vanessaguarino/TFflowers-diffusion-assessment

Flower Diffusion Assessment Results

This dataset contains a FiftyOne export of an evaluation performed on flower images generated by a text-to-image conditional UNet/DDPM diffusion model with Classifier-Free Guidance (CFG).

• Content: Generated TF-Flowers images (daisy, roses, sunflowers)
• Conditioning: Text prompts via CLIP embeddings
• Purpose: Quality, diversity, and representation analysis of conditional diffusion outputs
• Split: Test / Evaluation
• Format: FiftyOneDataset

This dataset includes 500 generated flower images, each enriched with embeddings and evaluation metrics to support both quantitative and qualitative analysis.

Below is a screenshot of the FiftyOne App used for interactive inspection of generated images, embeddings, CLIP scores, and brain metrics,

Dataset Contents

Each sample contains:

• Image: Generated flower image
• Prompt: Text prompt used for conditional generation
• CLIP Score: Semantic alignment between image and prompt
• UNet Embedding: Intermediate feature vector extracted from the UNet bottleneck
• FiftyOne Brain Metrics:
  • Uniqueness: Measures how distinct a sample is relative to others
  • Representativeness: Measures how central a sample is within the learned embedding space

Purpose

• Evaluate text-to-image generation quality under classifier-free guidance
• Analyze internal UNet representations learned during conditional diffusion
• Study how semantic prompts map to visual structure in low-resolution settings
• Inspect diversity, outliers, and prototype samples using embedding-based metrics

Key Observations

• Generated images exhibit strong class-level semantic alignment, particularly for radially symmetric flowers (e.g., daisies, sunflowers)
• CLIP scores indicate reliable color–class correspondence, while more detailed prompts remain challenging at 32×32 resolution
• Representativeness highlights prototypical samples (e.g., centered daisies), whereas uniqueness captures variations in color, multiplicity, and spatial layout
• Limited resolution and dataset size constrain fine-grained texture modeling, despite effective guidance

Model & Evaluation Setup

• Generator: UNet/DDPM with Classifier-Free Guidance
• Guidance Weight: Optimized at w = 3.0
• Evaluation Metrics: CLIP score, FID, UNet embeddings, FiftyOne Brain metrics
• Tracking: Experiments logged in the W&B project
  diffusion-model-assessment-v2

---

---

We can download the dataset snapshot using snapshot_download, then create a new FiftyOne dataset from the local files. Finally, we launch the FiftyOne app to visually confirm that all samples, metadata, and computed metrics have been preserved correctly.

from huggingface_hub import snapshot_download
HF_REPO_ID = "vanessaguarino/TFflowers-diffusion-assessment"

# Download the snapshot from Hugging Face
local_dir = snapshot_download(repo_id=HF_REPO_ID, repo_type="dataset")
fiftyone_dataset_name = "restored_flower_assessment"

if fiftyone_dataset_name in fo.list_datasets():
    fo.delete_dataset(fiftyone_dataset_name)
    
test_data_dir = os.path.join(local_dir, "test")

# Load it back into FiftyOne
restored_dataset = fo.Dataset.from_dir(
    dataset_dir=test_data_dir, # Also, local_dir would work, as it is a copy of test folder
    dataset_type=fo.types.FiftyOneDataset,
    name=fiftyone_dataset_name
)

# Verify
print(restored_dataset)
fo.launch_app(restored_dataset)