[DDIMInverseScheduler] inf values at first iteration when set_alpha_to_one=True and prediction_type="sample"

[andreabosisio]

Describe the bug

I got inf values in

diffusers/src/diffusers/schedulers/scheduling_ddim_inverse.py

Line 347 in 560fb5f

pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)

because beta_prod_t is 0.0 at the first iteration, when timestep is < 0 and beta_prod_t is 0.0 (because alpha_prod_t is set to 1.0 when set_alpha_to_one=True).

diffusers/src/diffusers/schedulers/scheduling_ddim_inverse.py

Line 338 in 560fb5f

beta_prod_t = 1 - alpha_prod_t

Reproduction

import torch as th 
from diffusers import DDIMInverseScheduler

ddim_inverse_scheduler = DDIMInverseScheduler(
    num_train_timesteps=1000,
    prediction_type="sample",
    set_alpha_to_one=True
)
ddim_inverse_scheduler.set_timesteps(num_inference_steps=50)

with th.no_grad():
    pred = th.randn((1, 1, 2, 2, 2))
    samples = th.randn((1, 1, 2, 2, 2))
    t = 0
    samples = ddim_inverse_scheduler.step(pred, t, samples).prev_sample
    assert not th.isinf(samples).any(), "samples contain inf values"

Logs

System Info

• 🤗 Diffusers version: 0.32.2
• Platform: macOS-15.2-arm64-arm-64bit
• Running on Google Colab?: No
• Python version: 3.10.8
• PyTorch version (GPU?): 2.5.1 (False)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Huggingface_hub version: 0.29.1
• Transformers version: not installed
• Accelerate version: not installed
• PEFT version: not installed
• Bitsandbytes version: not installed
• Safetensors version: 0.5.2
• xFormers version: not installed
• Accelerator: Apple M1 Pro
• Using GPU in script?: No (Using MPS)
• Using distributed or parallel set-up in script?: No

Who can help?

@yiyixuxu @sayakpaul