ancestral samplers are broken (on CPU)

[whitespace-rebel]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

ancestral samplers like euler_ancestral should work just fine

Actual Behavior

ancestral samplers like euler_ancestral produce a very noisy [preview] image. It looks like too much noise is being injected at step 2 for most of the affected samplers

Affected Samplers:
euler_ancestral; (starts at step 2 of 20)
ddpm; (starts at step 2 of 20)
dpm2_ancestral (starts at step 1.5 of 20)
dpmpp_2s_ancestral (starts at step 1.5 of 20)
res_multistep_ancestral; (starts at step 2 of 20)
er_sde; (starts at step 2 of 20)
sa_solver; (starts at step 5 of 20)
sa_solver_pece; (starts at step 4.5 of 20)
seeds_2; (starts between steps 1 and 2 of 20)
seeds_3; (starts between steps 1 and 2 of 20)

Steps to Reproduce

python main.py --cpu

Just use a basic SDXL workflow but change sampler to "euler_ancestral" + "simple" scheduler.

Debug Logs

Total VRAM 47285 MB, total RAM 47285 MB
pytorch version: 2.9.0.dev20250810+rocm6.4
Set vram state to: DISABLED
Device: cpu
Using sub quadratic optimization for attention, if you have memory or speed issues try using: --use-split-cross-attention
Python version: 3.13.5 (main, Jun 25 2025, 18:55:22) [GCC 14.2.0]
ComfyUI version: 0.3.59
ComfyUI frontend version: 1.27.3

Other

System:
Debian 13
AMD Ryzen 7 8700G

[chaObserv]

It appears similar here. In the CPU-only case, the noise from default_noise_sampler and the latent noise are both generated by the CPU and share the same seed. The noise sampler can produce noise identical to the initial one, causing noise dependency and breaking the results.

[whitespace-rebel]

@chaObserv , interesting, the difference I see is that I don't use any custom nodes, just default ComfyUI.

On the "same noise" issue: maybe something like a seed offset can be implemented, so that default_noise_sampler and latent noise are different.

[whitespace-rebel]

DiT/Flow based models seem to be affected too. Tested euler_ancestral with Wan2.2-T2V-14B (frame count=1, so it is basically text to image). Make sure to use at least 2 steps with the high noise model, and watch the noisy preview at step 2 on the High-noise KSampler.

###################################

I noticed another, possibly related, bug when using Wan2.2-T2V-14B (on CPU):
-set frame count to 1 (text to image)
-set UniPC sampler for both High/Low noise models
-use exactly 1 step for the High-noise model (I used the lightx2v Lightning LoRA)
-notice extremely noisy/pixelated preview on the Low-noise-KSampler, and final output.

I don't know if the UniPC bug also happens on GPU.

[whitespace-rebel]

I found a workaround for this problem:

1. install the Inspire-Pack custom nodes: https://github.com/ltdrdata/ComfyUI-Inspire-Pack
2. add the "Random noise (inspire)" to the workflow, and connect it to a "SamplerCustomAdvanced" node.
3. Set the right values:
   a) make sure noise_seed and internal_seed are NOT the same value (noise_seed != internal_seed)
   b) If you use the "DisableNoise" node (in multi stage workflows, like Wan2.2-High -> Wan2.2-Low), make sure internal_seed is NOT zero (internal_seed != 0)

Another custom node that works:
"Variation Noise 🍌" from https://github.com/laksjdjf/cgem156-ComfyUI

Here the variables are called base_seed instead of internal_seed,
and seed instead of noise_seed.

[comfyanonymous]

#11528