Guard against zero std_cfg in rescale_noise_cfg (#13425)

[jbbqqf]

Fixes #13425.

Summary

rescale_noise_cfg divides std_text by std_cfg without any numerical guard:

std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)   # ← nan / inf when std_cfg == 0

If noise_cfg has zero variance (e.g. a constant or zero-valued guided prediction at the start of a schedule, or numerical edge cases), std_cfg == 0, the division produces nan/inf, and the silent corruption propagates through the rest of the diffusion process — no exception is raised.

Fix

Clamp std_cfg to torch.finfo(noise_cfg.dtype).eps before the divide. In the degenerate case the rescaling becomes a no-op (0 * <finite> == 0) instead of nan. In the non-degenerate case the clamp is below numerical noise, so the rescaling result is unchanged in practice.

std_cfg = std_cfg.clamp(min=torch.finfo(noise_cfg.dtype).eps)
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)

The fix was applied to the canonical copy in pipelines/stable_diffusion/pipeline_stable_diffusion.py and propagated to all 31 "# Copied from" duplicates via python utils/check_copies.py --fix_and_overwrite. python utils/check_copies.py (no flags) is green.

Reproduce BEFORE/AFTER yourself (copy-paste)

# --- BEFORE: origin/main, expect NaN ---
git fetch origin main && git checkout origin/main
python - <<'PY'
import torch
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import rescale_noise_cfg
noise_cfg = torch.zeros(1, 4, 8, 8)            # std == 0
noise_pred_text = torch.randn_like(noise_cfg)
result = rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=1.0)
print('any nan:', torch.isnan(result).any().item())  # Expected: True
PY

# --- AFTER: this branch ---
git fetch origin pull/<PR>/head:fix && git checkout fix
python - <<'PY'
import torch
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import rescale_noise_cfg
noise_cfg = torch.zeros(1, 4, 8, 8)
noise_pred_text = torch.randn_like(noise_cfg)
result = rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=1.0)
print('any nan:', torch.isnan(result).any().item())  # Expected: False
print('any inf:', torch.isinf(result).any().item())  # Expected: False
print('max abs:', result.abs().max().item())          # Expected: 0.0
PY

I confirmed both runs locally:

• origin/main → any nan: True
• this branch → any nan: False, any inf: False, max abs: 0.0

A second sanity check with a non-degenerate noise_cfg = torch.randn(...) produces an unchanged result on this branch (the clamp is below numerical noise).

Notes

• 32 files touched, all the change is the same 4-line guard before the divide.
• No public API change; pure numerical hardening.

---

🤖 Disclosure: Authored with assistance from Claude (Anthropic). Reproducer was run against both origin/main and this branch end-to-end.