NutriVisRAG: Multimodal Exemplar Grounding for Nutrient Estimation with Vision-Language Models

Lubnaa Abdur Rahman; Ioannis Papathanail; Maria Panagiotou; Nefeli Dimitropoulou; Margret Harðardóttir; Elisabeth J. den Brok; Bastiaan E. de Galan; Stavroula Georgia Mougiakakou

cs.CVNov 5, 2026

NutriVisRAG: Multimodal Exemplar Grounding for Nutrient Estimation with Vision-Language Models

Authors: Lubnaa Abdur Rahman, Ioannis Papathanail, Maria Panagiotou, Nefeli Dimitropoulou, Margret Harðardóttir, Elisabeth J. den Brok, Bastiaan E. de Galan, Stavroula Georgia Mougiakakou

Organizations: Graduate School for Cellular and Biomedical Sciences, University of Bern, 3012 Bern, Switzerland · ARTORG Center for Biomedical Engineering Research, University of Bern, 3008 Bern, Switzerland · Department of Endocrinology, Maastricht University Medical Centre+, and Cardiovascular Research Institute Maastricht (CARIM), Maastricht University, Maastricht, the Netherlands

Abstract

Diet is a major determinant of health, yet accurate dietary assessment from free-living meal images remains challenging. Existing retrieval-based Vision-Language Model (VLM) approaches to nutrient estimation typically rely on textual food representations, while, to our knowledge, direct grounding on visually similar annotated meals has not been investigated. We present NutriVisRAG, a retrieval-augmented generation framework that uses a vision encoder to retrieve similar meals with known nutritional annotations and provide them as multimodal in-context exemplars for VLM nutrient estimation at inference time, without task-specific fine-tuning. We evaluate NutriVisRAG across four datasets, four VLMs, three retrieval encoders, and two prompting strategies, including free-living meals from adults with insulin-treated diabetes, where carbohydrate estimates are compared against two dietitians' annotations. Across all 96 configurations (4 datasets x 4 VLMs x 3 retrieval encoders x 2 prompting strategies), NutriVisRAG reduced the primary estimation error in 56, with improvements more frequent under whole-meal (35/48) than per-item prompting (21/48). Across the evaluated configurations, the lowest NutriVisRAG error was below the best Vanilla error in 13 of 16 VLM--dataset pairs. For GPT-5.4-mini, under whole-meal prompting, the best-performing retrieval encoder for each dataset reduced the primary error on all four datasets relative to Vanilla, with reductions of 1.7--35.7%. Overall, our results show that grounding in multimodal exemplars comprising images and respective annotations can improve nutrient estimation at inference time without task-specific fine-tuning, although the benefits remain configuration-dependent.

Figures & tables

Figure 1: Overview of the NutriVisRAG pipeline: the top- $K$ visually similar reference meals are retrieved for a query image, and their images and nutritional annotations are added to the prompt as multimodal context for the VLM.

Dataset	Evaluation queries	Indexed references	Evaluation targets
Nutrition5k [ 51 ]	815	2,445	Weight, energy, CHO, protein, fat
SNAPMe [ 24 ]	323	1,149	Energy, CHO, protein, fat
MADiMa [ 7 ]	160 (80 meals each at 40 and 60 cm)	–	Weight, energy, CHO, protein, fat
EDUCATE [ 12 ]	247	–	CHO

Table 1: Datasets used.

Table 3

Figure 2: Bland–Altman analysis of CHO estimates for 247 EDUCATE meals using GPT-5.4-mini with SigLIP 2 under whole-meal prompting, without (blue) and with (orange) NutriVisRAG. Differences are computed against the mean dietitian estimate; solid and dashed lines show the mean bias and 95% limits of agreement, respectively.

		Per-item				Whole-meal
			NutriVisRAG				NutriVisRAG
Query set	Model	Vanilla	PE-Core	SigLIP2	Nomic Vision	Vanilla	PE-Core	SigLIP2	Nomic Vision
Nutrition5k	Gemma-4-31B	53.6	53.7	54.6	59.8	63.5	51.0 ⋆	50.8 ⋆	55.8 ⋆
	GPT-5.4-mini	58.0	42.7 ⋆	43.4 ⋆	47.1 ⋆	58.1	43.4 ⋆	44.2 ⋆	48.0 ⋆
	Mistral-Small-3.2	72.1	111.2 ▼	119.5 ▼	145.1 ▼	78.7	62.3 ⋆	63.3 ⋆	74.3 ⋆
	GLM-5.3-Flash	54.0	54.1 ▼	53.4 ⋆	58.6 ▼	56.2	58.8 ▼	56.8 ▼	62.2 ▼

Table 4: Vanilla versus NutriVisRAG for all retrieval encoders and both prompting strategies, on the full query sets. Values are

\mathrm{\%MAE}_{\text{micro}}

(

\mathrm{\%MAE}_{\text{CHO}}

for EDUCATE); lower is better. Bold: lowest error within each prompting strategy per model and query set. ⋆ / ▼ : significant improvement/degradation over the corresponding Vanilla baseline (paired Wilcoxon signed-rank test, Benjamini–Hochberg corrected,

q<0.05

Figure 3: Effect of prompting granularity under Vanilla VLM inference.

\mathrm{\%MAE}_{\text{macro}}

(lower is better) is reported across VLMs and query sets; (*) for EDUCATE,

\mathrm{\%MAE}_{\text{CHO}}

against the mean dietitian estimate is shown.

Figure 4: Association of exemplar similarity with NutriVisRAG retrieval benefit under per-item and whole-meal prompting. Spearman correlations (

\rho

) relate retrieval benefit to mean visual similarity (left) and retrospective nutritional distance (right), by query set and retrieval encoder. Filled markers indicate Benjamini–Hochberg-corrected

q<0.05

; open markers indicate non-significant associations. Positive visual-similarity correlations indicate greater benefit for more visually similar exemplars, whereas negative nutritional-distance correlations indicate greater benefit for nutritionally closer exemplars.

			Per-item									Whole-meal
			$\mathrm{\%MAE}_{\text{weight}}$			$\mathrm{\%MAE}_{\text{energy}}$			$\mathrm{\%MAE}_{\text{micro/CHO}}$			$\mathrm{\%MAE}_{\text{weight}}$			$\mathrm{\%MAE}_{\text{energy}}$			$\mathrm{\%MAE}_{\text{micro/CHO}}$
Query set	Model	Encoder	$\tau{=}0$	Def.	$K{=}10$	$\tau{=}0$	Def.	$K{=}10$	$\tau{=}0$	Def.	$K{=}10$	$\tau{=}0$	Def.	$K{=}10$	$\tau{=}0$	Def.	$K{=}10$	$\tau{=}0$	Def.	$K{=}10$
Nutrition5k	Gemma-4-31B	PE-Core	68.6	64.0	68.4	53.7	50.1	53.0	60.1	60.4	61.8	66.9 ▼	50.7	70.5	52.8 ▼	44.0	57.2	57.4 ▼	50.0	65.4
		SigLIP2	70.8 ▼	45.7	75.7	54.2 ▼	44.8	60.4	61.1 ▼	49.8	73.1 ▼	64.0 ▼	47.9	53.3	55.3 ▼	44.3	50.5	58.8 ▼	48.4	55.4
		Nomic Vision	66.6	64.4	66.7	79.7	72.2	103.9	87.8	80.7	102.8	70.5 ▼	52.5	64.9	75.0	55.7	70.4	85.5	62.5	74.5
	GPT-5.4-mini	PE-Core	31.1	38.5	34.9	28.6	32.8	32.8	32.9	40.0	36.1	39.3	35.8	31.1	34.0	34.7	30.6	41.2	40.4	34.4

Table 5: Retrieval ablations on the fixed 100-query subsets, with per-item and whole-meal prompting side by side (%MAE, lower is better). Default is the main configuration (

\tau=0.75

K=5

); the other columns remove the threshold (

\tau=0

) or increase depth (

K=10

\mathrm{\%MAE}_{\text{micro/CHO}}

\mathrm{\%MAE}_{\text{micro}}

for Nutrition5k, SNAPMe, and MADiMa and

\mathrm{\%MAE}_{\text{CHO}}

for EDUCATE; dashes mark unavailable targets. Bold: lowest error within each group of three. ⋆ / ▼ : significantly lower/higher error than Default (paired Wilcoxon signed-rank test, Benjamini–Hochberg corrected,

q<0.05

Figure S1 : Effect of acquisition distance on Vanilla VLM inference for MADiMa.

\mathrm{\%MAE}_{\text{macro}}

and

\mathrm{\%MAE}_{\text{micro}}

(lower is better) are shown at 40 cm and 60 cm for per-item and whole-meal prompting across the four evaluated VLMs.

			Per-item							Whole-meal
Query set	Model	Setting	Weight	Energy	CHO	Protein	Fat	Macro	Micro	Weight	Energy	CHO	Protein	Fat	Macro	Micro
Nutrition5k	Gemma-4-31B	Vanilla	36.3	47.7	56.0	44.3	63.3	54.5	53.6	49.2	58.4	69.6	46.4	78.4	64.8	63.5
		NutriVisRAG (PE-Core)	48.4 $\uparrow$ ▼	46.0 $\downarrow$	61.0 $\uparrow$	45.8 $\uparrow$	52.9 $\downarrow$ ⋆	53.2 $\downarrow$	53.7 $\uparrow$	47.4 $\downarrow$ ⋆	44.9 $\downarrow$ ⋆	55.9 $\downarrow$ ⋆	43.9 $\downarrow$ ⋆	53.4 $\downarrow$ ⋆	51.0 $\downarrow$	51.0 $\downarrow$ ⋆
		NutriVisRAG (SigLIP2)	47.6 $\uparrow$ ▼	47.8 $\uparrow$	58.6 $\uparrow$	47.1 $\uparrow$	58.8 $\downarrow$ ⋆	54.9 $\uparrow$	54.6 $\uparrow$	46.7 $\downarrow$ ⋆	45.3 $\downarrow$ ⋆	54.6 $\downarrow$ ⋆	43.5 $\downarrow$ ⋆	55.1 $\downarrow$ ⋆	51.1 $\downarrow$	50.8 $\downarrow$ ⋆
		NutriVisRAG (Nomic Vision)	46.3 $\uparrow$ ▼	51.4 $\uparrow$	59.8 $\uparrow$	55.5 $\uparrow$	66.0 $\uparrow$	60.5 $\uparrow$	59.8 $\uparrow$	46.5 $\downarrow$ ⋆	49.6 $\downarrow$ ⋆	56.5 $\downarrow$ ⋆	51.2 $\uparrow$	61.4 $\downarrow$ ⋆	56.4 $\downarrow$	55.8 $\downarrow$ ⋆
	GPT-5.4-mini	Vanilla	35.8	47.2	66.2	45.8	62.2	58.0	58.0	40.1	48.1	60.8	47.2	69.9	59.3	58.1

Table S1 : Complete nutrient-estimation results with and without NutriVisRAG (%MAE, lower is better), for both prompting strategies side by side. Dashes indicate unavailable evaluation targets for the corresponding dataset. Bold: lowest Macro and Micro (or CHO) error within each model block and prompting strategy; arrows: increase (

\uparrow

) or decrease (

\downarrow

) relative to the Vanilla baseline of the same prompting strategy. ⋆ marks a statistically significant improvement over that baseline and ▼ a statistically significant degradation (paired Wilcoxon signed-rank test on per-meal absolute errors, Benjamini–Hochberg corrected,

q<0.05

); the Macro column is an average of per-nutrient errors and is therefore not tested. Arrows indicate the direction of change based on unrounded values; displayed values are rounded to one decimal place.

			Per-item							Whole-meal
Query set	Model	Setting	Weight	Energy	CHO	Protein	Fat	Macro	Micro	Weight	Energy	CHO	Protein	Fat	Macro	Micro
MADiMa	Gemma-4-31B	Vanilla	51.2	49.1	60.1	38.9	64.2	54.4	54.7	76.5	69.5	76.4	48.5	89.1	71.3	70.7
		NutriVisRAG (PE-Core)	38.8 $\downarrow$ ⋆	40.5 $\downarrow$ ⋆	54.6 $\downarrow$ ⋆	41.7 $\uparrow$	52.7 $\downarrow$ ⋆	49.7 $\downarrow$	50.5 $\downarrow$ ⋆	71.9 $\downarrow$ ⋆	62.4 $\downarrow$ ⋆	64.9 $\downarrow$ ⋆	43.1 $\downarrow$ ⋆	85.8 $\downarrow$	64.6 $\downarrow$	62.7 $\downarrow$ ⋆
		NutriVisRAG (SigLIP2)	38.3 $\downarrow$ ⋆	38.8 $\downarrow$ ⋆	51.1 $\downarrow$ ⋆	37.5 $\downarrow$	53.3 $\downarrow$ ⋆	47.3 $\downarrow$	47.5 $\downarrow$ ⋆	69.7 $\downarrow$ ⋆	64.6 $\downarrow$ ⋆	62.2 $\downarrow$ ⋆	45.1 $\downarrow$	93.8 $\uparrow$	67.0 $\downarrow$	63.5 $\downarrow$ ⋆
		NutriVisRAG (Nomic Vision)	41.5 $\downarrow$ ⋆	40.1 $\downarrow$ ⋆	53.8 $\downarrow$ ⋆	40.8 $\uparrow$	58.3 $\downarrow$	51.0 $\downarrow$	50.9 $\downarrow$ ⋆	74.3 $\downarrow$	57.5 $\downarrow$ ⋆	64.1 $\downarrow$ ⋆	40.8 $\downarrow$ ⋆	79.2 $\downarrow$ ⋆	61.4 $\downarrow$	60.3 $\downarrow$ ⋆
	GPT-5.4-mini	Vanilla	53.8	45.0	75.8	32.8	46.9	51.8	57.4	93.0	76.4	89.2	38.5	96.2	74.6	75.7

Table S1 : (continued) Complete nutrient-estimation results with and without NutriVisRAG.

VLM	Prompt	Vanilla	PE-Core	SigLIP 2	Nomic Vision
Gemma-4-31B	Per-item	49.8	50.1	48.7	51.9
	Whole-meal	59.3	52.1	51.6	52.6
GPT-5.4-mini	Per-item	52.2	43.0	44.0	45.4
	Whole-meal	58.5	45.0	46.0	47.5
Mistral-Small-3.2	Per-item	67.5	94.1	109.5	112.2
	Whole-meal	62.2	60.8	64.7	69.9

Table S2 : Descriptive unweighted cross-dataset mean

\mathrm{\%MAE}_{\mathrm{micro}}

for each VLM, prompting strategy, and inference setting across Nutrition5k, SNAPMe, and MADiMa (lower is better).

Figure S2 : t-SNE visualisation of the retrieval embedding spaces for PE-Core-G14, SigLIP 2, and Nomic Vision. Gray points represent the combined Nutrition5k–SNAPMe reference corpus, while coloured points represent evaluation queries from Nutrition5k, SNAPMe, EDUCATE, and MADiMa.

Figure S3 : VLM-specific Spearman associations between retrieval benefit and mean visual similarity (left) or retrospective nutritional distance (right) under per-item prompting. Rows represent VLMs, and markers represent retrieval encoders. Filled markers indicate Benjamini–Hochberg-corrected

q<0.05

; open markers indicate non-significant associations.

Figure S4 : VLM-specific Spearman associations between retrieval benefit and mean visual similarity (left) or retrospective nutritional distance (right) under whole-meal prompting. Rows represent VLMs, and markers represent retrieval encoders. Filled markers indicate Benjamini–Hochberg-corrected

q<0.05

; open markers indicate non-significant associations.

Figure S5 : Mean visual similarity versus normalised nutritional distance between queries and retrieved exemplars. Rows represent query sets and columns represent PE-Core, SigLIP 2, and Nomic Vision. Lower nutritional distance indicates greater nutritional proximity.

Figure S6 : Meal-level absolute error of Vanilla versus NutriVisRAG under per-item prompting. Error is $|\Delta\mathrm{CHO}|+|\Delta\mathrm{PROT}|+|\Delta\mathrm{FAT}|$ in grams (absolute CHO error for EDUCATE). Rows: query sets; columns: retrieval encoders, each with Gemma-4-31B, GPT-5.4-mini, Mistral-Small-3.2, and GLM-5.3-Flash. Points below the diagonal favour NutriVisRAG, and each panel reports the number of improved meals. Both axes are logarithmic, with $10^{-3}$ g added to all errors so that zero errors remain visible.

Figure S7 : Bland–Altman analysis of Vanilla and NutriVisRAG predictions under per-item prompting. The analysed quantity is $\mathrm{CHO}+\mathrm{PROT}+\mathrm{FAT}$ in grams (CHO for EDUCATE); panel layout as in Fig. S6 . Each point shows the difference (predicted minus reference) against the mean of the two. For each condition, solid and dashed lines show the mean difference and $\pm 1.96$ SD limits of agreement, respectively.

Figure S8 : Meal-level absolute error of Vanilla versus NutriVisRAG under whole-meal prompting. Error is $|\Delta\mathrm{CHO}|+|\Delta\mathrm{PROT}|+|\Delta\mathrm{FAT}|$ in grams (absolute CHO error for EDUCATE). Rows: query sets; columns: retrieval encoders, each with Gemma-4-31B, GPT-5.4-mini, Mistral-Small-3.2, and GLM-5.3-Flash. Points below the diagonal favour NutriVisRAG, and each panel reports the number of improved meals. Both axes are logarithmic, with $10^{-3}$ g added to all errors so that zero errors remain visible.

Figure S9 : Bland–Altman analysis of Vanilla and NutriVisRAG predictions under whole-meal prompting. The analysed quantity is $\mathrm{CHO}+\mathrm{PROT}+\mathrm{FAT}$ in grams (CHO for EDUCATE); panel layout as in Fig. S8 . Each point shows the difference (predicted minus reference) against the mean of the two. For each condition, solid and dashed lines show the mean difference and $\pm 1.96$ SD limits of agreement, respectively.

			Energy					Micro / CHO
Query set	Model	Encoder	NutriVisRAG	Lexical	Qwen3	Gemma	Nomic Text	NutriVisRAG	Lexical	Qwen3	Gemma	Nomic Text
Nutrition5k	GPT-5.4-mini	PE-Core	32.8	96.0 ▼	57.9	90.6 ▼	78.3 ▼	40.0	131.9 ▼	61.5	132.9 ▼	83.4 ▼
		SigLIP2	30.2	147.3 ▼	48.2	82.7 ▼	77.7 ▼	35.6	172.7 ▼	54.3	126.8 ▼	86.0 ▼
		Nomic Vision	29.9	127.7 ▼	54.5 ▼	90.4 ▼	80.8 ▼	37.6	161.2 ▼	57.6 ▼	137.4 ▼	81.5 ▼
	Gemma-4-31B	PE-Core	50.1	144.4 ▼	70.1	117.9 ▼	82.9 ▼	60.4	176.2 ▼	84.5	164.5 ▼	96.4 ▼
		SigLIP2	44.8	128.7 ▼	59.0	80.5 ▼	74.2 ▼	49.8	147.6 ▼	64.0	106.5 ▼	82.2 ▼

Table S3 : Post-hoc food-composition grounding under per-item prompting on the fixed 100-query subsets (%MAE, lower is better). NutriVisRAG denotes the original prediction; the other columns are the matching approaches defined in the text. Dashes indicate unavailable evaluation targets for the corresponding dataset. Bold indicates the lowest error per row and metric; ⋆ / ▼ indicate significant improvements/degradations relative to NutriVisRAG (paired Wilcoxon signed-rank test, Benjamini–Hochberg corrected,

q<0.05

			Per-item								Whole-meal
			Energy				Micro / CHO				Energy				Micro / CHO
Query set	Model	Encoder	Default	Jina	Qwen3	Nem.	Default	Jina	Qwen3	Nem.	Default	Jina	Qwen3	Nem.	Default	Jina	Qwen3	Nem.
Nutrition5k	GPT-5.4-mini	PE-Core	32.8	29.9	28.7	38.5	40.0	35.2	35.6	43.5	34.7	31.5	36.8	47.4	40.4	36.9	40.9	49.8
		SigLIP2	30.2	33.0	30.6	30.2	35.6	41.7	34.2	36.7	36.6	33.1	34.6	37.6	39.9	39.7	40.5	41.6
		Nomic Vision	29.9	30.5	27.6	37.6	37.6	38.2	36.6	45.1	40.1	38.5	36.6	39.6	46.8	44.1	40.3	47.6
	Gemma-4-31B	PE-Core	50.1	57.1	51.5	71.9 ▼	60.4	63.9	55.3	77.8 ▼	44.0	50.4	50.5	65.5 ▼	50.0	57.1	57.3	68.5 ▼

Table S4 : Multimodal reranking on the fixed 100-query subsets, with per-item and whole-meal prompting side by side (%MAE, lower is better); 50 candidates are reranked to five exemplars. NutriVisRAG denotes the default configuration without reranking. Dashes indicate unavailable evaluation targets for the corresponding dataset. Bold indicates the lowest error per row and metric; ⋆ / ▼ indicate significant improvements/degradations relative to NutriVisRAG (paired Wilcoxon signed-rank test, Benjamini–Hochberg corrected,

q<0.05

In-Context Learning

In-context learning studies how models adapt to new tasks from examples or instructions supplied at inference time, without changing their weights.

See all 418 papers

NutriVision: Ingredient-Conditioned Fusion and Prediction for Single-Image Food Nutrition Estimation
Sep 27, 2026Avinash Anand, Aman Kumar, Chaitanya Lakhchaura +4Feature Fusion Multi-Source Image Fusion
Singapore Institute of Technology, Singapore · Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, India · Rochester Institute of Technology, Rochester, NY, USA +1
NutriMLLM: Multimodal Large Language Models for Dietary Micronutrient Analysis
Jun 8, 2026Runze Yan, Minxiao Wang, Jiaying Lu +3Multimodal Large Language Models
Center for Data Science, Nell Hodgson Woodruff School of Nursing, Emory University, 1520 Clifton Rd NE, Atlanta, 30322, GA, United States. · Rollins School of Public Health, Emory University, 1518 Clifton Rd2026 NE, Atlanta, 30322, GA, United States.
Improving Image-Based Nutrition Estimation Through Multimodal Food-Item Verification and Recovery
Oct 8, 2026Jingbo Yue, Bruce Coburn, Jinge Ma +2Visual Grounding
Elmore Family School of Electrical and Computer Engineering, Purdue University West Lafayette, IN, USA

Sep 27, 2026J/K move · Enter open · S save

NutriVision: Ingredient-Conditioned Fusion and Prediction for Single-Image Food Nutrition Estimation

Avinash Anand, Aman Kumar, Chaitanya Lakhchaura, Ashutosh Kumar+3

Singapore Institute of Technology, Singapore · Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, India · Rochester Institute of Technology, Rochester, NY, USA+1

Feature Fusion Multi-Source Image Fusion

Open paper AskPDF arXiv

Nutrition estimation is a fundamental task in consumer diet tracking, clinical dietetics, chronic disease management, sports and hospital nutrition, and broader food computing systems. The existing approaches have progressed along two largely separate axes, vision models that rely on calibrated RGB-depth captures and ingredient-aware methods that use textual cues but use limited multimodal fusion. We introduce NutriVision, an end-to-end framework that leverages visual geometry and ingredient semantics to estimate calories, mass, fat content, carbohydrates, and protein from a single RGB image and an optional ingredient list. It obtains the unavailable depth modality using DepthAnything-V3 and encodes ingredient descriptions using CLIP. It integrates three complementary mechanisms: (1) an Ingredient-Conditioned Frequency-Aligned Fusion Module (IC-FAFM), which uses textual guidance to reweight and align RGB-depth frequency components; (2) an Ingredient-Aware Mask-based Prediction Head (IA-MPH), whose gating and channel masks are conditioned on food identity; and (3) modality-specific Internal Semantic Modeling (ISM) blocks. On the Nutrition5k dataset, NutriVision achieves a mean PMAE of

\mathbf{13.60\pm0.10\%}

, outperforming our IGSMNet implementation by

0.89

percentage points and OmniFood8k by

2.90

percentage points (both

p<0.001

). The module-level ablations identify the ingredient-aware prediction head as the primary architectural contributor, improving mean PMAE by

1.50\pm0.17

percentage points (

p<0.001

). These results demonstrate that ingredient-conditioned prediction and frequency-aware RGB-depth fusion provide measurable gains for single-image nutrient estimation. More broadly, NutriVision offers a practical route toward nutrition-assessment systems that exploit geometric and semantic cues without requiring specialized depth-sensing hardware

Abstract

In-Context Learning

Explore similar work

NutriVision: Ingredient-Conditioned Fusion and Prediction for Single-Image Food Nutrition Estimation

NutriMLLM: Multimodal Large Language Models for Dietary Micronutrient Analysis

Improving Image-Based Nutrition Estimation Through Multimodal Food-Item Verification and Recovery