GenAI Media Generation Challenge Workshop @ CVPR

Jun 17 (afternoon), 2024 Workshop Meeting Location: Summit 423-425

Challenge Overview

In the wake of the rapid advancements in generative model capabilities, the field of text-guided image generation and editing has seen unprecedented progress. This has resulted in the creation of images with unparalleled quality, aesthetics, and adherence to text guidance. Yet, a significant challenge remains: the absence of a universally-accepted, easily accessible benchmark for evaluating in-depth capabilities of the generated models. This issue stems from the lack of a comprehensive, large-scale evaluation dataset, standardized evaluation protocols, and the insufficiency of current automatic metrics.

We are proud to announce the launch of the GenAI Media Generation Challenge (MAGIC) that moves us towards addressing these issues. We host two challenge tracks: (1) text to image generation, and (2) text guided image editing, and will be providing the following set of data and resources:

• Benchmark Datasets: diverse and comprehensive datasets will be released publicly for both tracks, ensuring that all participants have access to the same information and resources.
• Evaluation Protocol: A standardized evaluation protocol and metrics will be established and shared with all participants. This metric will be designed keeping in mind the nuances and specifics of the tasks, ensuring fair and objective evaluation.
• Human Annotations for all Submissions: Recognizing the importance of meticulous human oversight, we commit to offering the requisite human annotation resources as stipulated by the evaluation protocol for the duration of the competition.
• Baselines: For aiding participants and setting a preliminary standard, baseline results will be shared.
• Awards for top performing participants for different tracks and challenges.

In addition to the benchmark dataset, after the challenge, we will be releasing all human annotations for all submissions for the participants. This hopefully would be a contribution to the community for future evaluation tooling development.

Tracks Overview

Track A - Text-to-image Generation

Benchmarking Datasets

For this track, we test the model's text to image generation capabilities using benchmark prompts with different scenarios and complexity. Benchmark prompts will be leveraging existing academic benchmarks and newly constructed prompts which expands to different levels of complexity. We aim to test different scenarios as below with different benchmark prompts:

1. Single object generation, focusing on generation accuracy and correctness of the object
2. Object composition, focusing on generation along with relationship and activity
3. Reasoning, focusing on concepts such as counting or logic
4. Stylization, focusing on correctness of global style
5. Rendered text, focusing on quality and correctness of the generated text in image.

Within these scenarios, the benchmark prompts would contain topics from various domains. For example, for single object generation, there prompts looking at generation of people, animals, food, and locations.

Evaluation Protocol

We leverage both human and automated evaluations. On human-based evaluations, we use human annotators to annotate the of the text2Image model performance on the following aspects:

1. Text Faithfulness - whether the visual elements in prompts such as objects, attributes, interactions are accurately generated in the image.
2. Visual quality - whether there are visual errors/defects in the generated image. Examples of errors/defects include misplaced body parts for people, unrealistic scale, bad arrangement, or blurry.
3. Visual appeal - whether the generated images are aesthetically pleasing to view. We judge whether models are either on-par or surpass professionally made ones in terms of abundant visual details, color harmony, and composition.

On automatic evaluation, we will leverage existing methods that have developed automatic metrics to help in assessing the outputs of the image based on the prompt.

To determine winners, we use automatic evaluation to help prune the total number entries to 10 finalists. Within the final 10 participants, we would use human evaluation results to determine the final winners.

Track B - Text-guided Image Editing

Benchmarking Dataset

For text guided image editing, we test the capacity of the model to change a given image's contents based on some text instructions. The specific type of instructions that we test for are the following:

1. Addition: Adding new objects within the images.
2. Remove: Removing objects
3. Local: Replace local parts of an object and later the object's attributes, i.e., make it smile
4. Global: Edit the entire image, i.e., let's see it in winter
5. Inpaint: Alter an object's visual appearance without affecting its structure
6. Background: Change the scene's background

Evaluation Protocol

We leverage both human and automated evaluations. On human-based evaluations, we use human annotators to annotate the we mainly evaluate the following aspects:

1. Edit faithfulness - whether the edited image follows the editing instruction
2. Content preservation - whether the edited image preserves the regions of the original image which should not be changed
3. Visual quality - whether the edited image is artifact-free, keeping the core visual features of the original image, etc

On automatic evaluation, similar to the text to image track, we will leverage existing methods that have developed automatic metrics to help in assessing the outputs of the image based on the prompt and instruction.

To determine winners, we use automatic evaluation to help prune the total number entries to 10 finalists. At 10, we would use human annotation and evaluation to determine the final winners.

Bonus Track - Movie Generation Challenge

The generated movie can be any video or topic, but should be created with generative AI solution. The movie should be no more than 5 minutes long. To submit to the bonus track, please use the link here: [Moviegen Challenge Submission]

Winner Prize

Top three contestants of each track will be awarded with the latest generation of AR/VR headsets from Meta.

---