Alibaba open-sources PromptEcho for T2I rewards using frozen VLMs
阿里开源PromptEcho:用冻结多模态大模型为文生图训练提供高质量Reward
Alibaba open-sourced PromptEcho, which uses one frozen Qwen3-VL-32B forward pass to score T2I training rewards. It computes token-level cross-entropy for the original prompt under teacher forcing, then uses the negative value as a continuous reward. In 5,000 poster tests, text accuracy rose from 68% to 75%.
Why it matters: Alibaba open-sources PromptEcho, which uses a frozen Qwen3-VL-32B to generate continuous reward signals without training a separate scorer, lifting text accuracy from 68% to 75% on 5,000 posters.