Open lwb2099 opened 11 months ago
Hi, thanks for reaching out! For OCR accuracy and Clip Score, we use the SimpleBench and CreativeBench. The glyph instructions are set as default without deliberate design (following the inference codes). We will release the benchmark testing codes this week. Sorry for the delay release.
While for the FID score testing, we use 10K samples which are not used for training in LAION-Glyph for evaluation. Hopefully we will release this eval set in a week or so.
The full prompts of two eval benchmarks, which are related to the OCR results shown in the paper, could be found in the text_prompts/paper folder.
are results tested without glyph instructions? also it better if you can provide eval dataset rather than prompts template for better reproduction