Quesiton about training and (possibly) finetuning the model

[costrice]

Hi, thanks for your impressive work on image generation!

I anticipate using it as a generative backbone of my future work, so I am a bit curious about how many GPU resources are needed to train the model from scratch. More importantly, is it possible to finetune the model (like we commonly did on diffusion-based models like SD) using fewer GPU resources? Could you please provide some information? Thanks very much!

[JeyesHan]

Training from scratch is very expensive especially for the 1024x1024 resolution. We highly recommand you finetune Infinity. According to our full-params fine-tuning test with 4 GPUs, an iteration takes around 6s and 50GB vRAM per GPU, where global batch size=16, resolution=1024x1024. You can estimate the GPU resources for your fine-tuning task.

[wxxhaoshuai]

hi, I would like to know how many computing resources are required to train the 125M model from scratch and how many are required to finetune？

[JeyesHan]

@wxxhaoshuai
Apart from the model size, computing resources also depend on the data size and the target resolution. I think 16 gpus (A100 or H100) are enough for training 125M model from scratch under 256x256 resolution.

[wxxhaoshuai]

Will you release the smaller checkpoint？such as 125M or 1B.

[JeyesHan]

@wxxhaoshuai These small models are trained with a small subset of the whole dataset and used for demonstrating the scaling capability of Infinity. They are not full trained with abundant data, resolutions , and iterations. Therefore, we have no plan to release samller models for Infinity😭. We plan to release Infinity-20B.