search

adymaharana / storydalle

MIT License
330 stars 27 forks source link

Difficulties about reproduce results on FlintStones #7

Closed xichenpan closed 2 years ago

xichenpan commented 2 years ago

Hi, I am currently working on reproduce StoryDALLE on Flintstones I use the provided code and trained mega model for 50 epochs in a lr of 1e-5. the FID is 32 and my generation results are poor compared to the figure 4 in your paper. blow are my generated images, the right ones are grount truth. 2_s0415_f0 2_s0415_f1 2_s0415_f2 2_s0415_f3 I think the reason maybe I am using the inference code in your training codebase. Am I using the correct method to generate images?

pixels = model.sample_images(texts, src_images).cpu().transpose(1, -1).transpose(-1, -2)

I wil appreciate it if you can give me any advise~

FeiiYin commented 1 year ago

@Flash-321 Hi, Flash, I faced problems during the inference stage. from dalle.models import PrefixTuningDalle, StoryDalle, PromptDalle the code in infer_t2i.py However, I did not find any files in the folder of Dalle. Can you share the files with me or did I miss anything?

Thanks in advance.

xichenpan commented 1 year ago

Hi there, sorry I am not the repo owner and I do not know the detail. For my reproduction, I use mega-storydalle and finetune all params. It worked fine.