training example for instruct pix2pix doesn't zero out embeds about diffusers

[sayakpaul]

#6552

[sayakpaul]

Thanks for bringing this up. Possible for you to show a comparison between what happens when you zero out like the way you mentioned compared to the existing approach?

I've checked the values of the embeds, and classifier-free guidance at inference time definitely makes use of the zero embed and not just "", which end up producing very different results.

The SD IP2P pipeline uses "", though when negative prompt is not provided:

diffusers/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py

Line 558 in ec9e881

uncond_tokens = [""] * batch_size

However, it makes use to zeros_like for the unconditional image embeddings:

diffusers/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py

Line 869 in ec9e881

uncond_image_latents = torch.zeros_like(image_latents)

[bghira]

the base model was trained using it, so i figured aligning with the base model's training and inference has better results.

from my own tests, i can now reduce the step count required when running the default config on the SDXL pipelines, eg. force zeroes is set to True

I also have much better learning. this model started from ptx0/terminus-xl-velocity-v1 and it was unable to spell.

1000 steps of tuning later:

the base model was trained using "" and it never really ends up with better CFG performance... but now it does!

[bghira]

see the base SDXL pipeline:

        # get unconditional embeddings for classifier free guidance
        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)

and the config:

{
   "_class_name": "StableDiffusionXLPipeline",
   "_diffusers_version": "0.19.0.dev0",
   "force_zeros_for_empty_prompt": true
}

[sayakpaul]

Ah makes a ton of sense. Do you want to take a step at opening a PR to fix this?

[bghira]

can i also open the pull request for all of the other training examples, to add general dropout capabilities to them?

[sayakpaul]

We can open that up for the community. This way everyone gets to participate.

[bghira]

like the ticket for updating the fp16 error? #6231