Skip to content

EDMEulerScheduler.set_timesteps fails with a list of sigmas, and custom sigmas are passed through the karras ramp #15266

Open

Description

@MaiAditya

Describe the bug

EDMEulerScheduler.set_timesteps is annotated sigmas: torch.Tensor | list[float] | None and documented as "Custom sigmas to use for the denoising process", but:

  1. A list raises AttributeError: 'list' object has no attribute 'to'. Only None, a single float and a tensor are handled (scheduling_edm_euler.py, the isinstance(sigmas, float) / sigmas.to(...) branch). retrieve_timesteps in the pipelines documents sigmas as list[float], so a list from a pipeline call hits this.
  2. A tensor is not used as the sigma schedule: it goes through _compute_karras_sigmas (or _compute_exponential_sigmas) as if it were the ramp in [0, 1]. Actual sigmas such as [80, 10, 1, 0.1] produce a meaningless schedule.

Reproduction

import torch
from diffusers import EDMEulerScheduler

sched = EDMEulerScheduler()  # sigma_schedule="karras" by default
try:
    sched.set_timesteps(sigmas=[80.0, 10.0, 1.0, 0.1])
except Exception as e:
    print("list sigmas ->", type(e).__name__, e)

custom = torch.tensor([80.0, 10.0, 1.0, 0.1])
sched.set_timesteps(sigmas=custom)
print("tensor sigmas in :", custom.tolist())
print("scheduler sigmas :", [round(x, 4) for x in sched.sigmas.tolist()])

Context

I was checking how the schedulers handle custom sigmas passed through retrieve_timesteps.

Logs

list sigmas -> AttributeError 'list' object has no attribute 'to'
tensor sigmas in : [80.0, 10.0, 1.0, 0.10000000149011612]
scheduler sigmas : [-263018948591616.0, -53744588.0, 0.002, 45.3137, 0.0]

System Info

diffusers 0.43.0 (same code on main 843f174), torch 2.15.0, Python 3.12.7, macOS.

Who can help?

@yiyixuxu

Which behaviour is intended for sigmas here: actual sigma values (used as given, like most other schedulers' sigmas=), or ramp positions in [0, 1] that the configured schedule maps to sigmas (which is how the default torch.linspace(0, 1, n) path works)? I'm happy to open a PR for whichever you prefer: accept lists in both cases, and either skip the ramp transform for custom sigmas or document that the values are ramp positions and validate the range.

AI disclosure: I used Claude Code (Anthropic) to scan the schedulers and draft the reproduction; I ran the reproduction myself and wrote this up after checking the code paths named above.

Activity

  1. sunhouy commented on Nov 15, 2026

    @sunhouy

    The list-of-sigmas crash and the karras ramp both come from set_timesteps running custom values through _compute_karras_sigmas — that transform only makes sense for the default torch.linspace(0, 1, n) ramp.

    Opened #15296 to use custom sigmas as given (lists included), and documented the expected order (largest to smallest) so it matches what the default schedule produces. With that, sigmas=[80.0, 9.723, 0.47, 0.002] and num_inference_steps=4 give the same schedule.

  2. MaiAditya commented on Nov 18, 2026

    @MaiAditya
    Author

    @sunhouy thanks, but I filed this and asked above which semantics sigmas should have (actual sigma values vs ramp positions), since the two lead to different fixes. Per CONTRIBUTING I'm waiting for a maintainer to pick one before opening a PR, so I'd rather leave that call to @yiyixuxu.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions