Appendix A Prompts for Q/A generation Tables 5 – 8 provide the full prompt templates used for taxonomy-guided evidence discovery and evidence-conditioned question–answer generation in Section 3 . We use separate templates for high-level semantics and low-level visual attributes, and apply the same templates independently to ground truth and super-resolved images. Table 5: High-level taxonomy-guided evidence discovery prompt. Look at this HR image ({image_id}) and identify high-level semantic evidence targets for short-answer Q/A generation. Taxonomy definitions: - scene_semantic: The overall scene, environment, or image-level context. This taxonomy must describe the image as a whole or a broad scene region, not a single object. - object_semantic: A distinct object or structure that contributes meaningfully to the image content, such as a tree, grass, tower, wall, house, vehicle, bridge, or signboard; or a recognizable visual property or part within an object or local region, such as texture, style, material, decoration, or a visible object part. - text_logo: Clearly readable text, words, numbers, signs, brands, or logos. Generate up to 1 scene_semantic target, up to 4 object_semantic targets, and up to 4 text_logo targets. A target is suitable when it is clearly visible in the HR image, supports a specific short-answer question, and may be altered by downsampling or super-resolution. Generate as many valid targets as are clearly supported by the image, within these limits. Reply with valid JSON only: [ { "evidence_id": "he01", "taxonomy": "scene_semantic", "region_scope": "whole_scene or localized region", "region_hint": "brief visible location", "target": "specific high-level semantic content", "hallucination_risk": "how this content may be altered or invented" } ] Table 6: High-level evidence-conditioned question–answer generation prompt. Look at this HR image ({image_id}) and generate one high-level short-answer Q/A pair for each provided {taxonomy} evidence item. Taxonomy definition: {taxonomy_definition} The question must identify the relevant image region unambiguously and ask only about the specified high-level semantic content. Use a precise semantic label or short noun phrase. Do not ask about overall image quality. Do not create yes/no questions. Each answer must be specific and 1-3 words only. Evidence items: {evidence_json} Reply with valid JSON only: [ {"evidence_id": "{example_evidence_id}", "q": "...", "a": "..."} ] Table 7: Low-level taxonomy-guided evidence discovery prompt. Look at this HR image ({image_id}) and identify low-level visual evidence targets for short-answer Q/A generation. Taxonomy definitions: - color_tone: Visible color, brightness, contrast, highlight, shadow, or tone relationship within an image region. - edge_shape: Visible orientation, curvature, thickness, continuity, or shape of a boundary or small structural part. - texture: Visible surface grain, material appearance, fine pattern, line structure, or texture density. Generate 3 color_tone targets, 3 edge_shape targets, and 3 texture targets for this image. A target is suitable when it is clearly visible in the HR image, supports a specific short-answer question, and may be altered by downsampling or super-resolution. Reply with valid JSON only: [ { "evidence_id": "le01", "taxonomy": "color_tone", "region_scope": "local", "region_hint": "brief visible location", "target": "specific low-level visual property", "hallucination_risk": "how this property may be altered or invented" } ] Table 8: Low-level evidence-conditioned question–answer generation prompt. Look at this HR image ({image_id}) and generate one highly specific low-level short-answer Q/A pair for each provided evidence item. The question must identify the relevant image region and ask only about the specified low-level visual property. Describe the finest reliably visible appearance. Avoid broad categorical answers when the image supports a more precise description. Use the following attribute structure: - color_tone: describe the visible hue together with its lightness, saturation, or local tonal character. - edge_shape: describe the visible boundary using its thickness, continuity, smoothness, geometry, or orientation. - texture: describe the visible texture using its scale, direction, regularity, density, or structural appearance. Combine two or three complementary visual attributes when they are clearly supported by the image. Do not add an attribute merely to make the answer longer or more specific. Use visually grounded everyday terms. Do not use uncommon or technical terminology solely for greater specificity. Do not ask about overall image quality or sharpness preference. Do not create yes/no questions. Each answer must be specific and 2-3 words long. Evidence items: {evidence_json} Reply with valid JSON only: [ {"evidence_id": "le01", "q": "...", "a": "..."} ]
Appendix B VHA+ training details We train the last four transformer blocks of the DINOv3 ViT-B/16 encoder and the affine scale and offset parameters, while keeping the earlier layers frozen. We set λcorr=3 in Eq. 7 and train VHA+ using AdamW ( Loshchilov and Hutter, 2019 ) for 20,000 steps with a batch size of 32. The learning rates are 10−6 for the encoder and 10−4 for the affine parameters. Score standardization and calibration. Let μH and σH denote the mean and standard deviation of the training-set target scores. We standardize each target score hi as h~i=σHhi−μH. (8) We train the affine head of VHA+ to predict the standardized target h~i . Both loss terms in Eq. 7 are computed between the model outputs and the standardized targets h~i . For numerical stability, we add 10−8 to the denominator when computing the Pearson correlation. At inference, we convert the affine head output to the original score scale using the stored training-set mean and standard deviation. For our training dataset, μH=0.7545 and σH=1.0254 . Appendix D Prompts for direct VLM rating Table 9 provides the full prompt used for the Direct rating baseline in the comparison of VLM-based hallucination assessment methods in Section 5.2 (Table 1 ). Table 9: Full prompt for Direct rating. You will receive two images: 1. Ground Truth (GT): The reference high-resolution image. 2. Super-Resolved Image (SR): An image restored by an AI super-resolution model. We define hallucinations as the loss or alteration of ground-truth evidence in SR, or the addition of content unsupported by GT. Rate the amount of hallucination in SR compared with GT on a scale from 1 to 5. A lower score means less hallucination, and a higher score means more hallucination. Output only the following JSON: { "score": <integer from 1 to 5> }
Appendix E Additional ablations E.1 VHA scoring threshold ablations Table 10: Ablations of VHA scoring thresholds on VHA-bench. ρ and r denote Spearman and Pearson correlations. (a) Forward score threshold τg ρ r 0 0.5919 0.6295 −1 0.6190 0.6427 −2 0.6179 0.6449 −3 0.6103 0.6408 −4 0.6085 0.6396 −5 0.6075 0.6355 −6 0.6079 0.6354 (b) Reverse score threshold τs ρ r −3 0.6160 0.6426 −2 0.6164 0.6431 −1 0.6159 0.6437 0 0.6179 0.6449 1 0.6144 0.6430 2 0.6087 0.6383 3 0.5941 0.6295 (c) Reverse margin threshold τm ρ r −1 0.5909 0.6366 −0.5 0.6045 0.6407 0 0.6173 0.6443 0.5 0.6179 0.6449 1 0.6133 0.6423 Table 10 shows the effects of the GT answer-score threshold τg in Eq. 3 and the SR answer-score threshold τs and margin τm in Eq. 4 . VHA is relatively insensitive to variations in these hyperparameters. However, overly strict or permissive thresholds can reduce correlations with human judgments. These results suggest that appropriate thresholds help suppress noisy Q/A contributions while preserving informative ones. E.2 VHA+ training ablations Table 11: VHA+ training ablations on VHA-bench. ρ and r denote Spearman and Pearson correlations, respectively; bold indicates the selected settings. (a) Correlation loss weight λcorr ρ r 0 0.6370 0.6680 1 0.6540 0.6771 2 0.6586 0.6763 3 0.6594 0.6753 4 0.6593 0.6741 5 0.6582 0.6735 (b) Trainable blocks Last K blocks ρ r 2 0.5962 0.6212 4 0.6594 0.6753 6 0.6556 0.6638 8 0.6494 0.6660 10 0.6389 0.6631 Table 11 presents VHA+ training ablations. Table 11 ( 11(a) ) shows the results of varying the Pearson correlation loss weight λcorr in Eq. 7 . An appropriate correlation loss weight improves agreement with human judgments compared with MSE alone. Table 11 ( 11(b) ) shows the results for different numbers of trainable transformer blocks. Updating fewer or more than four blocks yields lower correlations, which may reflect underfitting and overfitting, respectively. Appendix F Fine-Tuning SR Models with VHA+ F.1 Additional Training Details VOSR ( Wu et al., 2026 ) first trains a multi-step SR model and then distills it into a one-step student. We fine-tune the pretrained one-step student with the additional VHA+ and MUSIQ losses while keeping the released multi-step model fixed as the teacher. We use the full LSDIR training set ( Li et al., 2023 ) and all 800 DIV2K training images ( Agustsson and Timofte, 2017 ) . Following VOSR’s training setup, we use random 512×512 crops and its degradation pipeline. The fine-tuning objective is L=Ldistill+γvhaLVHA++γmusiqLMUSIQ, (9) where Ldistill is VOSR’s original distillation loss. The auxiliary losses LVHA+ and LMUSIQ are the score predicted by the frozen VHA+ model and the negative MUSIQ score ( Ke et al., 2021 ) , respectively. We set γvha=0.1 and γmusiq=0.001 . We fine-tune the student for 1,000 updates using AdamW ( Loshchilov and Hutter, 2019 ) , a learning rate of 10−5 , and a batch size of 16. We use an exponential moving average (EMA) with decay 0.99 . F.2 SR fine-tuning ablations Table 12: Ablations of VHA+ and MUSIQ loss weights for VOSR fine-tuning. Pixel sim. Feature sim. NR-IQA GFR-IQA Ours Dataset γvha γmusiq PSNR ↑ SSIM ↑ LPIPS ↓ DISTS ↓ MUSIQ ↑ CLIPIQA ↑ VQ-R1 ↑ A-FINE ↓ VHA ↓ VHA+ ↓ RealSR 0 0.001 23.04 0.6418 0.3601 0.2643 75.89 0.6820 3.723 11.29 1.372 1.172 0.05 0.001 23.96 0.6791 0.2925 0.2089 72.17 0.6373 3.930 11.03 0.9164 0.6856 0.10 0.001 24.29 0.6914 0.2820 0.1985 70.98 0.6074 3.856 11.24 0.9494 0.6413 0.15 0.001 24.25 0.6914 0.2698 0.1907 70.25 0.5910 3.769 13.17 0.9557 0.6304 0.20 0.001 24.38 0.6988 0.2616 0.1892 69.39 0.5610 3.791 13.94 0.9355 0.6211 [2pt/1.5pt] 0.10 0 25.78 0.7431 0.2237 0.1737 51.31 0.3457 2.709 48.03 0.8744 0.5645 0.10 0.0005 24.53 0.7026 0.2618 0.1892 68.87 0.5710 3.761 16.08 0.9410 0.6341 0.10 0.0010 24.29 0.6914 0.2820 0.1985 70.98 0.6074 3.856 11.24 0.9494 0.6413 0.10 0.0015 23.98 0.6816 0.2899 0.2026 71.83 0.6255 3.943 10.10 1.0331 0.6536 0.10 0.0020 23.70 0.6653 0.3042 0.2087 73.02 0.6391 3.945 8.811 0.9854 0.6641 DRealSR 0 0.0010 24.51 0.6482 0.4357 0.3002 73.57 0.6899 3.184 12.27 1.271 1.256 0.05 0.0010 25.51 0.6954 0.3573 0.2349 68.52 0.6545 3.678 18.59 0.7262 0.6788 0.10 0.0010 25.83 0.7076 0.3473 0.2263 66.97 0.6422 3.534 21.63 0.7161 0.6473 0.15 0.0010 25.84 0.7060 0.3349 0.2175 66.01 0.6237 3.471 25.25 0.6745 0.6101 0.20 0.0010 25.95 0.7139 0.3266 0.2146 65.11 0.6042 3.459 27.07 0.6342 0.6043 [2pt/1.5pt] 0.10 0 27.24 0.7714 0.2539 0.1837 48.03 0.4260 2.489 52.57 0.6347 0.5812 0.10 0.0005 26.08 0.7173 0.3247 0.2140 64.51 0.6123 3.462 28.58 0.7118 0.6473 0.10 0.0010 25.83 0.7076 0.3473 0.2263 66.97 0.6422 3.534 21.63 0.7161 0.6473 0.10 0.0015 25.52 0.6947 0.3592 0.2310 68.30 0.6541 3.666 20.93 0.6843 0.6469 0.10 0.0020 25.30 0.6799 0.3733 0.2386 69.49 0.6648 3.673 19.24 0.6466 0.6527 Table 12 shows the results of varying the VHA+ and MUSIQ loss weights in Eq. 9 . Increasing the VHA+ loss weight tends to lower perceptual-quality scores, whereas increasing the MUSIQ loss weight can reduce image similarity and increase VHA+ scores. These results suggest that appropriate loss weights are needed to reduce hallucinations while maintaining perceptual quality.