Prompt rewriting for text-to-image models just got a visual feedback loop.
Researchers have introduced FaithRewriter, a framework designed to close the gap between what a user types and what a model actually generates. The current approach in most prompt-enhancement tools is to polish text for fluency — fix grammar, add descriptors, make it sound more "professional." FaithRewriter takes a different path: it first generates an image from the original prompt using a multimodal large language model, then feeds that image back alongside the original text into a larger LLM to produce augmentations grounded in what the output actually looks like. That rewritten prompt is then distilled into a smaller model for practical deployment.
The insight here is that existing rewriters are essentially guessing. Without any visual reference, they over-infer missing details, which can send the final image further from what the user intended rather than closer. Anchoring the rewrite to an actual generated image gives the system something concrete to reason from — a meaningful shift from text-only refinement pipelines.
Experiments show FaithRewriter produces prompts that are both more faithful to user intent and more visually plausible than competing baselines, according to the paper.
Text-to-image prompt engineering has quietly become its own cottage industry, with users hand-crafting elaborate strings of keywords to coax models into compliance. FaithRewriter's approach automates a version of what experienced users already do manually — generate, evaluate, adjust — which is either a genuine quality-of-life improvement or a sign that the underlying models still need significant help understanding plain language.