[bug]: ERNIE-Image Heun scheduler refuses every partial denoise (img2img, inpaint, Canvas) #10003

Is there an existing issue for this problem?

• I have searched the existing issues

Install method

Manual (dev checkout)

Operating system

Windows

GPU vendor

Not GPU-specific (reproduced on CPU)

GPU model

No response

GPU VRAM

No response

Version number

main at cbc5ab0595

What happened

ERNIE-Image with the Heun scheduler refuses every partial denoise. Any img2img, inpaint or Canvas pass with a denoising strength below 1 fails before the first step:

ValueError: FlowMatchHeunDiscreteScheduler does not accept an explicit sigma schedule, so it cannot honor denoising_start/denoising_end. Use the euler or lcm scheduler for partial denoising.

The check is in invokeai/backend/ernie_image/denoise.py (around line 88). Euler and LCM pass the clipped sigmas through set_timesteps(sigmas=...). FlowMatchHeunDiscreteScheduler.set_timesteps takes only a step count, so the node refuses any schedule that does not run from 1 to 0.

Generate offers Heun for ERNIE-Image (scheduler_set="flow" in invokeai/backend/architectures/defs/ernie_image.py), so a user can select it and only finds out when the Canvas pass fails.

What you expected to happen

Heun starts at denoising_start and stops at denoising_end, as it does for FLUX.1 and Z-Image since #9851.

How to reproduce the problem

1. Select an ERNIE-Image model and the Heun scheduler.
2. Run img2img or a Canvas pass with denoising strength below 1.
3. The queue item fails with the ValueError above.

Without a model, the backend call reproduces it directly:

import torch
from invokeai.backend.ernie_image.denoise import denoise
from invokeai.backend.flux.schedulers import ERNIE_IMAGE_SCHEDULER_MAP

img = torch.zeros(1, 4, 4, 4)
denoise(
    model=None, img=img, text_bth=None, text_lens=None,
    timesteps=[0.6, 0.4, 0.2, 0.0], step_callback=lambda s: None, cfg_scale=[1.0] * 3,
    scheduler=ERNIE_IMAGE_SCHEDULER_MAP["heun"](), init_latents=img,
)

Additional context

FLUX.1 and FLUX.2 already drive Heun on a clipped schedule with set_heun_timesteps_from_sigmas in invokeai/backend/flux/schedulers.py. ERNIE-Image could use the same helper. One difference to account for: ERNIE passes raw timesteps and lets the scheduler apply its own shift inside set_timesteps, while the helper expects the sigmas the model runs on.

Discord username

No response