Summary
This paper explores the effectiveness of training a text-to-image (T2I) model using synthetic examples generated by existing T2I models. The authors find that on the order of 10M image-text pairs are necessary to approach the quality of a good model like PixArt-Alpha, while using only 1M or 100k examples results in serious deterioration.
To improve sample efficiency, the authors introduce their online learning framework, EvolveDirector, which leverages a vision-language model (VLM) to curate a better online training set. The VLM is used to compare the current trained checkpoint's generated images against the reference T2I model's images, in order to determine which training examples to drop from the training set, and which training example prompts to generate more variants of for further training. Occasionally, entirely new training prompts are introduced to increase diversity. Overall, this results in a training process which matches the quality of PixArt-Alpha according to human comparisons using only 100k examples.
The authors perform further ablations to show the necessity of each part of the framework, and train a model using four different advanced T2I models as targets, resulting in a final T2I model, named Edgen, which surpasses each of the target models in overall human preferences.
Strengths
**Originality.** The paper demonstrates the possibility of a training framework using synthetic data which leverages expert T2I models and VLMs to maximize data efficiency. This opens up an avenue of research into better use of expert models for informing the training procedure in T2I.
**Quality.** In certain areas, the authors perform thorough ablations to determine the best procedures, and to provide concrete numbers to compare against. For example, they evaluate the candidates for an expert VLM along several axes to justify their choice, and train baseline T2I models at several data scales to demonstrate the performance drop with decreasing scale.
**Clarity.** The high-level approach and framework are explained clearly.
**Significance.** The paper shows that existing T2I model performance can be replicated using only about 1% of the data needed with naive knowledge distillation.
Weaknesses
**Quality.** While most of the experimental details are explained thoroughly, the paper omits any description of the text data used. The text data is a crucial part of the proposed framework, since it is used to seed the training data, as well as for the mutation operation, presumably. There is also no mention of the source of the data used for evaluation. Along the lines of evaluation, the dataset used to compute FID score should also be mentioned.
**Clarity.** The paper fails to clarify many useful details earlier in the paper (abstract + introduction), leaving the overall understanding of the method vague until reaching Section 3. For example, the key components of expansion, deletion, and mutation are mentioned but not explained at all in the introduction.
Questions
1. What is the source of your training data? The training images are generated by the expert/advanced T2I models, but where do the input prompts come from?
2. Following Q1, how do you know the seed prompts for training (i.e. the initial training set) cover a diversity of prompts? While using LLMs to create variations of prompts can result in new combinations of objects or concepts, the resulting new prompts are unlikely to capture the same diversity of text present in natural datasets.
3. Where does the *mutation* operation source the new prompts from? Are they also generated by the same VLM, or taken from the same source as your initial training data? If they come from an existing dataset, what is the benefit of adding them midway through the training process instead of just starting with a initial training set?
4. To my understanding, LLaVA-NeXT (and many other current VLMs) are not trained on examples involving multiple images. Are the images for comparison concatenated or passed separately to LLaVA? Do you notice any statistical confounders, such as choosing the first image more often than the second?
5. The instructions for "expansion" to the VLM, to generate new variations on the prompt, specify only to replace nouns in the prompt. Why is this the strategy used to generate similar prompts? What happens if the T2I model is struggling to generate a specific noun, and the syntactic structure of the prompt is not the weak point? In that case, it feels like the new prompts would not be relevant for further training.
6. When selecting which VLM to use, you say "The output [of the VLM] is scored by 0 for wrong or 1 for correct by human raters". Why not ask the human raters to rank the generations themselves, and then compute how many examples the VLM and the humans agreed on? Having the human raters rate the VLM output seems like it could induce bias in the raters.
Limitations
The authors note that their method could result in a T2I model which inherits biases from both the teacher T2I model and the VLM used for evaluation. They also describe the potential positive and negative broader societal impacts of the research, including improving model training costs and potential for misuse.