EEVEE: Multi-Domain ML Artifacts

• EEVEE is an overloaded name referring to distinct artifacts in video virtual try-on, LLM prompt learning, and NLP annotation.
• In video virtual try-on, Eevee includes a high-resolution dataset, benchmark, and the novel VGID metric to assess detailed garment fidelity.
• EEVEE also drives innovations in self-improving LLM agents via router-conditioned prompt learning and in NLP with a lightweight browser-based annotation tool.

EEVEE is an overloaded research name used for several unrelated artifacts across machine learning and natural language processing. In the recent arXiv literature, it denotes at least three distinct systems: a high-resolution dataset, benchmark, and evaluation protocol for close-up video-based virtual try-on; a router-conditioned framework for test-time prompt learning in self-improving LLM agents; and a lightweight browser-based annotation tool for NLP datasets (Zeng et al., 24 Nov 2025, Xu et al., 9 Jun 2026, Sorensen et al., 2024). A related but terminologically distinct case is the eye-tracking dataset EVE, whose paper explicitly states that the acronym “EEVEE” is not used there (Park et al., 2020). The term therefore requires domain-sensitive disambiguation.

1. Disambiguation and scope

The name EEVEE has no single canonical meaning in the literature represented here. Instead, it appears in three separate lines of work.

Domain	Artifact	Principal role
Video virtual try-on	Eevee	High-resolution dataset, benchmark, and evaluation protocol with VGID
LLM agents	EEVEE	Router-conditioned test-time prompt learning framework
NLP data annotation	EEVEE	Browser-based annotation tool using tab-separated files

The virtual try-on work defines Eevee as a dataset and benchmark for video-based VTON, with explicit emphasis on close-up marketing videos, detailed garment imagery, and a garment-centric consistency metric called VGID (Zeng et al., 24 Nov 2025). The agent-learning work defines EEVEE as a multi-dataset test-time prompt learning framework in which a router assigns inputs to specialized prompt slots under heterogeneous task streams (Xu et al., 9 Jun 2026). The NLP paper defines EEVEE as an annotation environment that runs directly in the browser, uses tab-separated files, and supports sequence labeling, span labeling, text classification, and seq2seq (Sorensen et al., 2024).

This distribution of meanings suggests that “EEVEE” functions more as a project-specific label than as a stable technical acronym shared across subfields. In practice, the surrounding domain vocabulary—VTON, prompt learning, or annotation tooling—is essential for interpretation.

2. Eevee in video-based virtual try-on

In computer vision and fashion technology, Eevee refers to “Towards Close-up High-resolution Video-based Virtual Try-on,” a high-resolution dataset, benchmark, and evaluation protocol designed specifically for video-based virtual try-on with a focus on close-up marketing videos (Zeng et al., 24 Nov 2025). The motivating problem is that prior video VTON datasets rely heavily on a single in-shop garment image and on full-shot-only videos, which limits the preservation of realistic texture details, prints, material cues, and other garment-specific micro-structure.

The dataset provides both full-shot and close-up try-on videos of real human models. It also supplies high-fidelity in-shop garment images, additional detailed close-up garment images, textual descriptions per garment, and person images. The annotation and utility stack is heterogeneous: person parsing and masks for full-shot video and person images are OpenPose-derived; close-up garment masks use Grounded SAM-2; garment contour maps use AniLines; DensePose UV coordinates come from Detectron2; and text descriptions and category classification use Qwen-VL-Max (Zeng et al., 24 Nov 2025).

The reported scale is substantial: 9,364 person–garment pairs in total, split into 8,364 train and 1,000 test pairs; 28,092 images total excluding parsing annotations; and 1,905,973 total video frames. Category counts are given as upper-body 4492/500, lower-body 2308/250, and dresses 1564/250 for train/test. Garment images reach up to $2400 \times 1800$, and videos up to $1088 \times 816$ (Zeng et al., 24 Nov 2025). The data are collected from e-commerce websites, and the paper does not specify file formats, frame rates, licensing, or standardized capture hardware and lighting protocols.

A central methodological claim is that existing video generation models can exploit Eevee’s detailed garment imagery only after adaptation. The paper states that simple feed-in of detailed garment images to a pre-trained video generator did not help by itself; instead, the implemented VACE fine-tuning pipeline formulates try-on as masked video-to-video editing, encodes detailed garment images through the Wan VAE, prepends their latents to video latents, and uses LoRA fine-tuning (Zeng et al., 24 Nov 2025). The mask-based decoupling is written as

$F_r = F \odot M, \qquad F_i = F \odot (1 - M).$

Benchmarking covers ViViD, MagicTryOn, VACE, and an Eevee-fine-tuned VACE variant under paired and unpaired settings, full-shot and close-up scenarios, two resolutions, and 49-frame clips. In full-shot, $1088 \times 816$, paired evaluation, the fine-tuned VACE reports $\mathrm{VFID}_R^p = 0.076$, $\mathrm{VFID}_I^p = 6.555$, $\mathrm{VGID}^p = 0.534$, $\mathrm{SSIM} = 0.910$, and $\mathrm{LPIPS} = 0.078$. In close-up, $1088 \times 816$, paired evaluation, it reports $1088 \times 816$0, $1088 \times 816$1, $1088 \times 816$2, $1088 \times 816$3, and $1088 \times 816$4 (Zeng et al., 24 Nov 2025). The paper further states that VBench shows little separation across models on Subject Consistency, Background Consistency, Aesthetic Quality, and Imaging Quality, implying that global metrics are less sensitive to garment-specific fidelity differences in close-ups.

The close-up emphasis is not merely a presentational choice. The paper explicitly frames it as a business-critical requirement for fashion e-commerce, where buyers expect to inspect collars, seams, knits, print fidelity, logos, stitching, and material cues. In that sense, Eevee reorients video VTON from full-body plausibility toward fine-grained garment preservation.

3. VGID and garment-centric evaluation

The same Eevee work introduces VGID (Video Garment Inception Distance) as a garment-centric consistency metric intended to quantify preservation of garment texture and structure, especially in close-up shots (Zeng et al., 24 Nov 2025). Despite its name, the paper explicitly notes that VGID is not a Fréchet distance: it does not model Gaussian feature distributions and does not compute means and covariances. Instead, it is a cosine similarity over masked DINO-V2 features.

The feature extractor is DINO-V2, denoted $1088 \times 816$5. A soft garment mask is obtained from $1088 \times 816$6 patch attention from the last Transformer block, then upsampled and normalized to produce $1088 \times 816$7. For a source garment image $1088 \times 816$8 and generated frame $1088 \times 816$9, the method computes feature maps $F_r = F \odot M, \qquad F_i = F \odot (1 - M).$0 and $F_r = F \odot M, \qquad F_i = F \odot (1 - M).$1, resizes soft masks to feature-map resolution, masks the features elementwise, global-average-pools the masked representations, and computes cosine similarity: $F_r = F \odot M, \qquad F_i = F \odot (1 - M).$2 The video-level score is the average of per-frame VGID values over $F_r = F \odot M, \qquad F_i = F \odot (1 - M).$3 frames (Zeng et al., 24 Nov 2025).

The paper contrasts this construction with classical FID,

$F_r = F \odot M, \qquad F_i = F \odot (1 - M).$4

and argues that existing metrics such as VFID and VBench prioritize overall video realism or temporal stability rather than source-garment faithfulness (Zeng et al., 24 Nov 2025). That distinction is particularly important for close-up VTON, where structurally correct but texture-blurred outputs may still score reasonably on generic visual metrics.

A notable limitation is the soft-mask failure mode for upper-body and lower-body outfits. The paper states that attention-derived masks can bleed across adjacent items, so VGID is reported primarily for the dresses category to avoid cross-garment contamination (Zeng et al., 24 Nov 2025). This is an important corrective to any misconception that VGID provides a uniformly reliable garment-isolation signal across all apparel categories.

4. EEVEE for self-improving LLM agents

In LLM research, EEVEE refers to “Towards Test-time Prompt Learning in the Real World for Self-Improving Agents,” described as the first multi-dataset test-time prompt learning framework for LLM agents under heterogeneous real-world task streams (Xu et al., 9 Jun 2026). The paper’s problem setting is explicitly test-time prompt learning: the agent adapts a deployed foundation model by editing prompts online, without updating model weights.

The framework addresses a failure mode of single-prompt adaptation under heterogeneous streams. Inputs may come from multiple datasets, domains, and task distributions; in that regime, a single evolving prompt can suffer cross-dataset interference, catastrophic interference, and retention loss as new tasks are added (Xu et al., 9 Jun 2026). EEVEE’s response is architectural decoupling through specialization: a router partitions incoming inputs into task clusters and assigns them to specialized prompt slots.

The inference rule is

$F_r = F \odot M, \qquad F_i = F \odot (1 - M).$5

where $F_r = F \odot M, \qquad F_i = F \odot (1 - M).$6 is the router, $F_r = F \odot M, \qquad F_i = F \odot (1 - M).$7 is the prompt set, and $F_r = F \odot M, \qquad F_i = F \odot (1 - M).$8 is the frozen target model (Xu et al., 9 Jun 2026). Routing is hard rather than soft; the framework does not introduce soft prompts or adapters. Prompts are discrete instruction-style texts and may encode procedures, answer-format contracts, and domain-specific guidelines.

A central design feature is router–prompt co-evolution. Because prompt specialization depends on which examples the router sends to each slot, and router quality depends on what each prompt slot can already solve, the paper alternates router evolution and prompt evolution in cycles. It writes the cycle as

$F_r = F \odot M, \qquad F_i = F \odot (1 - M).$9

The router objective combines downstream accuracy, consistency, and balance: $1088 \times 816$0 The paper states that routing decisions are informed especially by per-example correctness vectors across prompts, and that the scoring weights anneal from regularization toward downstream accuracy (Xu et al., 9 Jun 2026).

Initialization is itself structured. EEVEE first learns a diverse prompt pool on mixed data, reduces it to a Pareto frontier using per-example correctness, and then greedily selects $1088 \times 816$1 prompts by coverage of frontier membership across validation examples: $1088 \times 816$2 This procedure, together with per-slot Pareto-front prompt pools, is intended to preserve complementary behaviors rather than collapse onto a single dominant prompt (Xu et al., 9 Jun 2026).

Empirically, the paper reports a four-benchmark suite consisting of GPQA Diamond, Formula, TheoremQA, and HumanEval. On Qwen3-4B-Instruct, EEVEE achieves a 51.75 average, improving the unadapted baseline by $1088 \times 816$3 points; on DeepSeek-V3.2, it achieves 64.07 average, $1088 \times 816$4 over baseline (Xu et al., 9 Jun 2026). It also states that EEVEE surpasses GEPA and ACE by up to $1088 \times 816$5 and $1088 \times 816$6. In the task-retention experiment, GEPA and ACE end at $1088 \times 816$7 and $1088 \times 816$8 cumulative retention, whereas EEVEE remains positive throughout and ends at $1088 \times 816$9 (Xu et al., 9 Jun 2026).

Ablations are central to the paper’s argument. A default router without learned routing reaches 43.58 average, a manual router 37.18, and a no-co-evolution two-stage variant 42.88, all below the full EEVEE result of 51.75 (Xu et al., 9 Jun 2026). The paper therefore attributes performance not merely to having multiple prompts, but to learned routing and interleaved router–prompt optimization.

5. EEVEE as an NLP annotation tool

In NLP infrastructure, EEVEE denotes “An Easy Annotation Tool for Natural Language Processing,” a browser-based annotation tool designed to simplify the creation of NLP datasets (Sorensen et al., 2024). Its defining choice is to operate on tab-separated, CoNLL-like files rather than on character-offset or task-specific formats.

The data model is token-centric. One token appears per line; empty lines separate utterances; columns are tab-separated; and optional comment lines above utterances, beginning with “#”, store metadata or utterance-level labels (Sorensen et al., 2024). The tool does not enforce fixed headers. Instead, the user specifies which column is input text and which column stores output labels. A representative schema given in the paper includes token index, token string, a sequence-label column, and a span-label column. Spans are represented tokenwise using BIO-style tags such as B-MISC, I-PER, and O, so no character offsets are required (Sorensen et al., 2024).

EEVEE supports four task types on a single dataset: sequence labeling, span labeling, text classification, and seq2seq (Sorensen et al., 2024). Sequence labeling assigns one label per token. Span labeling applies labels to contiguous token sequences; the interface selects entire tokens rather than character substrings. Text classification stores utterance-level labels in comments such as # intent = inform. Seq2seq stores the target text in comments as well, for example # translation = "Eevee entwickelt sich!" (Sorensen et al., 2024).

Configuration is expressed through JSON. A task definition includes title, type, input index, output index, and labels. The paper gives, for example, $\mathrm{VFID}_R^p = 0.076$0 for NER (Sorensen et al., 2024). Multiple tasks can be configured over the same TSV file, labels can be imported automatically if already present, and default labels can be assigned to empty cells.

The interface is optimized for speed. The setup page defines tasks, columns, labels, and defaults. The annotation page uses a minimal UI with keyboard mode and search mode: number keys select labels directly, while search mode activates automatically when a task has more than 10 labels (Sorensen et al., 2024). Navigation supports per-utterance status labels—completed, wrong, unsure, or cleared—and a progress bar provides a compact view of completion and uncertainty.

The paper emphasizes simplicity, privacy, and setup-free deployment. EEVEE runs directly in the browser, requires no installation, can also be saved as a desktop application for offline use, and does not store or send data to the network (Sorensen et al., 2024). User management is intentionally outside scope; the tool does not provide centralized accounts, built-in adjudication, inter-annotator agreement computation, or embedded guideline systems.

The reported empirical evaluation is a usability-oriented case study. Two annotators performed NER and German dialect identification on eight documents totaling 14.2K tokens over 16 hours per annotator, after a 12-minute tutorial. The resulting System Usability Scale scores were 75.0 and 87.5, with an average of 81.25 (Sorensen et al., 2024). Qualitative feedback highlighted ease of learning, keyboard-centric speed, and willingness to reuse the tool.

The limitations are explicit. EEVEE assumes existing tokenization; some languages may be unsupported if Unicode font handling is inadequate; the scope is limited to the four task types; and advanced features such as active learning, relation extraction, overlapping spans, or project-management subsystems are not described (Sorensen et al., 2024).

6. Relation to EVE and recurrent naming ambiguity

A separate vision paper, “Towards End-to-end Video-based Eye-Tracking,” is relevant because its title contains the phrase “End-to-end Video-based Eye-Tracking,” which could superficially invite the shorthand “EEVEE.” However, the paper explicitly states that the acronym “EEVEE” is not used; the official name is EVE (Park et al., 2020). This distinction matters for bibliographic accuracy.

EVE is a synchronized multimodal eye-tracking dataset consisting of user-facing camera frames, gaze and pupil size from a Tobii Pro Spectrum eye tracker, and time-synchronized screen-content video (Park et al., 2020). It includes 54 participants, 12,308,334 frames, approximately 105 hours of recordings, and 1327 stimuli. The method couples an appearance-based gaze estimator, EyeNet, with GazeRefineNet, a screen-content-conditioned refinement network that improves point-of-gaze estimation without supervised per-user adaptation.

The reported final performance is 2.49 degrees angular error and 2.75 cm, or 95.59 px, point-of-gaze error on a 25-inch 1080p display, representing up to a 28 percent reduction in angular error relative to a baseline that does not consider screen content (Park et al., 2020). The methodological lesson is that conditioning on synchronized visual stimuli and temporal context can compensate for person-specific biases such as angle kappa without test-user labels.

The inclusion of EVE in a discussion of EEVEE is therefore terminological rather than nominative. It documents a plausible source of confusion: one line of work uses the title phrase “End-to-end Video-based Eye-Tracking,” but the dataset and method are formally named EVE, not EEVEE (Park et al., 2020).

7. Comparative significance across domains

Across these papers, EEVEE labels artifacts that occupy very different positions in the research stack. In virtual try-on, Eevee is a dataset/benchmark/metric package centered on high-resolution visual fidelity and garment consistency (Zeng et al., 24 Nov 2025). In LLM-agent research, EEVEE is a test-time adaptation framework centered on routing, prompt specialization, and retention under heterogeneous task streams (Xu et al., 9 Jun 2026). In NLP tooling, EEVEE is an annotation environment centered on tabular data formats, browser execution, and lightweight workflow design (Sorensen et al., 2024).

Several contrasts are especially sharp. The VTON Eevee introduces a new evaluation metric, VGID, because existing video quality measures were said to miss garment-specific texture and structural fidelity in close-ups (Zeng et al., 24 Nov 2025). The agent-learning EEVEE instead focuses on online prompt adaptation without model weight updates, using correctness-vector-driven routing and Pareto-front preservation to mitigate cross-dataset interference (Xu et al., 9 Jun 2026). The annotation-tool EEVEE avoids algorithmic automation almost entirely, preferring low setup overhead, privacy, and keyboard efficiency over active learning or centralized project management (Sorensen et al., 2024).

A plausible implication is that the repeated use of the name reflects a broader pattern in contemporary ML literature: short memorable project names are often reused independently across subfields, even when the underlying objects—datasets, algorithms, metrics, and software tools—are unrelated. For precise scholarly communication, “EEVEE” should therefore almost always be accompanied by its domain qualifier, paper title, or arXiv identifier.