick color correct? Is the face slimmer? Did the gender/age/hairstyle transform successfully?).\n   1  Failed to execute or wrong operation applied (e.g., changed features when asked to remove spots).\n   2  Effect too weak/strong, partial success, or clearly mismatched style/parameters.\n   3  Main intent achieved but inconsistent (e.g., spots not fully gone, background warped by slimming, uneven makeup).\n   4  Goal achieved perfectly; only minor deviations in strength or edge transitions.\n   5  Perfect match: The edit (whether micro-retouch or macro-transform) matches the requested scope, intensity, and style exactly.\n\n2. Visual Naturalness & Detail Quality\n   - Evaluate the blending of edited regions, texture preservation, and artifacts. Focus on skin texture for micro-tasks and lighting coherence for macro-tasks.\n   1  Severe distortion: Plastic skin, obvious smearing, color banding, halos, hard edges, or total loss of texture.\n   2  Obvious artificial traces: Over-smoothing blurring features, floating makeup, unnatural lighting, or destroyed skin texture.\n   3  Generally acceptable but visible retouch traces upon close inspection (slight blur, uneven tone, hard edges).\n   4  Natural-looking results; edited areas blend well; only tiny flaws visible when zoomed.\n   5  Professional quality: Seamless editing, perfect texture retention/reconstruction, correct lighting logic, looks like a native photo or pro retouch.\n\n3. Identity Consistency & Structural Integrity\n   - For Non-Transform Tasks (Beauty/Makeup): Strictly preserve original identity and bone structure.\n   - For Transform Tasks (Gender/Age/Body/Hair): Maintain logical coherence and anatomical correctness under the new setting. No limb breakage or facial collapse allowed.\n   1  Identity confusion or structural collapse: Looks like a different person (for non-transform tasks), or missing limbs, melted features, severe deformities.\n   2  Severe identity drift or structural anomalies: Distorted features, weird joint angles, disproportionate bodies, or incoherent transformation.\n   3  Identity mostly preserved (or transformation logic holds), but with local deformations (shifted ears, twisted neck, slight asymmetry).\n   4  Identity clear (or transformation coherent); anatomy is normal with no obvious deformities, only minor proportion fluctuations.\n   5  Perfect consistency/coherence: For non-transform tasks, identity is unchanged; for transform tasks, the new form is structurally perfect, proportional, and logically self-consistent.\n\n*** CRITICAL CONSTRAINT ***\nThe scores for Dimension 2 (Visual Naturalness) and Dimension 3 (Identity/Structure) MUST NOT be higher than the score for Dimension 1 (Instruction Compliance)!!!\nIf the instruction is not followed, the overall quality is capped by the compliance score.\n\nExample Response Format:\nBrief reasoning: A short explanation based on the criteria above, max 25 words.\nInstruction Compliance: [1-5]\nVisual Naturalness & Detail Quality: [1-5]\nIdentity Consistency & Structural Integrity: [1-5]\nEditing instruction is: <edit_prompt>\n\nBelow are the reference and edited images:\n",
+    "Low Level": "\nYou are an expert data rater specializing in grading low-level image enhancement edits. Your scope includes: Colorization, Old Photo Restoration, Low-light Enhancement, De-raining, De-snowing, De-hazing, Deblurring, Super-Resolution, Inpainting, Outpainting, and any combination of these tasks.\n\nYou will be provided with two images (Input and Output) and a specific editing instruction. Your task is to evaluate the enhancement quality on a 5-point scale based on the following three dimensions:\n\n1. Instruction Compliance\n   - 1: No enhancement applied, or the wrong type of enhancement was used (e.g., deblurring requested but colorization applied).\n   - 2: Correct direction but the effect is extremely weak or severely overdone; key artifacts remain or new major problems are introduced.\n   - 3: Clear improvement visible, but uneven across regions; some details are still lost or slightly over-processed.\n   - 4: Enhancement matches the instruction well; only minor strength mismatches or small residual noise/blur remain.\n   - 5: Perfect match to the instruction: intended defects are completely removed, and details are recovered naturally without deviation.\n\n2. Visual Naturalness\n   - 1: Image severely damaged by processing: heavy ringing, mosaics, unnatural hallucinated patterns, or severe color shifts.\n   - 2: Obvious over-sharpening or over-smoothing; textures look plastic/waxy, or lighting looks physically impossible.\n   - 3: Mostly natural; minor ringing, slight waxy areas, or subtle color drift noticeable upon inspection.\n   - 4: Natural look retained; artifacts are only visible when zoomed in significantly.\n   - 5: Clean and authentic: The result looks like a higher-quality original capture with no signs of artificial processing.\n\n3. Physical & Detail Consistency\n   - 1: Important structures changed or warped; faces/objects deformed; content logic violated (especially in inpainting/outpainting).\n   - 2: Noticeable geometry/edge inconsistencies or strong texture hallucination that contradicts the scene context.\n   - 3: Structures largely preserved; small local inconsistencies are acceptable but noticeable.\n   - 4: Fine details recovered while strictly preserving original structures, materials, and geometric integrity.\n   - 5: Excellent fidelity: Micro-textures restored accurately without hallucination; structure, identity, and perspective fully preserved.\n\nCRITICAL CONSTRAINT: The scores for \"Visual Naturalness\" and \"Physical & Detail Consistency\" must NOT be higher than the score for \"Instruction Compliance\".\n\nOutput Format:\nBrief reasoning: [A concise explanation of the score based on the criteria above, max 20 words]\nInstruction Compliance: [Integer 1-5]\nVisual Naturalness: [Integer 1-5]\nPhysical & Detail Consistency: [Integer 1-5]\n\nEditing Instruction: <edit_prompt>\n\nBelow are the images before and after editing:\n",
+    "Subject-Driven Generation": "\nYou are a data rater specializing in grading Subject-Driven Generation tasks. You will be given a reference image (source subject/style) and a generated image, along with an editing instruction. Your task is to evaluate the result on a 5-point scale from three perspectives:\n\n1. Subject/Style Consistency (Identity & Concept Adherence)\n   - For ID Preservation tasks: Does the specific person/object retain their unique features (face, body, distinct markings) in the new context?\n   - For Style/Transformation tasks: Does the output accurately reflect the target style (e.g., Lego, plushie, specific art style) while maintaining the core essence of the source subject?\n   1  Complete failure: Wrong subject, wrong style, or unrecognizable transformation.\n   2  Strong drift: Vague resemblance only; key identity features lost or style incorrectly applied.\n   3  Main traits present but noticeable errors in details, style texture, or specific identity markers.\n   4  Clearly the same subject/style; only minor deviations in fine details or accessory consistency.\n   5  Perfect adherence: The subject/style is flawlessly preserved or transformed exactly as instructed.\n\n2. Structural Stability (Deformation & Logic)\n   1  Severe collapse: Broken limbs, melted faces, impossible geometry, or missing critical parts.\n   2  Obvious structural issues: Warped features, unnatural joint angles, or logic breaks in the transformation (e.g., floating parts).\n   3  Generally stable but with local deformation (slightly crooked lines, minor asymmetry, small artifacts).\n   4  Anatomy and structure look normal; only tiny distortions visible on close inspection.\n   5  Completely stable structure: Clean geometry, correct proportions, and no visible collapse artifacts.\n\n3. Overall Realism & Image Quality (or Style Coherence)\n   - Note: For non-photorealistic styles (e.g., Lego, cartoon), \"Realism\" refers to the quality and coherence of that specific style.\n   1  Very poor quality: Extreme noise, blur, heavy artifacts, or incoherent rendering.\n   2  Clearly synthetic/low-effort: Harsh edges, unnatural textures, or inconsistent lighting/shadows within the style.\n   3  Acceptable quality but with visible issues (waxy texture, slight halo, mild over-sharpening).\n   4  High quality: Natural texture (or consistent stylized texture), believable lighting, and good background integration.\n   5  Excellent quality: Looks like a professional creation in the target style; seamless blending and perfect texture.\n\nCRITICAL CONSTRAINT: The scores for \"Structural Stability\" and \"Overall Realism & Image Quality\" must NOT be higher than the \"Subject/Style Consistency\" score.\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nSubject/Style Consistency: [Number 1-5]\nStructural Stability: [Number 1-5]\nOverall Realism & Image Quality: [Number 1-5]\n\nEditing instruction: <edit_prompt>\n\nBelow are the reference and generated images:\n",
+    "Feature Extraction": "\nYou are a data rater specializing in evaluating computer vision feature extraction results (including Canny edges, MLSD lines, HED edges, depth maps, sketches, segmentation masks, object detection boxes, normal maps, human poses, SMPL models, and keypoints). You will be provided with an input image, the extracted feature maps/results, and the corresponding instruction. Your task is to rate the extraction quality on a 1-5 scale across the following three dimensions:\n\nInstruction Compliance\n1  Incorrect feature type extracted, or output is completely empty.\n2  Correct feature type but missing major structures; contains extensive false positives or false negatives.\n3  Captures major structures but exhibits obvious breaks, noise, or incomplete boundaries.\n4  High accuracy; only minor detail missing or slight boundary inaccuracies.\n5  Perfect match: All required features are precisely extracted with no omissions or hallucinations.\n\nVisual Naturalness & Clarity\n1  Unrecognizable output: severe noise, chaotic patterns, or illegible maps.\n2  Extremely noisy or very blurry; feature lines/masks are broken or boundaries are unclear.\n3  Generally clear; some fragmentation or minor artifacts present.\n4  Clean and sharp; artifacts visible only upon close inspection.\n5  Clear and precise: clean lines, smooth masks, or exact coordinates with no noise interference.\n\nPhysical & Structural Consistency\n1  Severe geometric distortion; features completely misaligned with input image content.\n2  Obvious misalignment; object shapes distorted, or depth/normal values inverted.\n3  Basically aligned; acceptable minor local offsets or slight topological errors.\n4  Excellent alignment; preserves fine details while maintaining global structure.\n5  Perfect fidelity: Pixel-perfect alignment with the source image; strictly maintains topology and geometric structure.\n\nNote: Scores for the second and third dimensions must not exceed the score for the first dimension!!!\n\nExample Response Format:\nBrief Reason: A short explanation based on the criteria above, no more than 20 words.\nInstruction Compliance: A number between 1 and 5.\nVisual Naturalness & Clarity: A number between 1 and 5.\nPhysical & Structural Consistency: A number between 1 and 5.\nEdit Prompt: <edit_prompt>.\n\nBelow are the input image and the extracted feature results:\n",
+    "Structure-Guided Generation": "\nYou are a data rater specializing in grading structure-guided image generation (generating images conditioned on Canny, MLSD, HED, Depth, Sketch, Segmentation, Detection, Normal, Pose, SMPL, Landmarks, or Layout). You will be given the input structural guide, the generated image, and the corresponding instruction. Your task is to evaluate the generation quality on a 5-point scale from three perspectives:\n\nInstruction Compliance & Structural Adherence\n1  Complete failure: Generated image ignores the structural guide entirely or uses the wrong condition type.\n2  Poor adherence: Major structures (edges, poses, depth layers) are distorted, missing, or significantly misaligned.\n3  Moderate adherence: Global structure is preserved, but local details drift from the guide; some elements misplaced.\n4  Strong adherence: Structure is well-followed; only minor deviations in fine details or boundary alignment.\n5  Perfect adherence: The generated content strictly follows the geometric/topological constraints of the input guide while fulfilling the prompt.\n\nVisual Naturalness & Aesthetic Quality\n1  Severe artifacts: Heavy distortion, nonsensical textures, or broken objects making the image unusable.\n2  Obvious flaws: Unnatural lighting, weird textures, or inconsistent styles that break immersion.\n3  Generally good: Minor texture inconsistencies or slight awkwardness in complex regions.\n4  High quality: Realistic textures and lighting; artifacts only visible upon close inspection.\n5  Photorealistic/Stylistically Perfect: Flawless integration of content; looks like a high-quality original photograph or artwork.\n\nSemantic & Detail Consistency\n1  Semantic collapse: Objects do not match the prompt or guide logic (e.g., extra limbs, floating objects).\n2  Noticeable errors: Wrong materials, inconsistent shadows relative to depth/normals, or illogical object interactions.\n3  Mostly consistent: Small logical errors or slight material mismatches acceptable.\n4  Highly consistent: Details align perfectly with structural cues (e.g., shadows match depth, folds match pose).\n5  Exceptional fidelity: Micro-details are logically derived from the structure; perfect harmony between prompt semantics and structural geometry.\nThe second and third scores should not be higher than the first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nInstruction Compliance & Structural Adherence: A number from 1 to 5.\nVisual Naturalness & Aesthetic Quality: A number from 1 to 5.\nSemantic & Detail Consistency: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the structural guide and the generated image:\n",
+    "Ref-Based Motion": "\nYou are a data rater specializing in grading reference-based motion edits. You will be given a source image, a reference image (defining the target pose/motion), and the generated result. Your task is to evaluate the motion transfer on a 5-point scale from three perspectives:\n\nMotion Fidelity (Pose & Action Transfer)\n1  Target motion not applied; wrong pose from reference; or subject identity severely altered.\n2  Partial motion transfer; key limbs/joints mismatch the reference; action direction incorrect.\n3  Main pose matches reference but details off (finger angles, spine curvature, dynamic balance).\n4  Pose accurately transferred; only minor deviations in extreme joints or subtle timing nuances.\n5  Exact match: Every limb, gesture, and body posture perfectly aligns with the reference image while keeping source identity.\n\nIdentity Preservation\n1  Subject unrecognizable; face/body replaced by reference subject.\n2  Strong drift: Key facial features, clothing, or body shape heavily altered to match reference too closely.\n3  Mostly same identity; moderate changes in some features but still recognizable as source.\n4  Identity clearly preserved; only subtle lighting or texture shifts due to pose change.\n5  Perfect preservation: Source face, hairstyle, skin tone, and clothing details remain intact despite new pose.\n\nVisual & Anatomical Coherence\n1  Severe artifacts: Broken/duplicated limbs, extreme distortion, heavy noise/blur.\n2  Noticeable cut-out halos, proportion errors, or lighting inconsistent with the new pose.\n3  Generally plausible; minor joint shading issues or small noise acceptable.\n4  Clean render; anatomy, lighting, and depth consistent with the transferred motion; flaws only on close inspection.\n5  Flawless realism: Perfect anatomy, natural shadows/highlights for the new pose, and seamless texture continuity.\n\nCRITICAL CONSTRAINT: The scores for \"Identity Preservation\" and \"Visual & Anatomical Coherence\" must NOT be higher than the \"Motion Fidelity\" score.\n\nExample Response Format:\nBrief reasoning: A short explanation based on the criteria above, max 20 words.\nMotion Fidelity: [Number 1-5]\nIdentity Preservation: [Number 1-5]\nVisual & Anatomical Coherence: [Number 1-5]\n\nEditing instruction: <edit_prompt>\nBelow are the source, reference, and generated images:\n",
+    "Ref-Based Removal": "\nYou are a data rater specializing in grading reference-based object removal. You will be given a source image, a reference image (indicating the target object needed to be removed in the source image), and the edited result. Your task is to evaluate the removal quality on a 5-point scale:\n\nInstruction Compliance (Removal Accuracy)\n1  Target object not removed, or unrelated objects deleted based on reference.\n2  Object partly removed; remnants visible; or wrong instance deleted compared to reference intent.\n3  Target mostly removed but extra background elements deleted; faint traces remain.\n4  Only specified objects removed; tiny background items mistakenly deleted; count matches reference.\n5  Perfect: All and only requested objects removed exactly as implied by the reference; scene otherwise untouched.\n\nVisual Naturalness (Inpainting Quality)\n1  Image broken: Large holes, strong artifacts, or obvious copy-paste from reference causing mismatches.\n2  Clear erase marks; color/resolution mismatch between filled area and original; background not restored naturally.\n3  General look acceptable; lighting/color clashes slightly; blur or noise visible in filled regions.\n4  Style consistent; minor edge issues visible only when zoomed; seamless blend with surrounding context.\n5  Seamless: Removal is virtually impossible to spot; filled content looks native to the scene.\n\nPhysical & Detail Integrity\n1  Severe physical errors: Floating items, wrong perspective/light; background heavily warped to fit reference.\n2  Large unfilled gaps or obvious background shifts inconsistent with scene geometry.\n3  Lighting, perspective, and contacts mostly correct; flaws small and tolerable.\n4  Background reconstruction clean; existing details preserved; minute changes outside removal area.\n5  Physically flawless: Accurate light/shadow/texture infill that respects scene logic; high-quality micro-details.\n\nCRITICAL CONSTRAINT: Scores for \"Visual Naturalness\" and \"Physical & Detail Integrity\" must NOT exceed \"Instruction Compliance\".\n\nExample Response Format:\nBrief reasoning: Max 20 words explaining the score.\nInstruction Compliance: [1-5]\nVisual Naturalness: [1-5]\nPhysical & Detail Integrity: [1-5]\n\nEditing instruction: <edit_prompt>\nBelow are the source, reference, and edited images:\n",
+    "Ref-Based Addition": "\nYou are a data rater specializing in grading reference-based object addition. You will be given a source image, a reference image (containing the object to add), and the edited result. Evaluate the added object on a 5-point scale:\n\nPrompt Compliance (Object & Attributes)\n1  Nothing added, or added content is corrupt/unrelated to the reference object.\n2  Added object is wrong class or lacks key attributes (color, size, orientation) from reference.\n3  Correct class added, but key attributes (position, scale, specific details) deviate from reference.\n4  Main attributes correct matching reference; only minor details off or 1-2 small features missing.\n5  Every stated attribute correct; added object perfectly matches the reference instance in all aspects.\n\nVisual Naturalness (Integration)\n1  Image badly broken; obvious paste marks; style/resolution/palette strongly mismatch source.\n2  Strong clash in lighting or colors; added object looks like a foreign cut-out.\n3  General style similar, but lighting direction or color temperature clearly conflicts with source.\n4  Style almost uniform; small edge issues or slight shadow mismatch visible only when zoomed.\n5  Perfect blend; added object interacts naturally with source lighting; no visible difference.\n\nPhysical & Detail Coherence\n1  Severe physical errors: Floating, wrong perspective, occlusion handled poorly; background distorted.\n2  Contact shadows missing or wrong; background visibly shifted or noisy around insertion point.\n3  Lighting, perspective, and contact mostly correct; remaining flaws small and acceptable.\n4  Shadows, reflections, and material response believable; no loss of original detail; background stable.\n5  Added object enhances realism: Precise highlights, accurate shadows, ambient effects; background untouched.\n\nCRITICAL CONSTRAINT: Scores for \"Visual Naturalness\" and \"Physical & Detail Coherence\" must NOT exceed \"Prompt Compliance\".\n\nExample Response Format:\nBrief reasoning: Max 20 words.\nPrompt Compliance: [1-5]\nVisual Naturalness: [1-5]\nPhysical & Detail Coherence: [1-5]\n\nEditing instruction: <edit_prompt>\nBelow are the source, reference, and edited images:\n",
+    "Ref-Based Change": "\nYou are a senior data rater specializing in grading reference-based physical and environmental changes. You will be given a Source Image, one Reference Images, and the Edited Result. \n\nYour task covers specific non-style transfer tasks including:\n1. Object Attribute Transfer: Changing an object's color, material, texture, or pattern in the source to match specific attributes found in the reference.\n2. Background Replacement: Replacing the source background with the background from the reference image while keeping the source foreground intact.\n3. Environmental Transfer: Adapting the source scene's lighting, weather, season, or time of day to exactly match the conditions shown in the reference image.\n\nEvaluate the result on a 5-point scale from three perspectives:\n\n1. Instruction Compliance & Attribute Fidelity\n   - Assess if the specific attribute (color, material, background content, weather condition, lighting direction) was accurately transferred from the reference to the source.\n   1  Failed to apply the reference attribute; wrong object/region edited; or completely ignored the reference content (e.g., sunny instead of rainy).\n   2  Correct category attempted but wrong target value (e.g., wrong color shade, wrong background scene, wrong weather type); partial application.\n   3  Main attribute transferred (e.g., correct season/weather/background) but noticeable deviations in specific details (e.g., wrong cloud density, slightly off material hue).\n   4  Attribute accurately matched; only minor discrepancies in intensity, coverage, or specific sub-features compared to the reference.\n   5  Perfect fidelity: The specific attribute (material/color/weather/light/background) matches the reference image precisely in all aspects; no unintended changes.\n\n2. Visual Naturalness & Integration\n   - Evaluate how seamlessly the new attribute blends with the source image. Focus on edge blending for backgrounds, texture continuity for materials, and atmospheric cohesion for weather/light.\n   1  Severe artifacts: Obvious cut-out halos, heavy banding, mismatched resolution, or \"pasted\" look that breaks immersion.\n   2  Clear visual conflicts: Lighting direction clashes between source subject and new environment; color bleeding; abrupt transitions at boundaries.\n   3  Generally acceptable but visible inconsistencies: Slight tone mismatch, soft edges, or grain/noise differences between the transferred element and the original scene.\n   4  Smooth integration: Transitions are natural; lighting and color grading feel unified; artifacts only visible upon close inspection.\n   5  Seamless fusion: The change looks native to the source image; perfect blending of edges, textures, and atmospheric effects; indistinguishable from a real capture.\n\n3. Physical & Logical Consistency\n   - Crucial for environmental/attribute changes. Check for correct shadow casting, reflection updates, refraction, perspective alignment, and logical interaction (e.g., snow accumulating on objects, wet surfaces reflecting light).\n   1  Physically impossible: Shadows point wrong way for new light source; floating subjects; background perspective mismatch; weather effects ignoring geometry (e.g., rain passing through roofs).\n   2  Noticeable logic errors: Missing contact shadows; incorrect reflection colors; background scale/perspective inconsistent with foreground; lighting intensity mismatch.\n   3  Mostly plausible with minor flaws: Some shadows missing or weak; slight inconsistency in how weather interacts with specific materials; small perspective drift.\n   4  High physical accuracy: Shadows, reflections, and occlusions correctly updated to match the new environment/lighting; perspective aligns well.\n   5  Physically flawless: Perfect light interaction (specular/diffuse), accurate atmospheric scattering, correct shadow lengths/directions, and logical material response (e.g., wetness, frost) consistent with the reference condition.\n\nCRITICAL CONSTRAINT: The scores for \"Visual Naturalness\" and \"Physical & Logical Consistency\" MUST NOT be higher than the score for \"Instruction Compliance\". If the attribute is not correctly transferred, the quality scores are capped.\n\nExample Response Format:\nBrief reasoning: A concise explanation based on the criteria above, max 25 words.\nInstruction Compliance & Attribute Fidelity: [1-5]\nVisual Naturalness & Integration: [1-5]\nPhysical & Logical Consistency: [1-5]\n\nEditing instruction: <edit_prompt>\nBelow are the source, reference, and edited images:\n",
+    "Ref-Based Text Modification": "\nYou are a senior data rater specializing in grading reference-based text editing tasks. You will be provided with:\n1. A Source Image (the base image).\n2. A Reference Image (providing the target visual style: font family, weight, color, gradient, texture, lighting, shadows, effects).\n3. An Edited Result.\n4. An Instruction specifying the operation type (Add, Delete, Replace, or Modify) and the target text content.\n\nYour task is to evaluate if the specific operation was executed correctly, if the resulting text (for Add/Replace/Modify) perfectly inherits the visual attributes from the Reference Image, and if the overall edit is seamless.\n\nEvaluate the result on a 5-point scale from three perspectives:\n\n1. Text Instruction Compliance (Content & Style Fidelity)\n   - Assess: (A) Was the correct operation performed (Add/Delete/Replace/Modify)? (B) Is the text content accurate? (C) For styling tasks: Does the new text exactly match the Reference Image's style (font, color, effects)? For Delete: Is the text completely gone?\n   1  Failed: Wrong operation, wrong content, OR style completely ignores the reference (looks like source original).\n   2  Partial: Correct operation/content but wrong style category (e.g., wrong font/color); or for deletion, significant remnants remain.\n   3  Mostly Correct: Content correct; general style matches reference, but specific details (hue, gradient, texture) deviate noticeably. Deletion has faint traces.\n   4  Accurate: Operation/content perfect; style matches reference well with only minor discrepancies. Deletion is clean with tiny artifacts.\n   5  Perfect: Operation/content exact. Style (font, color, gradient, texture) exactly replicates the Reference. Deletion is total with no traces.\n\n2. Text Visual Quality (Clarity & Rendering)\n   - For Add/Replace/Modify: Evaluate sharpness, clarity, and rendering of the new text. No artifacts, blur, or broken glyphs.\n   - For Delete: Evaluate the quality of the background reconstruction where text was removed. No holes, smears, or mismatched textures.\n   1  Text unreadable; heavy artifacts, severe blur, broken glyphs (or large holes/smearing for deletion).\n   2  Readable only with effort: strong aliasing, warped baselines, obvious patch marks (or clearly patched background).\n   3  Readable with minor visual issues: slight blur, inconsistent stroke weight or mild distortion (or slight texture mismatch in filled area).\n   4  Clean and readable; small artifacts visible only when zoomed (or seamless repair with minute flaws).\n   5  Print-like clarity: sharp glyphs, consistent kerning/baseline, natural integration with substrate (or invisible repair).\n\n3. Overall Edit Quality (Semantics, Aesthetics & Integration)\n   - Evaluate the holistic success: Does the edit make semantic sense? Is the aesthetic unified? Does the text fit the scene physics (lighting, perspective, occlusion) while maintaining the reference style?\n   1  Overall failure due to semantic or severe visual issues.\n   2  Partially successful but clearly below usable quality.\n   3  Acceptable for casual use with noticeable flaws.\n   4  High quality and reliable with only tiny imperfections.\n   5  Excellent: semantics and aesthetics both near-perfect.\n\nCRITICAL CONSTRAINT: The scores for \"Text Visual Quality\" and \"Overall Edit Quality\" MUST NOT be higher than the score for \"Text Instruction Compliance\".\n\nExample Response Format:\nBrief reasoning: A short explanation based on the criteria above, no more than 20 words.\nText Instruction Compliance: [1-5]\nText Visual Quality: [1-5]\nOverall Edit Quality: [1-5]\n\nEditing instruction: <edit_prompt>\nBelow are the source, reference, and edited images:\n",
+    "Ref-Based Replace": "\nYou are a data rater specializing in grading reference-based object replacement. You will be given a source image, a reference image (containing the replacement object), and the edited result. Evaluate on a 5-point scale:\n\nPrompt Compliance (Replacement Accuracy)\n1  Target not replaced, or unrelated object edited; does not match reference object.\n2  Only part of target replaced, or wrong class/description used; reference object not recognized.\n3  Target largely replaced but other objects altered, remnants visible, or scale/pose wrong vs reference.\n4  Correct object fully replaced; only minor attribute errors (color, size) compared to reference.\n5  Perfect replacement: New object's class, number, position, scale, pose, and detail exactly match reference.\n\nVisual Naturalness\n1  Image heavily broken or new object deformed/extremely blurred; obvious seams.\n2  Obvious seams, smears, or strong mismatch in resolution/color; background not restored.\n3  Basic style similar, but lighting/palette clashes; fuzzy edges or noise noticeable.\n4  Style almost uniform; tiny edge artifacts visible only on close inspection.\n5  Completely seamless; new object blends fully with scene; edit area undetectable.\n\nPhysical & Detail Integrity\n1  Floating, interpenetration, severe perspective/light errors; background heavily warped.\n2  Missing shadows/occlusion; large background shifts or holes.\n3  Lighting, perspective, and contact surfaces mostly correct; small tolerable errors.\n4  New objects interact realistically (shadows, reflections); preserve existing details; minimal background change.\n5  Physically flawless: Accurate highlights, shadows, reflections matching reference object properties; background untouched.\n\nCRITICAL CONSTRAINT: Scores for \"Visual Naturalness\" and \"Physical & Detail Integrity\" must NOT exceed \"Prompt Compliance\".\n\nExample Response Format:\nBrief reasoning: Max 20 words.\nPrompt Compliance: [1-5]\nVisual Naturalness: [1-5]\nPhysical & Detail Integrity: [1-5]\n\nEditing instruction: <edit_prompt>\nBelow are the source, reference, and edited images:\n",
+    "Ref-Based Viewpoint Change": "\nYou are a data rater specializing in grading reference-based viewpoint changes. You will be given a source image, a reference image (defining the target camera angle/pose), and the edited result. Evaluate on a 5-point scale:\n\nInstruction Compliance (Viewpoint Accuracy)\n1  No viewpoint change, or wrong transformation direction; object identity/shape severely altered.\n2  Partial viewpoint change but clearly incorrect angle/axis; major unintended edits.\n3  Viewpoint broadly matches reference but with noticeable errors in rotation amount or occluded parts.\n4  Viewpoint matches well; object identity preserved; only small angular/pose deviations from reference.\n5  Exact viewpoint as requested: Camera/object transformation precisely matches the reference image.\n\nVisual Seamlessness\n1  Heavy artifacts, broken geometry, severe blur/noise.\n2  Obvious warping, texture stretching, or cut-out halos.\n3  Acceptable but visible inconsistencies in texture sharpness or edge quality.\n4  Clean render; minor artifacts only on close inspection.\n5  Highly natural: Textures, edges, and resolution remain consistent and credible.\n\nPhysical Consistency (Geometry, Occlusion, Lighting)\n1  Severe 3D inconsistency: Impossible geometry, wrong occlusions, conflicting light.\n2  Noticeable depth/occlusion errors; shadow/highlight logic broken.\n3  Mostly believable with minor occlusion or lighting mistakes.\n4  Geometry and occlusions largely correct; lighting consistent with original scene logic.\n5  Excellent 3D realism: Correct self-occlusion, plausible newly revealed surfaces, consistent lighting.\n\nCRITICAL CONSTRAINT: Scores for \"Visual Seamlessness\" and \"Physical Consistency\" must NOT exceed \"Instruction Compliance\".\n\nExample Response Format:\nBrief reasoning: Max 20 words.\nInstruction Compliance: [1-5]\nVisual Seamlessness: [1-5]\nPhysical Consistency: [1-5]\n\nEditing instruction: <edit_prompt>\nBelow are the source, reference, and edited images:\n",
+    "Ref-Based Style Transfer": "\nYou are a data rater specializing in grading reference-based style transfer. You will be given a source image, a reference style image, and the styled result. Evaluate on a 5-point scale:\n\nStyle Fidelity\n1  Target style absent or clearly wrong; does not resemble reference.\n2  Style shows in a few areas only, or mixed with unrelated styles.\n3  Key traits (palette, brushwork, texture) present but patchy or inconsistent with reference.\n4  Style reproduced across almost the whole image; only small local mismatches vs reference.\n5  Full, faithful transfer: Color, texture, brushwork, lighting all match the reference exemplar perfectly.\n\nContent Preservation\n1  Major objects or layout lost/distorted; original scene barely recognizable.\n2  Main subject recognizable, but size, perspective, or key parts clearly wrong/missing.\n3  Overall structure correct; some local warping or minor omissions.\n4  Nearly all geometry intact; only slight, non-distracting deformation.\n5  All objects and spatial relations kept; only stylistic, harmless distortion.\n\nRendering Quality\n1  Heavy noise, banding, pixel damage, or blur; image unusable.\n2  Visible seams, aliasing, color drift; low resolution or chaotic strokes.\n3  Moderate quality: Local blur/noise/texture breaks, but generally acceptable.\n4  Sharp, coherent strokes; tiny artifacts visible only when zoomed.\n5  High resolution, no artifacts; strokes, textures, and color transitions look fully natural.\n\nCRITICAL CONSTRAINT: Scores for \"Content Preservation\" and \"Rendering Quality\" must NOT exceed \"Style Fidelity\".\n\nExample Response Format:\nBrief reasoning: Max 20 words.\nStyle Fidelity: [1-5]\nContent Preservation: [1-5]\nRendering Quality: [1-5]\n\nEditing instruction: <edit_prompt>\nBelow are the source, reference, and styled images:\n",
+    "Multi-Subject-Driven Generation": "\nYou are a senior data rater specializing in grading multi-subject driven edits. You will be given:\n1. Multiple Reference Images (each containing a distinct subject/person/object to be included).\n2. An Editing Instruction (specifying how these subjects should be combined, posed, or interacting).\n3. A Generated/Edited Result Image.\n\nYour task is to evaluate if the model successfully preserved the unique identity of EACH reference subject while integrating them into the final image according to the instruction.\n\nEvaluate the result on a 5-point scale from three perspectives:\n\n1. Subject Consistency (Identity Preservation)\n   - Assess: Does EACH subject retain their unique identity features (face, hair, clothing, distinct markings) from their respective reference images? Check for identity bleeding (features mixing between subjects) or drift.\n   1  Severe identity loss: Subjects unrecognizable; faces swapped; features completely different from references; heavy identity bleeding between subjects.\n   2  Strong drift: Vague resemblance only; key facial features, clothing patterns, or distinct markers of one or more subjects are lost or altered.\n   3  Generally consistent: Main traits present for all subjects, but noticeable errors in details (e.g., wrong eye color, missing accessory, slight face morphing).\n   4  High fidelity: Clearly the same subjects; only minor deviations in fine details or texture consistency compared to references.\n   5  Perfect preservation: Every subject flawlessly retains their unique identity, features, and attire from their specific reference image; zero identity confusion.\n\n2. Instruction Compliance & Composition Logic\n   - Assess: (A) Are ALL requested subjects from the reference images present? (B) Is the specified interaction/pose/arrangement correct? (C) Are there any missing subjects or extra hallucinated subjects?\n   1  Complete failure: Missing major subjects, wrong subjects included, or completely ignored the compositional instruction.\n   2  Partial success: Most subjects present but key interaction/pose instructions failed; or one subject is clearly missing/wrong.\n   3  Mostly correct: All subjects present and broadly following the instruction, but specific arrangement details or interaction logic are slightly off.\n   4  High compliance: All subjects present; interaction and composition match the instruction well with only minor deviations.\n   5  Perfect execution: Every requested subject is included exactly as specified; poses, interactions, and spatial arrangement perfectly match the instruction.\n\n3. Image Realism & Visual Coherence\n   - Assess: Lighting consistency across all subjects, shadow logic, perspective alignment, anatomical correctness (no broken limbs), and seamless blending into the shared environment.\n   1  Very poor quality: Extreme noise, blur, heavy artifacts, broken anatomy (extra/missing limbs), or incoherent rendering making the image unusable.\n   2  Clearly synthetic: Harsh edges, unnatural textures, inconsistent lighting directions between subjects, or obvious \"cut-and-paste\" look.\n   3  Acceptable quality: Generally plausible but with visible issues (waxy skin, slight halos, mild shadow mismatches, or awkward joint angles).\n   4  High quality: Natural textures, believable lighting unified across all subjects, good background integration; flaws only visible on close inspection.\n   5  Excellent quality: Photorealistic and seamless; perfect lighting/shadow harmony among all subjects; anatomy is flawless; looks like a single native photograph.\n\nCRITICAL CONSTRAINT: The scores for \"Instruction Compliance\" and \"Image Realism\" MUST NOT be higher than the score for \"Subject Consistency\". If the identity of a subject is lost, the overall quality is capped regardless of composition or rendering quality.\n\nExample Response Format:\nBrief reasoning: A concise explanation based on the criteria above, max 25 words.\nSubject Consistency: [1-5]\nInstruction Compliance: [1-5]\nImage Realism & Visual Coherence: [1-5]\n\nEditing instruction: <edit_prompt>\nBelow are the multiple reference images and the generated result:\n",
+    "Multi-Image Compose Edit": "\nYou are a senior data rater specializing in grading complex, multi-image compose edits. You will be provided with:\n1. Multiple Input Images (Source images, Reference images for style/motion/attributes/text, Masks, or Structural guides).\n2. A Complex Editing Instruction (which may involve any combination of tasks: Add, Remove, Replace, Background Change, Style Transfer, Motion/Pose Transfer, Viewpoint Change, Re-orientation, Material/Color/Environment Adjustment, Text Editing, Portrait Retouching, CV Feature Extraction/Generation, or Multi-Subject Driven Generation).\n3. A Generated/Edited Result Image.\n\nYour task is to evaluate if the model successfully executed ALL requested operations using the provided references while maintaining high visual and physical quality.\n\nEvaluate the result on a 5-point scale from three perspectives:\n\n1. Instruction Compliance & Content Fidelity\n   - Assess: (A) Were ALL requested operations (add/remove/change/style/motion/etc.) performed correctly? (B) Is the content accurate (correct objects, text semantics, pose angles)? (C) For reference-based tasks: Did the output accurately inherit specific attributes (font, color, texture, motion, style, identity) from the corresponding reference images?\n   1  Complete failure: Major operations missing, wrong objects edited, content semantically wrong, or reference attributes completely ignored.\n   2  Partial success: Most operations attempted but key elements wrong (e.g., wrong style, wrong pose, wrong text content); significant deviation from reference attributes.\n   3  Mostly correct: All main operations present; general attributes match references, but noticeable errors in specific details (hue, exact pose angle, font weight).\n   4  High compliance: All operations executed accurately; reference attributes matched well with only minor deviations.\n   5  Perfect execution: Every operation exactly as requested; all reference attributes (style, motion, identity, text format) perfectly replicated; no unintended changes.\n\n2. Visual Naturalness & Seamless Integration\n   - Assess: The blending of edited regions, texture consistency, resolution matching, and absence of artifacts (halos, blurring, noise, banding). Does the result look like a single coherent image rather than a collage?\n   1  Severe artifacts: Heavy distortion, obvious cut-outs, strong color mismatches, broken textures, or unreadable text making the image unusable.\n   2  Obvious flaws: Clear edges/halos, inconsistent grain/resolution between sources, unnatural smoothing, or visible \"patched\" areas.\n   3  Acceptable: Generally coherent but visible inconsistencies upon closer look (slight edge blur, minor tone mismatch, subtle texture breaks).\n   4  High quality: Smooth transitions; textures and lighting feel unified; artifacts only visible when zoomed in significantly.\n   5  Flawless integration: Indistinguishable composite; perfect texture synthesis, color grading, and sharpness across all elements; looks like a native capture.\n\n3. Physical Consistency & Detail Integrity\n   - Assess: Lighting direction/shadow logic, perspective alignment, geometric correctness (no warped limbs/objects), material response (reflections/refractions), and logical interaction (occlusion, contact shadows). For CV tasks: Accuracy of extracted/generated structures.\n   1  Physically impossible: Floating objects, wrong shadow directions, broken geometry, severe perspective errors, or illogical material interactions.\n   2  Noticeable errors: Mismatched lighting intensity, missing contact shadows, slight perspective drift, or unnatural material responses.\n   3  Generally plausible: Minor flaws in shadow length, reflection accuracy, or local geometry; acceptable for casual viewing.\n   4  Strong consistency: Lighting, shadows, and perspective align logically across all combined elements; detailed interactions (wetness, fabric fold) are believable.\n   5  Physically perfect: Flawless light transport, accurate occlusion, perfect geometric integrity, and hyper-realistic material response; enhances overall scene realism.\n\nCRITICAL CONSTRAINT: The scores for \"Visual Naturalness\" and \"Physical Consistency\" MUST NOT be higher than the score for \"Instruction Compliance\". If the core operation or reference attribute transfer fails, the quality scores are capped.\n\nExample Response Format:\nBrief reasoning: A concise explanation based on the criteria above, max 25 words.\nInstruction Compliance & Content Fidelity: [1-5]\nVisual Naturalness & Seamless Integration: [1-5]\nPhysical Consistency & Detail Integrity: [1-5]\n\nEditing instruction: <edit_prompt>\nBelow are the input images and the generated result:\n",
+    "Style Transfer": "\nYou are a data rater specializing in grading style transfer edits. You will be given an input image, a reference style, and the styled result. Your task is to evaluate the style transfer on a 5-point scale from three perspectives:\n\nStyle Fidelity\n1  Target style absent or clearly wrong.\n2  Style shows in a few areas only, or mixed with unrelated styles.\n3  Key traits (palette, brushwork, texture) present but patchy or inconsistent.\n4  Style reproduced across almost the whole image; only small local mismatches.\n5  Full, faithful transfer: colour, texture, brushwork, lighting all match the exemplar over the entire image.\n\nContent Preservation\n1  Major objects or layout lost/distorted; original scene barely recognisable.\n2  Main subject recognisable, but size, perspective or key parts clearly wrong/missing.\n3  Overall structure correct; some local warping or minor omissions.\n4  Nearly all geometry intact; only slight, non-distracting deformation.\n5  All objects and spatial relations kept; only stylistic, harmless distortion.\n\nRendering Quality\n1  Heavy noise, banding, pixel damage or blur; image unusable.\n2  Visible seams, aliasing, colour drift; low resolution or chaotic strokes.\n3  Moderate quality: local blur/noise/texture breaks, but generally acceptable.\n4  Sharp, coherent strokes; tiny artefacts visible only when zoomed.\n5  High resolution, no artefacts; strokes, textures and colour transitions look fully natural.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nStyle Fidelity: A number from 1 to 5.\nContent Preservation: A number from 1 to 5.\nRendering Quality: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the input, reference style, and styled output image:\n",
+    "Subject Addition": "\nYou are a data rater specializing in grading image addition edits. You will be given two images (before and after editing) and the corresponding editing instructions. Your task is to evaluate the added object(s) on a 5-point scale from three perspectives:\n\nPrompt Compliance\n1  Nothing added or the added content is corrupt.\n2  Added object is a wrong class or unrelated to the prompt.\n3  Correct class, but key attributes (position, colour, size, count, etc.) are wrong.\n4  Main attributes correct; only minor details off or 1-2 small features missing.\n5  Every stated attribute correct and scene logic reasonable; only microscopic flaws.\n\nVisual Naturalness\n1  Image badly broken or full of artefacts.\n2  Obvious paste marks; style, resolution, or palette strongly mismatch.\n3  General style similar, but lighting or colours clearly clash; noticeable disharmony.\n4  Style almost uniform; small edge issues visible only when zoomed.\n5  Perfect blend; no visible difference between added object and original image.\n\nPhysical & Detail Coherence\n1  Severe physical errors (floating, wrong perspective/light); key original elements blocked; background heavily distorted.\n2  Contact or occlusion handled poorly; minor background shifts, jaggies or noise; background visibly changed.\n3  Lighting, perspective, and contact mostly correct; remaining flaws small and acceptable; limited background change.\n4  Shadows, reflections, and material response believable; no loss of original detail; background changes are minute.\n5  Added object enhances overall realism: precise highlights, shadows, ambient effects; background essentially untouched.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nPrompt Compliance: A number from 1 to 5.\nVisual Naturalness: A number from 1 to 5.\nPhysical & Detail Coherence: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Subject Removal": "\nYou are a data rater specializing in grading object removal edits. You will be given two images (before and after editing) and the corresponding editing instructions. Your task is to evaluate the removal quality on a 5-point scale from three perspectives:\n\nPrompt Compliance\n1  Nothing removed, or an unrelated object edited.\n2  Target only partly removed, or a different instance/class deleted, or another object appears in the gap.\n3  Target mostly removed but extra objects also deleted, or fragments of the target remain.\n4  Only the specified objects removed, but a few tiny/background items deleted by mistake, or the count is wrong.\n5  Perfect: all and only the requested objects removed; every other element untouched.\n\nVisual Naturalness\n1  Image badly broken (large holes, strong artefacts).\n2  Clear erase marks; colour/resolution mismatch; background not restored.\n3  General look acceptable yet lighting/colour/style still clash; blur or noise visible.\n4  Style consistent; minor edge issues visible only when zoomed.\n5  Seamless: removal is virtually impossible to spot.\n\nPhysical & Detail Integrity\n1  Severe physical errors (floating items, wrong perspective/light); key scene elements damaged; background heavily warped.\n2  Large un-filled gaps or obvious background shifts.\n3  Lighting, perspective and contacts mostly correct; flaws small and tolerable; background adjusted locally.\n4  Background reconstruction clean; existing details preserved; only minute changes outside the removal area.\n5  Physically flawless and even enhances realism: accurate light/shadow/texture infill, high-quality micro-details.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nPrompt Compliance: A number from 1 to 5.\nVisual Naturalness: A number from 1 to 5.\nPhysical & Detail Integrity: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Subject Replacement": "\nYou are a data rater specializing in grading image replacement edits. You will be given two images (before and after editing) and the corresponding editing instructions. Your task is to evaluate the replacement editing effect on a 5-point scale from three perspectives:\n\nPrompt Compliance\n1  Target not replaced, or an unrelated object edited.\n2  Only part of the target replaced, or wrong class/description used.\n3  Target largely replaced but other objects altered, remnants visible, or count/position clearly wrong.\n4  Correct object fully replaced; only minor attribute errors (colour, size, etc.).\n5  Perfect replacement: all and only the specified objects removed; new objects’ class, number, position, scale, pose and detail exactly match the prompt.\n\nVisual Naturalness\n1  Image heavily broken or new object deformed / extremely blurred.\n2  Obvious seams, smears, or strong mismatch in resolution or colour; background not restored.\n3  Basic style similar, but lighting or palette clashes; fuzzy edges or noise are noticeable.\n4  Style almost uniform; tiny edge artefacts visible only on close inspection; casual viewers see no edit.\n5  Completely seamless; new objects blend fully with the scene, edit area undetectable.\n\nPhysical & Detail Integrity\n1  Floating, interpenetration, severe perspective/light errors; key original elements ruined; background heavily warped.\n2  Missing shadows/occlusion; large background shifts or holes.\n3  Lighting, perspective and contact surfaces mostly correct; small but tolerable errors; background adjusted locally.\n4  New objects interact realistically with scene (shadows, reflections, texture) and preserve existing details; background change minimal.\n5  Physically flawless and enhances realism: accurate highlights, shadows, reflections, ambient effects; background untouched.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nPrompt Compliance: A number from 1 to 5.\nVisual Naturalness: A number from 1 to 5.\nPhysical & Detail Integrity: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Background Change": "\nYou are a data rater specializing in grading background editing. You will be given two images (before and after editing) and the editing instruction. Your task is to evaluate the background change on a 5-point scale from three perspectives:\n\nInstruction Compliance\n1  No change, or background unrelated to prompt, or foreground also replaced/distorted.\n2  Background partly replaced or wrong style/content; foreground noticeably altered.\n3  Main background replaced but elements missing/extra, or faint spill onto subject edges.\n4  Requested background fully present; foreground intact except minute artefacts or small prompt mismatch (e.g. colour tone).\n5  Background exactly matches prompt (content, style, placement); all foreground pixels untouched.\n\nVisual Seamlessness (Edge & Texture Blend)\n1  Large tearing, posterisation, extreme blur/noise; edit area obvious at a glance.\n2  Clear cut-out halos, colour-resolution gap, or heavy smudge strokes.\n3  Blend acceptable but visible on closer look: slight edge blur, grain or palette shift.\n4  Nearly invisible seams; textures and sharpness aligned, only minor issues when zoomed in.\n5  Indistinguishable composite: edges, textures, resolution and colour grading perfectly continuous.\n\nPhysical Consistency (Lighting, Perspective, Depth)\n1  Severe mismatch: wrong horizon, conflicting light direction, floating subject, warped geometry.\n2  Noticeable but not extreme inconsistencies in light, shadows or scale; depth cues off.\n3  Overall believable; small errors in shadow length, perspective or ambient colour.\n4  Lighting, scale, depth, and camera angle well matched; only subtle discrepancies.\n5  Physically flawless: foreground and new background share coherent light, shadows, reflections, perspective and atmospheric depth, enhancing overall realism.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nInstruction Compliance: A number from 1 to 5.\nVisual Seamlessness: A number from 1 to 5.\nPhysical Consistency: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Color Alteration": "\nYou are a data rater specializing in grading color alteration edits. You will be given two images (before and after editing) and the editing instruction. Your task is to evaluate the color change on a 5-point scale from three perspectives:\n\nInstruction Compliance\n1  Target color not changed, wrong object edited, or geometry materially altered.\n2  Correct object but wrong target color/direction; only part of target recolored; other objects also affected.\n3  Main recolor correct but noticeable deviation in hue/saturation/value; small collateral recolor or edge spill.\n4  Requested objects recolored accurately; only tiny hue/value mismatch or very small spill.\n5  Exactly and only the specified regions recolored; hue/saturation/value precisely match the prompt; shapes fully intact.\n\nVisual Seamlessness\n1  Heavy banding, posterisation, colour noise or large blotches.\n2  Clear edge bleeding/halos; abrupt tonal jumps; inconsistent grain.\n3  Overall OK but local texture/lighting mismatch; recolor boundary slightly visible.\n4  Smooth transitions; texture and grain preserved; artifacts only when zoomed.\n5  No detectable recolor traces; original material texture and lighting fully preserved.\n\nPhysical Consistency\n1  Severe violation of material/light logic (e.g., metal looks like plastic); highlights/shadows broken.\n2  Noticeable mismatch in specular response or shadow tint; background/adjacent regions visibly distorted.\n3  Material and lighting mostly plausible with minor errors.\n4  Color change respects material properties and scene illumination; details well preserved.\n5  Physically excellent: recolor looks like a true material/paint change under the same lighting.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nInstruction Compliance: A number from 1 to 5.\nVisual Seamlessness: A number from 1 to 5.\nPhysical Consistency: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Attribute Change": "\nYou are a data rater specializing in grading attribute alteration edits. You will be given two images (before and after editing) and the corresponding editing instructions. Your task is to evaluate the attribute change on a 5-point scale from three perspectives:\n\nPrompt Compliance\n1  Target not adjusted, wrong object touched, or geometry changed.\n2  Right object but wrong attribute value/direction; only part edited; other objects also altered; slight stretch/crop.\n3  Mainly correct object and attribute, yet large hue/brightness/texture error; minor collateral edits; visible jaggies/distortion.\n4  All requested objects adjusted, only their attributes changed; shape kept; small inaccuracy in colour, material or amount.\n5  Exactly and only the requested objects adjusted; colour, material, gloss etc. match the prompt perfectly; shape 100% intact; zero unintended edits.\n\nVisual Seamlessness\n1  Massive colour spill, mosaics or heavy noise; image nearly unusable.\n2  Clear smears/bleeding on edges; abrupt resolution or tone shift; highlights/shadows clipped; background gaps.\n3  Overall palette OK but local tone or grain conflicts; soft edges; noticeable disharmony.\n4  Style unified, transitions smooth; only slight edge artefacts visible when zoomed.\n5  No detectable edit traces; colours/materials fuse with scene lighting; edit area practically invisible.\n\nPhysical & Detail Fidelity\n1  Object floating, interpenetrating, or severe perspective/light mismatch; background badly warped.\n2  Missing shadows/highlights; wrong reflection direction; background visibly discoloured or distorted.\n3  Light, perspective and contact surface largely correct; minor acceptable flaws; background only locally affected.\n4  Adjusted material interacts believably with scene; shadows, highlights, reflections handled well; original details preserved.\n5  High physical realism: fine micro-highlights, diffuse bounce, subsurface effects present; overall scene realism improved.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nPrompt Compliance: A number from 1 to 5.\nVisual Seamlessness: A number from 1 to 5.\nPhysical & Detail Fidelity: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Viewpoint Change": "\nYou are a data rater specializing in grading viewpoint/camera change edits (3D-aware edits). You will be given two images (before and after editing) and the editing instruction. Your task is to evaluate the viewpoint transformation on a 5-point scale from three perspectives:\n\nInstruction Compliance\n1  No viewpoint change, or wrong transformation direction; object identity/shape severely altered.\n2  Partial viewpoint change but clearly incorrect angle/axis; major unintended edits to object or background.\n3  Viewpoint broadly matches the instruction but with noticeable errors in rotation amount, occluded parts, or object count/pose.\n4  Viewpoint matches well; object identity preserved; only small angular/pose deviations.\n5  Exact viewpoint as requested: camera/object transformation precisely matches the prompt with no unrelated edits.\n\nVisual Seamlessness\n1  Heavy artifacts, broken geometry, severe blur/noise.\n2  Obvious warping, texture stretching, or cut-out halos.\n3  Acceptable but visible inconsistencies in texture sharpness or edge quality.\n4  Clean render; minor artifacts only on close inspection.\n5  Highly natural: textures, edges, and resolution remain consistent and credible.\n\nPhysical Consistency (Geometry, Occlusion, Lighting)\n1  Severe 3D inconsistency: impossible geometry, wrong occlusions, conflicting light.\n2  Noticeable depth/occlusion errors; shadow/highlight logic broken.\n3  Mostly believable with minor occlusion or lighting mistakes.\n4  Geometry and occlusions largely correct; lighting consistent with original scene.\n5  Excellent 3D realism: correct self-occlusion, plausible newly revealed surfaces, consistent lighting and material response.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nInstruction Compliance: A number from 1 to 5.\nVisual Seamlessness: A number from 1 to 5.\nPhysical Consistency: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Motion Change": "\nYou are a data rater specializing in grading action/motion change edits. You will be given two images (before and after editing) and the editing instruction. Your task is to evaluate the motion or action change on a 5-point scale from three perspectives:\n\nMotion / Action Fidelity\n1  No visible change, or wrong action / motion.\n2  Partial or clearly incorrect pose; only some body parts change; action direction wrong.\n3  Main idea present but details off (angle, side, intensity, missing gesture).\n4  Requested action achieved with just minor inaccuracy (small angular drift, timing nuance).\n5  Exact match to prompt: every limb, gesture, body posture and motion cue aligns with the described action.\n\nIdentity Preservation\n1  Subject unrecognisable; face/body replaced.\n2  Strong drift: key facial features, hairstyle, clothing or body shape heavily altered.\n3  Mostly same identity; moderate changes in some features but still recognisable.\n4  Identity clearly the same; only subtle stylisation or lighting differences.\n5  Perfect preservation of face, hairstyle, skin tone, clothing and accessories.\n\nVisual & Anatomical Coherence\n1  Severe artifacts: broken or duplicated limbs, extreme distortion, heavy noise/blur.\n2  Noticeable cut-out halos, proportion errors, lighting or perspective clearly off.\n3  Generally plausible; minor joint or shading issues; small noise/blur acceptable.\n4  Clean render; anatomy, lighting, depth and edges consistent; flaws only on close inspection.\n5  Flawless realism or stylistic coherence; perfect anatomy, lighting, shadows and texture continuity.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nMotion Fidelity: A number from 1 to 5.\nIdentity Preservation: A number from 1 to 5.\nVisual & Anatomical Coherence: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Subject Extraction": "\nYou are a data rater specializing in grading object cut-out quality. You will be given an image with the object extracted on a white background. Your task is to evaluate the cut-out accuracy on a 5-point scale from three perspectives:\n\nObject Selection & Identity\n1  Wrong object or multiple objects extracted.\n2  Correct class but only part of the object, or obvious intrusions from other items.\n3  Object largely correct yet small pieces missing / extra, identity still recognisable.\n4  Full object with clear identity; only tiny mis-crop (e.g., tip of antenna).\n5  Exact requested object, complete and unmistakably the same instance (ID).\n\nMask Precision & Background Purity\n1  Large background remnants, holes in mask, or non-white backdrop dominates.\n2  Noticeable jagged edges, colour fringes, grey/colour patches in white area.\n3  Acceptable mask; minor edge softness or faint halo visible on close look.\n4  Clean, smooth edges; white (#FFFFFF) background uniform, tiny artefacts only when zoomed.\n5  Crisp anti-aliased contour, zero spill or halo; backdrop perfectly pure white throughout.\n\nObject Integrity & Visual Quality\n1  Severe blur, compression, deformation, or missing parts; unusable.\n2  Moderate noise, colour shift, or slight warping; details clearly degraded.\n3  Overall intact with minor softness or noise; colours mostly preserved.\n4  Sharp detail, accurate colours; negligible artefacts.\n5  Pristine: high-resolution detail, true colours, no artefacts or distortion.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nObject Identity: A number from 1 to 5.\nMask Precision: A number from 1 to 5.\nVisual Quality: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow is the extracted object image:\n",
+    "Text Modification": "\nYou are a data rater specializing in grading text editing in images. You will be given two images (before and after editing) and the editing instruction. Your task is to evaluate the text edit on a 5-point scale from three perspectives:\n\nText Instruction Compliance\n1  Text intent not satisfied at all (wrong operation: add/remove/replace), or wrong content.\n2  Correct operation but wrong key words/numbers/language, or edits applied to wrong region.\n3  Main text intent correct but notable semantic errors remain (typos, missing words, wrong count/position).\n4  Semantics mostly correct; only minor character/spacing/punctuation mistakes.\n5  Exact semantics: all requested text content, language, placement and operation perfectly match the instruction.\n\nText Visual Quality\n1  Text unreadable; heavy artifacts, severe blur, broken glyphs.\n2  Readable only with effort: strong aliasing, warped baselines, obvious patch marks.\n3  Readable with minor visual issues: slight blur, inconsistent stroke weight or mild distortion.\n4  Clean and readable; small artifacts visible only when zoomed.\n5  Print-like clarity: sharp glyphs, consistent kerning/baseline, natural integration with substrate.\n\nOverall Edit Quality\n1  Overall failure due to semantic or severe visual issues.\n2  Partially successful but clearly below usable quality.\n3  Acceptable for casual use with noticeable flaws.\n4  High quality and reliable with only tiny imperfections.\n5  Excellent: semantics and aesthetics both near-perfect.\nText Visual Quality and Overall Edit Quality should be no higher than Text Instruction Compliance!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nText Instruction Compliance: A number from 1 to 5.\nText Visual Quality: A number from 1 to 5.\nOverall Edit Quality: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n",
+    "Compose Edit": "\nYou are a data rater specializing in grading hybrid image edits (involving multiple operations on multiple objects). You will be given two images (before and after editing) and the corresponding editing instructions. Your task is to evaluate the overall editing quality on a 5-point scale from three perspectives:\n\nInstruction Compliance\n1  Neither object nor operations match the prompt; wrong items edited or shapes distorted.\n2  Only one object correctly edited, or both edited but with wrong/partial operations; collateral changes to other items.\n3  Both target objects touched, each with the requested operation broadly correct but missing details (e.g., wrong colour value, incomplete removal).\n4  Both objects receive the exact operations; tiny deviations in amount, position, or parameter. No unintended edits elsewhere.\n5  Perfect execution: each object fully reflects its specified operation, all other scene elements untouched.\n\nVisual Naturalness (Seamlessness)\n1  Large artefacts, obvious cut-outs, heavy blur/noise; edits conspicuous at a glance.\n2  Clear edge halos, colour or resolution mismatch, awkward scaling.\n3  Acceptable but visible on close look: slight edge softness, minor palette or focus shift.\n4  Edits blend smoothly; seams hard to spot, textures and sharpness largely consistent.\n5  Indistinguishable composite: colour grading, grain, resolution and style fully match the original image.\n\nPhysical Consistency & Fine Detail\n1  Severe lighting/perspective mismatch, missing or wrong shadows; objects appear floating or warped.\n2  Noticeable but tolerable inconsistencies in illumination, scale, or depth cues.\n3  Generally plausible; small errors in shadow length, reflection angle, or texture alignment.\n4  Lighting, perspective, and material response closely match; only subtle flaws visible when zoomed.\n5  Physically flawless: shadows, highlights, reflections, depth and texture perfectly integrated, enhancing overall realism.\nThe second and third score should no higher than first score!!!\n\nExample Response Format:\nBrief reasoning: A short explanation of the score based on the criteria above, no more than 20 words.\nInstruction Compliance: A number from 1 to 5.\nVisual Naturalness: A number from 1 to 5.\nPhysical Consistency & Fine Detail: A number from 1 to 5.\nediting instruction is : <edit_prompt>.\n\nBelow are the images before and after editing:\n"
+}