Looking into this further, I am wondering if there was a specific reason to not follow the original CFG++ paper implementation on this. Left is the current approach (including my initial fix), and right would be the one following the CFG++ paper.
I did not upscale it or change anything else about it, so the differences are entirely current approach vs reference implementation. Difficult to say if one is better than the other, personally I say that reference has slightly better eyes and the ears come out a bit better. If more nature = better is pretty subjective though.
|
If aligning with the paper is desired, the change would be:

uncond_denoised = temp[0]
x_2 = (sigma_fn(s) / sigma_fn(t)) * x - (-h * r).expm1() * uncond_denoised
denoised_2 = model(x_2, sigma_fn(s) * s_in, **extra_args)
uncond_denoised_2 = temp[0]
x = denoised_2 - torch.exp(-h) * uncond_denoised_2 + (sigma_fn(t_next) / sigma_fn(t)) * x

Happy to add this to the PR or open a separate one if preferred.

Test Evidence Check

⚠️ Warning: Test Explanation Missing

If this PR modifies behavior that requires testing, a test explanation is required. PRs lacking applicable test explanations may not be reviewed until added. Please add test explanations to ensure code quality and prevent regressions.

⚠️ Warning: Visual Documentation Missing

If this PR changes user-facing behavior, visual proof (screen recording or screenshot) is required. PRs without applicable visual documentation may not be reviewed until provided.

You can add it by:

• GitHub: Drag & drop media directly into the PR description
• YouTube: Include a link to a short demo