January Food Vision V1

• january/food-vision-v1 is a specialized vision-only model designed for meal-level food analysis, identifying meal names, ingredients, and macronutrients from images.
• The model uses a Vision Transformer backbone and deterministic output without prompt engineering, achieving an Overall Score of 86.2 on the January Food Benchmark.
• It offers practical applications in diet tracking and public health by outperforming general-purpose vision-language models in ingredient recognition and nutritional accuracy.

january/food-vision-v1 is a specialized vision-only model for food image analysis, designed to infer a meal name, an ingredient list, and macronutrient values directly from a food photograph without prompt engineering or temperature sampling. It is introduced within the “January Food Benchmark (JFB): A Public Benchmark Dataset and Evaluation Suite for Multimodal Food Analysis,” where it serves as a domain-specialized baseline against general-purpose Vision-LLMs for automated nutritional analysis. On that benchmark, the model attains an Overall Score of 86.2, exceeding the best-performing general-purpose configuration by 12.1 points (Hosseinian et al., 13 Aug 2025).

1. Model definition and operational scope

january/food-vision-v1 is described as a specialized vision-only model, rather than a general-purpose, prompt-driven VLM. Its task scope is explicitly food image analysis at the meal level: meal recognition, ingredient identification, and macronutrient estimation. Architecturally, the model uses a Vision Transformer backbone, processes images at a resolution of $1024 \times 1024$ pixels, and was trained on a large private food dataset on the order of $10^5$ images, with no overlap with the public JFB benchmark (Hosseinian et al., 13 Aug 2025).

The model is also presented as deterministic and non-prompt-based. Given an image, it directly emits structured predictions for meal name, ingredient list, and macronutrient values, rather than relying on natural-language prompting, JSON-format coercion, or stochastic decoding. In the benchmark context, this determinism is treated as central to stable and repeatable evaluation.

Its intended application domain is automated nutritional analysis from user photos. The target use cases include diet tracking and food logging, personalized nutrition tools, and possible public health or epidemiological deployment where reliable calorie, carbohydrate, protein, and fat estimates are needed at scale. The model is therefore positioned less as a general visual reasoning system than as a narrow, production-oriented component for mobile health workflows.

2. Benchmark substrate: the January Food Benchmark

The benchmark used to define and evaluate january/food-vision-v1 is the January Food Benchmark (JFB), a public dataset of 1,000 food images sourced from users of a mobile health application who consented to research use. The images are explicitly real-world: variable lighting, diverse camera angles, and complex backgrounds are part of the benchmark design. JFB also excludes trivial cases by including only meals with more than two ingredients (Hosseinian et al., 13 Aug 2025).

Each image carries expert-validated annotations for meal name, ingredient list, quantities, and macronutrients. The annotation procedure begins with two cohorts: a “Liked Cohort,” where users accepted AI output, and a “Disliked Cohort,” where users corrected it. After filtering out meals with two or fewer ingredients, a human annotator reviews and corrects meal names, ingredient lists, quantities, and macronutrient values to create final ground-truth labels. The paper characterizes this as a rigorous expert-validation protocol.

JFB is not merely a food classification dataset. It is built for multi-task evaluation comprising meal name similarity, ingredient recognition, macronutrient estimation, and operational metrics such as latency and cost. Its cuisine distribution is also reported as diverse: American is the largest category at 31.7%, followed by Other/mixed at 29.8%. This makes the benchmark explicitly meal-centric and nutritionally annotated, rather than a collection of isolated food-category labels.

3. Evaluation methodology and composite scoring

The benchmark defines a multi-metric evaluation suite intended to reflect practical dietary-assessment requirements. Meal-name accuracy is measured by semantic similarity, not exact string match. If $E(\cdot)$ denotes the embedding function and $\text{pred}$ and $\text{gt}$ denote predicted and ground-truth meal names, the score is

$$\text{score} = \cos\big(E(\text{pred}), E(\text{gt})\big).$$

The implementation uses a single shared embedding model, OpenAI text-embedding-3-small, across all evaluated systems (Hosseinian et al., 13 Aug 2025).

Ingredient recognition is framed as semantic set matching between predicted ingredient list $P$ and ground-truth list $G$. A cost matrix is constructed from embedding cosine distances, the Hungarian algorithm is used for optimal one-to-one assignment, and a matched pair is counted as correct when cosine similarity exceeds $T = 0.75$. Precision, recall, and F1 are then computed, with Ingredient F1 used as the reported score.

Macronutrient estimation is evaluated across calories, carbohydrates, protein, and fat using Weighted Mean Absolute Percentage Error (WMAPE):

$$\text{WMAPE} = \frac{\sum_{m \in M} \left| \text{Actual}_m - \text{Predicted}_m \right|}{\sum_{m \in M} \text{Actual}_m}.$$

Lower values are better, and the weighting ensures that larger meals contribute proportionally more to the aggregate error.

A notable contribution of the paper is the Overall Score, defined as a weighted geometric mean over meal name similarity, ingredient F1, macronutrient accuracy via $10^5$0, and normalized cost and latency terms. The weights are $10^5$1 for meal name similarity, $10^5$2 for ingredient recognition, $10^5$3 for macronutrients, and $10^5$4 each for cost and latency. The use of a geometric mean is intended to penalize a single weak dimension, so that a model must remain balanced across recognition quality, nutritional accuracy, and operational constraints. The weights were chosen through an expert survey using SMART, followed by sensitivity analysis showing highly stable rankings under perturbations of $10^5$5 per weight, with Spearman’s $10^5$6 (Hosseinian et al., 13 Aug 2025).

4. Comparative performance against general-purpose VLMs

On JFB, january/food-vision-v1 records an Overall Score of 86.2. The best general-purpose baseline, GPT-4o (Best-of-4 prompts), records 74.1, yielding the reported 12.1-point improvement. The paper further states that this gap is statistically significant under a paired $10^5$7-test on per-image scores with $10^5$8 and $10^5$9 (Hosseinian et al., 13 Aug 2025).

The task-wise breakdown is similarly favorable. january/food-vision-v1 achieves meal name similarity 0.886, ingredient F1 0.883, macronutrient WMAPE 14.2%, latency 10.5 s/call, and cost \$E(\cdot)$00.0065/call. Other reported baselines include GPT-4o (Avg prompts) at 70.6 overall, GPT-4o-mini (Best) at 66.4, Gemini 2.5 Pro (Best) at 60.7, and Gemini 2.5 Flash (Best) at 60.7.

The most consequential gap lies in ingredient recognition, which is also the most heavily weighted component of the Overall Score. The model’s 0.883 F1 exceeds GPT-4o (Best) by approximately 0.146. Macronutrient estimation also improves materially: the reduction from 23.5% WMAPE to 14.2% corresponds to an error reduction of roughly 40%. Latency is reported as broadly comparable to GPT-4o, while cost is slightly higher per call. The paper additionally reports tighter interquartile ranges in per-image score distributions for january/food-vision-v1, indicating greater consistency across individual meals.

The paper attributes this performance difference to three factors: domain-specific training data, task specialization for meals, ingredients, and macros, and deterministic structured output rather than prompt-sensitive free text. The emphasis on ingredient recognition is explicit: errors at the ingredient stage are said to cascade into macronutrient estimation.

5. Capabilities, design choices, and limitations

The reported results support several concrete capabilities. january/food-vision-v1 can identify dishes at a semantically meaningful level, recognize constituent ingredients with high fidelity under embedding-based matching, and estimate calories, carbohydrates, protein, and fat at substantially lower error than the evaluated general-purpose VLMs. Its cost and latency profile is also described as suitable for interactive mobile applications (Hosseinian et al., 13 Aug 2025).

The disclosed design choices are limited but consequential. The Vision Transformer backbone and $E(\cdot)$1 input resolution indicate an emphasis on fine-grained visual detail, including small ingredients and subtle textures. The exact training losses are not provided, and the paper explicitly states that proprietary details are withheld. This suggests, but does not establish, a multi-task optimization regime aligned to meal recognition, ingredient prediction, and macronutrient regression.

The paper does not enumerate model-specific failure modes, but it identifies general challenges that remain for any food-analysis system. These include mixed dishes and occlusion, the difficulty of inferring portion size from a 2D image, high intra-class variance among visually similar dishes, possible underrepresentation of rare cuisines, and the persistent invisibility of hidden ingredients such as oil or sauces. The reported 14.2% WMAPE is therefore strong but not exact, and the paper notes that highly sensitive clinical applications may still require further improvement or human oversight.

6. Position in the research landscape

Within the broader dietary-assessment literature, january/food-vision-v1 occupies a specific point in the design space: it is a domain-specialized, deterministic, vision-only model paired with a public benchmark and a task-composite evaluation scheme. The paper argues that JFB fills a gap left by datasets such as Food-101, which lacks ingredients and nutrition, Recipe1M+, which lacks complete validated nutrition, and MEAL, which is only partially validated and assembled from mixed sources (Hosseinian et al., 13 Aug 2025).

This placement is significant when compared with adjacent strands of food-vision research. Earlier vision-based dietary assessment systems are often organized as multi-stage pipelines—food image analysis, volume estimation, and nutrient derivation—or as end-to-end regressors from images to nutrition, with both paradigms limited by meal complexity and annotation scarcity (Wang et al., 2021). Contemporary evaluations of general-purpose VLMs for dietary assessment report strong coarse-grained recognition in single-product images but weaker reliability for fine-grained distinctions such as cooking style and multi-product scenes, which suggests why january/food-vision-v1 emphasizes structured ingredient and macronutrient outputs rather than prompt-based free-text reasoning (Romero-Tapiador et al., 9 Apr 2025). Subsequent work such as Food-R1 extends the field toward unified food VLMs trained with multi-task learning, Chain-of-Thought cold-start instruction tuning, and reinforcement fine-tuning, indicating a later research direction beyond the vision-only specialization represented by january/food-vision-v1 (Zhu et al., 3 Jun 2026).

JFB itself is reported as CC-BY-4.0, with dataset files, JSON annotations, metric scripts, prompts, and evaluation code available at https://github.com/January-ai/food-scan-benchmarks. In that sense, january/food-vision-v1 is not only a model result but also a reference point for subsequent work on automated nutritional analysis, particularly where balanced evaluation across semantic recognition, nutrient accuracy, latency, and cost is required.