A Glance Is All You Need: Single-Pass Fine-Grained
Image Captioning with SimLoss

Suryaansh Jain1, Rahasya Barkur1, Vishal G1, Ryan Rossi, Franck Dernoncourt2, Jack Wang2, Koustava Goswami2, Nedim Lipka2, Puneet Mathur2, Samyadeep Basu2, Seunghyun Yoon2
1 University of Massachusetts Amherst  ·  2 Adobe Research
Work done while the author was at Adobe Research.
UMass Amherst Adobe Research
A photo of a cat sitting on a wooden table between a ceramic lamp and a spiral notebook, annotated with human-written fine-grained region descriptions for the lamp, wall, cat, table, book, and floor.
Fine-grained captioning requires visually specific details. A human-written hyper-detailed description (from ImageInWords) names the cone-shaped lamp with its ceramic base, the short-haired white-and-black cat sitting between the lamp and the spiral notebook, the wood grain of the table, and the wall's lined texture. A generic caption identifies the main objects while omitting the attributes, materials, counts, textures, and spatial relationships that distinguish the image. SimLoss recovers this detail in a single pass — matching the F1 of a five-stage verification pipeline at roughly 20× lower latency, without any caption targets during training.

Abstract

An image may be worth a thousand words, but most captioning models describe it in only a few. Modern vision-language models produce fluent high-level captions, yet routinely miss the attributes, counts, textures, materials, and spatial relations that make an image visually specific. Recent multi-stage systems recover some of these details through generation, decomposition, verification, and rewriting — at the expense of substantially higher inference latency.

We propose SimLoss, a reference-free embedding-space objective for single-pass fine-grained image captioning. SimLoss trains a vision-language model to align its projected hidden-state representation with a frozen image embedding through an InfoNCE contrastive loss, supplying a dense visual supervision signal before any text is decoded — and requiring neither human-written fine-grained captions nor pseudo-captions from a multi-stage pipeline. We instantiate it as SimLoss FFT, which backpropagates through a locally available embedding model, and SimLoss GRPO, which treats that model as a black-box reward.

Compared with single-pass, multi-stage verification, reward-optimized, and perception-aware baselines, SimLoss FFT achieves the highest precision while nearly matching the F1 of the multi-stage method — all while retaining single-pass inference and running roughly 20× faster than the multi-stage pipeline. SimLoss GRPO attains the strongest recall. Together, these results show that embedding-space supervision can recover the quality of multi-stage verification at the latency of a single-pass captioner.

The Problem: Fluent, but Visually Incomplete

Humans can glance at an image once and retain fine visual detail — the material of a lamp base, the pattern on a fabric, the number of repeated objects. Vision-language models, despite their fluency, often miss this level of specificity: “a lamp on a table” is not wrong, just incomplete. The task is not to generate longer captions — asking for more detail is known to increase hallucination. It is to improve visual coverage while maintaining factual grounding.

The supervision gap

Standard caption supervision is not designed for this setting. MS COCO captions are intentionally concise — 10.0 words on average in our sampled batch — while the hyper-detailed human descriptions of IIW-400 average 171.2 words, roughly 17× longer. Multi-stage pipelines can generate detailed pseudo-captions, but not clean supervision: verified CapMAS captions average 29.3 atomic propositions per image, of which 7.0 are judged false — a factuality ratio of 0.766. Distilling from them would transfer their unverifiable claims and stylistic artifacts along with their detail.

Key idea

View a fine-grained captioner as preserving discriminative visual information. A generic caption remains compatible with many similar images; a detailed caption should make the source image easy to identify. Ideally a caption ŷ maximizes the mutual information

I(v; ŷ) = H(v) − H(v | ŷ),

but this is intractable over discrete captions. SimLoss applies the same principle before decoding, in continuous representation space: align the VLM's pooled hidden state with a frozen image embedding that already captures visual similarity — no caption targets at all.

A diagnostic: where supervision falls short

Embedding-space similarities in a frozen Qwen3-VL-Embedding space sharpen the motivation. The base captioner is not disconnected from the image — its generations already sit close to the source image (0.6982 cosine). The gap is between what supervision provides and what evaluation demands: COCO's short human captions align with their own images at only 0.4794, far below the fine-grained regime IIW evaluates. SimLoss closes this gap with the image embedding itself, rather than with fine-grained human captions or pipeline-generated captions as targets.

IIW-400 · mean matched-pair cosine similarity
IIW humanCapMASQwen2.5-VL
Image0.66160.70000.6982
IIW human0.69280.6803
CapMAS0.7307
MS COCO · mean matched-pair cosine similarity
COCO humanQwen2.5-VL
Image0.47940.6982
COCO human0.5055

Method: Embedding-Space Distillation

During training, a frozen multimodal embedding model encodes the image while the trainable VLM processes the same image and prompt. The VLM's hidden states are mean-pooled, projected into the embedding space, and aligned contrastively with the frozen image embedding. At inference, the embedding model and projector are removed — captioning uses only the adapted VLM, in a single pass.

TRAINING INFERENCE single-pass generation No caption targets — supervision comes from image embeddings and batch identities Image batch {vi}, i = 1…N ❄︎ Frozen encoder EI Qwen3-VL-Embed Image target ziI = EI(vi) Fixed prompt x Qwen2.5-VL-7B ❄︎ frozen base trainable LoRA Δθ Token states {Hi,t} Mean pool hθ(vi, x) Projector gφ (2-layer MLP) Identity contrast sij = cos(ziI, zjV) i = j → positive pair j ≠ i → in-batch negatives InfoNCE SimLoss ziI vi zjV gradients update only LoRA Δθ and the projector gφ Image v + prompt x “Describe this image in detail.” Qwen2.5-VL + learned LoRA one forward pass Fine-grained caption ŷ ≈115 words, high precision Removed after training: encoder · projector · loss ❄︎ frozen (no gradients) trainable forward pass gradients
SimLoss fully differentiable fine-tuning (FFT). The frozen encoder produces image targets ziI, while Qwen2.5-VL's token states are mean-pooled and projected to zjV. Cosine scores for all batch pairs are separated into matched identities (sii, positive) and other identities (sij, j ≠ i, negatives) before entering InfoNCE. The loss updates only the LoRA adapters and the projector, without caption targets. At inference, the encoder, projector, and contrastive loss are removed.

The objective

Let EI be the frozen image encoder and gφ a learned two-layer MLP projector from the VLM hidden dimension into the image-embedding space. For a batch of N image–prompt pairs with ziI = EI(vi), ziV = gφ(hθ(vi, x)) and cosine similarities sij = cos(ziI, zjV), SimLoss optimizes the InfoNCE loss

SimLoss = − 1N Σi=1N  log  exp(sii / τ)Σj exp(sij / τ)

A generic representation that discards visual detail is hard to match uniquely to its source image among in-batch alternatives; a representation that preserves fine-grained attributes, counts, textures, materials, and spatial relations aligns strongly with the correct image embedding. SimLoss therefore rewards the captioner for retaining discriminative visual information before text generation occurs — no human-written fine-grained captions, no pseudo-captions, and, when the embedding model is differentiable, no sampling of captions during training.

SimLoss FFT

fully differentiable fine-tuning

local embedding modelgradient supervision

Gradients flow from the InfoNCE loss through the projector into the LoRA adapters; the base weights and encoder stay frozen. “Fully differentiable” describes the training signal, not full-parameter tuning. The signal is dense and applied before any discrete text is sampled.

best precision · 0.8485shortest captions · 114.9 ± 13.1 words

SimLoss GRPO

black-box embedding reward

API-only embedding modelreinforcement learning

When the embedding model offers no gradients (here: Gemini 2 Embed), sampled captions are scored with the reward rSim(v, ŷ) = cos(zI, zT) — image and caption embedded in the same space — and the policy is optimized with GRPO over sampled caption groups.

best recall · 0.6015best CLAIR · 0.858

Results on IIW-400

Following the CapMAS dual protocol: precision is the fraction of a caption's atomic propositions a multimodal judge finds supported by the image and the IIW reference description; recall is the fraction of human-verified multiple-choice questions answerable from the caption alone, with the image withheld. CLAIR is a reference-based LLM rating. Training uses MS COCO images only — the captions are discarded.

MethodModel / signalCLAIRPrecisionRecallF1Length (words)
Inference-time baselines
PlainQwen2.5-VL-7B0.8420.78840.60020.6815347.55 ± 46.82
CapMAS5-stage pipeline0.8540.84670.60030.7025189.71 ± 49.03
PAPO baselines
PAPOQwen2.5-7B0.8430.79050.59500.6790181.86 ± 30.47
PAPO + YOLOQwen2.5-7B0.8420.79360.59690.6813182.41 ± 32.71
Reward-optimized baselines
FeedQuillQwen2.5-VL-7B0.8330.79660.59670.6823164.15 ± 69.48
Ours
SimLoss FFTQwen3-VL-Embed0.8440.84850.59910.7023114.86 ± 13.06
SimLoss GRPOGemini 2 Embed0.8580.82270.60150.6949132.87 ± 34.02
SimLoss-PAPO-YOLOQwen2.5-7B0.8550.83400.59700.6959149.84 ± 29.56

Bold = best per column, underlined = second best. Length is mean ± standard deviation.

Quality–latency tradeoff

IIW-400 F1 vs. measured A100 seconds per image (log scale). Better is lower-right.
SimLoss (ours) SimLoss FFT Single-pass baselines Multi-stage pipeline
SimLoss FFT reaches 0.7023 F1 at 5.77 s/image — CapMAS needs 115.31 s for 0.7025. Hover a point for exact values.

Recall is flat; precision decides F1

Per-method precision (filled) and recall (hollow), sorted by precision.
Ours Single-pass baselines Multi-stage Recall
Recall varies by only 0.0065 across all eight methods — the entire F1 spread is driven by precision.

Qualitative Examples

Across eight IIW-400 examples, SimLoss more consistently describes scene depth, paired objects, visible text, and fine-grained materials, with less repetition. Pick an image to read the SimLoss FFT caption and what to notice — including one honest failure case.

IIW-400 test image aar_test_04859

Layered scene structure

aar_test_04859
SimLoss FFT · 122 wordsThis image captures a serene and majestic mountain landscape. In the foreground, there is a wooden bench placed on a grassy area, inviting viewers to sit and enjoy the view. The bench is simple in design, with a backrest and seat made of natural wood, blending harmoniously with the natural surroundings. The middle ground is dominated by a sea of clouds that blanket the lower slopes of the mountains, creating a misty and ethereal atmosphere. The clouds appear thick and low-lying, partially obscuring the lower parts of the mountains and adding a sense of depth and mystery to the scene. In the background, the mountains rise majestically, their peaks sharp and jagged, with some areas covered in snow, indicating higher altitudes. …
Most baselines list bench, mountains, and clouds as co-equal objects — and Plain VLM calls the sky “clear, with a light blue hue,” a straightforward error given the prominent cloud bank. Both SimLoss variants organize the scene into three depth layers, which directly covers more spatial and relational questions. Measured precision here: 100% for SimLoss FFT vs 96.3% (Plain) and 85.2% (CapMAS).

Same coverage, a third of the words

On the two images below, Plain VLM, CapMAS, and SimLoss FFT answer an identical set of evaluation questions — coverage is exactly equal — yet caption length varies by 3×. The baselines' surplus words are summaries, speculation, and imposed structure, and they coincide with lower precision. Conciseness is a byproduct of optimizing for information content: the most efficient way to keep a caption discriminative is to fill it with grounded detail and little else.

IIW-400 test image aar_test_04779

The peace-tie sign

All three methods answer the same 20 of 23 evaluation questions — coverage is identical. (aar_test_04779)

“The image shows a wooden sign attached to a wooden door or panel. The sign is white with black text and features an illustration of a knife. The text on the sign reads, ‘All weapons must be peace tied….’ The knife depicted on the sign has a wooden handle and a metallic blade. Below the sign, there is a piece of wood that appears to be part of the door or panel, and it is secured with some black zip ties. The background consists of weathered wooden planks, giving the sign a rustic appearance. The overall tone of the sign suggests a humorous or ironic message about the use of weapons, possibly indicating a playful or lighthearted context rather than a serious one.”

Plain VLM spends nearly a third of its words on a closing summary that repeats what was already said, plus speculation about the setting (“possibly a campsite, a cabin”). CapMAS infers the box's intended use without evidence. SimLoss FFT covers the same questions in 123 words at 100% measured precision against Plain VLM's 90.9%.

Measured precision —Plain VLM: 90.9% precisionSimLoss FFT: 100% precision
IIW-400 test image aar_test_04629

The stone carving

All three methods answer the same 18 of 24 evaluation questions — coverage is identical. (aar_test_04629)

“The image shows an intricate circular design etched into a textured surface, likely concrete or stone. The design features two intertwined, stylized faces with exaggerated, flowing lines that give them a dynamic and almost animated appearance. Each face has a large, central eye and a smaller eye on the side, both depicted with simple, curved lines. The eyes are surrounded by elaborate, swirling patterns that extend outward, creating a sense of movement and fluidity. The overall effect is reminiscent of ancient or tribal art, with a focus on symmetry and organic forms. The texture of the surface adds depth to the design, making it appear three-dimensional.”

Plain VLM structures its 395 words as a numbered breakdown with headers and a full closing summary, none of which adds answerable content. CapMAS interprets the figures as symbols of “unity, balance, or the interconnectedness of life” — unverifiable from the image. SimLoss FFT answers the same questions in 106 words at 100% precision against 94.1% (Plain) and 89.7% (CapMAS).

Measured precision —Plain VLM: 94.1% precisionCapMAS: 89.7% precisionSimLoss FFT: 100% precision

Why It Works

Why does FFT improve precision?

It aligns the VLM's continuous hidden state with a frozen image embedding — a dense signal applied before text generation. Unlike caption imitation, it never copies a pseudo-reference's wording or omissions; unlike RL, it avoids high-variance updates from sampled text. The model is pushed to preserve image-specific evidence in its internal representation.

Why does GRPO gain recall but lose precision?

The black-box reward — cosine similarity between caption and image embeddings — encourages broader coverage (best recall, 0.6015). But because it is applied after discrete generation, the model can also raise similarity by adding visually plausible content, which costs precision relative to FFT.

Why is CapMAS still marginally best in F1?

Explicit decompose–verify–rewrite is well suited to suppressing false positives. But the margin is 0.0002 F1 — bought with a five-stage pipeline at 115.31 s/image against 5.77 s. SimLoss FFT occupies the more favorable quality–latency point.

LLM-judge supervision is fragile

CapMAS filters propositions with a multimodal judge; FeedQuill rewards captions with one. LLM judges show a systematic agreeableness bias — accepting valid outputs at ~96% but flagging invalid ones below 25% — missing exactly the plausible hallucinations that matter. SimLoss's supervision is a frozen image embedding, not a judge verdict.

Limitations & outlook

SimLoss trades dependence on caption targets for dependence on a frozen embedding teacher: its biases and visual granularity determine which distinctions are rewarded — and greater specificity can amplify fine-category errors (the driftwood moose). The remaining gap to CapMAS suggests explicit verification still helps precision–recall balancing. Natural next steps: comparing or ensembling embedding teachers across domains, combining FFT's precision with GRPO's recall, and lightweight factuality checks that preserve the single-pass deployment path.

BibTeX

@misc{jain2026glanceneedsinglepassfinegrained,
      title={A Glance Is All You Need: Single-Pass Fine-Grained Image Captioning with SimLoss},
      author={Suryaansh Jain and Rahasya Barkur and Vishal G and Ryan Rossi and Franck Dernoncourt and Jack Wang and Koustava Goswami and Nedim Lipka and Puneet Mathur and Samyadeep Basu and Seunghyun Yoon},
      year={2026},
      eprint={2609.00591},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.00591},
}