Image generation config

[tcm03]

Hi,

Thanks for your work and the publishing of resources.

I would like to ask that how do we customize the text-to-image inference process, involving:

• specifying the output resolution.
• specifying the number of inference steps.
• turning off flash-attn when running.

Also, I find the performance of the Dif-DTok version impressive. Have you published the code for this diffusion detokenizer version?

Thank you in advance.

[csuhan]

• Load different DToks. For example, change

  Tar/t2i_inference.py

  Line 90 in 47da8e1

  config.ar_path = hf_hub_download("csuhan/TA-Tok", "ar_dtok_lp_1024px.pth")

  from ar_dtok_lp_1024px.pth to ar_dtok_lp_512px.pth.
• It cannot be changed since AR-DTok generates the token one by one. But you can specify it in Dif-DTok.
• passing attention params to

  Tar/t2i_inference.py

  Line 37 in 47da8e1

  self.model = Qwen2ForCausalLM.from_pretrained(self.config.model_path, torch_dtype=self.config.dtype).to(self.device)

  
  add attn_implementation="sdpa" to from_pretrained

We now support Dif-DTok. Play it with: https://github.com/csuhan/Tar/blob/main/t2i_inference_sana.py