Unified Multimodal Eval

---
name: unified-multimodal-eval
description: Evaluates a unified discrete diffusion framework's capability to jointly generate and reason over image-text pairs. It probes unconditional and conditional generation quality, the effectiveness of classifier-free guidance, training and inference efficiency, and cross-modal retrieval and reasoning performance. Use when the user wants to benchmark on DataComp1B, CC12M, MS-COCO30k, Flickr, Winoground, or asks about evaluating this task. Reports FID.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2503.20853
  bibtex_key: swerdlow2025unified
  confidence: high
---

# unified-multimodal-eval

> Unified Multimodal Discrete Diffusion — Swerdlow et al. (2025) (arXiv:2503.20853, 2025)

## What this evaluates

Evaluates a unified discrete diffusion framework's capability to jointly generate and reason over image-text pairs. It probes unconditional and conditional generation quality, the effectiveness of classifier-free guidance, training and inference efficiency, and cross-modal retrieval and reasoning performance.

## Datasets

- **DataComp1B** — total ?; splits: val (-1)
- **CC12M** — total ?; splits: val (-1)
- **MS-COCO30k** — total 30000; splits: test (30000)
- **Flickr** — total ?; splits: test (-1)
- **Winoground** — total ?; splits: test (-1)

## Metrics

- `FID` **(primary)** — range: other
  - Fréchet Inception Distance. Measures the distance between feature distributions of real and generated images to quantify quality and diversity.
- `CLIP score` — range: [-1, 1]
  - Cosine similarity between image and text embeddings in the CLIP latent space. Used to evaluate image-text coherence.
- `Joint perplexity` — range: other
  - Exponential of the average negative log-likelihood across both image and text tokens. Lower indicates better fitting.
- `Retrieval accuracy` — range: [0, 1]
  - Fraction of queries where the model assigns the highest probability to the correct candidate image, text, or joint pair among N options.

## Input / output format

**Input**: Conditioned on an image to generate a text caption, or conditioned on a text caption to generate an image. For retrieval tasks, given a query (text or image) paired with N candidate images or texts.

**Output**: Generated sequence of discrete tokens (text or image patches), or probability scores p(x^img|x^txt), p(x^txt|x^img), or p(x_img,x_txt) for ranking candidates.

## Scoring recipe

```python
def compute_retrieval_accuracy(probs, gold_idx, n_candidates=16):
    # probs: list of model probabilities for each candidate given the query
    # gold_idx: index of the ground-truth correct candidate
    predicted_idx = probs.index(max(probs))
    return 1.0 if predicted_idx == gold_idx else 0.0
```

## Common pitfalls

- Solely relying on generative perplexity, which can be artificially low due to token repetition without capturing generation diversity.
- Assuming classifier-free guidance (CFG) scales identically for AR and diffusion models; AR is highly sensitive to CFG weighting with a narrow optimal range.
- Applying FID to text evaluation; the protocol explicitly uses CLIP score for text-image coherence since no direct FID equivalent exists for text.

## Evidence (verbatim from paper)

> We consider the following three evaluation metrics, most commonly used in previous works: i) Joint perplexity indicates a model’s ability to fit to different validation sets. Note that this metric is jointly calculated across image-text tokens. ... ii) Fréchet inception distance (FID) *Heusel et al. ([2017])* is a popular metric in image-generation to quantify the quality and diversity of image generation.iii) CLIP score is used for calculating image-text coherence.

## Citation

```bibtex
@misc{swerdlow2025unified,
  title={Unified Multimodal Discrete Diffusion},
  author={Swerdlow et al. (2025)},
  year={2025},
  note={arXiv:2503.20853}
}
```

- arXiv: 2503.20853