Scheduler selection in main UI

[AUTOMATIC1111]

Description

• adds scheduler selection to main UI
• reworks sampler selection UI section to be in a separate file (modules/processing_scripts/sampler.py)
• reworks the list of samplers to no longer include schedulers
• writes schedule type as Schedule type infotext field (this was previously a setting)
• schedule type of Automatic chooses the default appropriate scheduler for the sampler

I tested selecting different samplers and schedulers, changing rho and other params in settings, infotext pasting from old and current versions, all seems to work. Comments welcome.

Screenshots/videos:

[MoonRide303]

OMG, finally, hahah! Much better this way - thank you @AUTOMATIC1111! 😊

[MoonRide303]

Test results - prompt:

photography of a nice girl
Steps: 8, Sampler: Euler, Schedule type: SGM Uniform, CFG scale: 1, Seed: 0, Size: 896x1152, Model hash: 43f0501ac4, Model: sdxl_lightning_8step, Clip skip: 2, RNG: CPU, SGM noise multiplier: True, Version: v1.8.0-238-g31306ce6

and the output:

[gel-crabs]

On the topic of scheduler selection, the Restart sampler in sd_samplers_extra.py currently forces Karras via the get_sigmas_karras function; I have tested with exponential/polyexponential and can confirm it has no issue with either.

[w-e-w]

you forgot hires samples

[kaalibro]

UPD: Fixed, see PR below

Trying to use the "X/Y/Z plot" results in failure. On the dev branch works fine.

dev 57727e5

scheduler_selection 31306ce

Console logs

venv "D:\AI\SD\stable-diffusion-webui\venv\Scripts\Python.exe"
Python 3.10.11 (tags/v3.10.11:7d4cc5a, Apr  5 2023, 00:38:17) [MSC v.1929 64 bit (AMD64)]
Version: v1.8.0-238-g31306ce6
Commit hash: 31306ce6721e6f1ae323da64767da91897b3a1c2
Launching Web UI with arguments: --xformers --opt-sdp-attention --opt-channelslast --medvram --port 7860 --listen --ckpt-dir D:\AI\Data\SD\models\Stable-diffusion --vae-dir D:\AI\Data\SD\models\VAE --embeddings-dir D:\AI\Data\SD\embeddings --textual-inversion-templates-dir D:\AI\Data\SD\textual_inversion_templates --clip-models-path D:\AI\Data\SD\models\CLIP --gfpgan-models-path D:\AI\Data\SD\models\GFPGAN --codeformer-models-path D:\AI\Data\SD\models\Codeformer --esrgan-models-path D:\AI\Data\SD\models\ESRGAN --bsrgan-models-path D:\AI\Data\SD\models\BSRGAN --realesrgan-models-path D:\AI\Data\SD\models\RealESRGAN --ldsr-models-path D:\AI\Data\SD\models\LDSR --lora-dir D:\AI\Data\SD\models\Lora --hypernetwork-dir D:\AI\Data\SD\models\hypernetworks --scunet-models-path D:\AI\Data\SD\models\ScuNET --swinir-models-path D:\AI\Data\SD\models\SwinIR --localizations-dir D:\AI\Data\SD\localizations --styles-file D:\AI\Data\SD\styles.csv --disable-extra-extensions
*** "--disable-extra-extensions" arg was used, will only load built-in extensions ***
Loading weights [14953e03f0] from D:\AI\Data\SD\models\Stable-diffusion\SDXL⚡\lightningbloomxl_v10.safetensors
Running on local URL:  http://0.0.0.0:7860
Creating model from config: D:\AI\SD\stable-diffusion-webui\repositories\generative-models\configs\inference\sd_xl_base.yaml

To create a public link, set `share=True` in `launch()`.
Startup time: 18.0s (prepare environment: 3.4s, import torch: 4.4s, import gradio: 1.0s, setup paths: 2.2s, initialize shared: 0.8s, other imports: 0.5s, load scripts: 0.9s, reload hypernetworks: 0.2s, create ui: 0.4s, gradio launch: 4.2s).
Loading VAE weights from user metadata: D:\AI\Data\SD\models\VAE\sdxl_vae.safetensors
Applying attention optimization: xformers... done.
Model loaded in 13.6s (calculate hash: 0.2s, load weights from disk: 0.5s, create model: 0.6s, apply weights to model: 2.1s, apply channels_last: 0.2s, apply fp8: 7.3s, load VAE: 0.6s, load textual inversion embeddings: 0.3s, calculate empty prompt: 1.8s).
X/Y/Z plot will create 2 images on 1 2x1 grid. (Total steps to process: 8)
                                  generating image for xyz plot: AttributeError
Traceback (most recent call last):
  File "D:\AI\SD\stable-diffusion-webui\modules\processing.py", line 784, in process_images
    opts.set(k, v, is_api=True, run_callbacks=False)
  File "D:\AI\SD\stable-diffusion-webui\modules\options.py", line 151, in set
    option = self.data_labels[key]
KeyError: 'k_sched_type'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "D:\AI\SD\stable-diffusion-webui\scripts\xyz_grid.py", line 723, in cell
    res = process_images(pc)
  File "D:\AI\SD\stable-diffusion-webui\modules\processing.py", line 802, in process_images
    setattr(opts, k, v)
  File "D:\AI\SD\stable-diffusion-webui\modules\options.py", line 102, in __setattr__
    if info.do_not_save:
AttributeError: 'NoneType' object has no attribute 'do_not_save'

*** Error completing request
*** Arguments: ('task(8p6hou7rvuah98z)', <gradio.routes.Request object at 0x000001B9B668D630>, 'a girl sitting on a throne, head tilt, white hair, black tight dress, white fur trim, red fire behind, realistic, colorful', '', [], 1, 1, 1, 832, 1216, False, 0.2, 2, '4x-UltraSharp', 0, 0, 0, 'Use same checkpoint', 'Use same sampler', '', '', [], 3, 4, 'DPM++ SDE', 'SGM Uniform', 3168277607, False, -1, 0, 0, 0, False, 'CodeFormer', 0, False, 'second pass', False, '', 0.8, False, False, 'positive', 'comma', 20, False, False, 'start', '', 17, '', ['exponential', 'sgm_uniform'], 0, '', [], 0, '', [], True, False, False, False, False, False, False, 20, False) {}
    Traceback (most recent call last):
      File "D:\AI\SD\stable-diffusion-webui\modules\call_queue.py", line 57, in f
        res = list(func(*args, **kwargs))
      File "D:\AI\SD\stable-diffusion-webui\modules\call_queue.py", line 36, in f
        res = func(*args, **kwargs)
      File "D:\AI\SD\stable-diffusion-webui\modules\txt2img.py", line 105, in txt2img
        processed = modules.scripts.scripts_txt2img.run(p, *p.script_args)
      File "D:\AI\SD\stable-diffusion-webui\modules\scripts.py", line 768, in run
        processed = script.run(p, *script_args)
      File "D:\AI\SD\stable-diffusion-webui\scripts\xyz_grid.py", line 764, in run
        processed = draw_xyz_grid(
      File "D:\AI\SD\stable-diffusion-webui\scripts\xyz_grid.py", line 352, in draw_xyz_grid
        process_cell(x, y, z, ix, iy, iz)
      File "D:\AI\SD\stable-diffusion-webui\scripts\xyz_grid.py", line 296, in process_cell
        processed: Processed = cell(x, y, z, ix, iy, iz)
      File "D:\AI\SD\stable-diffusion-webui\scripts\xyz_grid.py", line 747, in cell
        grid_infotext[subgrid_index] = processing.create_infotext(pc, pc.all_prompts, pc.all_seeds, pc.all_subseeds)
      File "D:\AI\SD\stable-diffusion-webui\modules\processing.py", line 728, in create_infotext
        "Seed": p.all_seeds[0] if use_main_prompt else all_seeds[index],
    TypeError: 'NoneType' object is not subscriptable

---

[37OMKAR]

its too dev++ completed, can any one help us how can we get the old sampler back, in a work way easy way

[37OMKAR]

Description

• adds scheduler selection to main UI
• reworks sampler selection UI section to be in a separate file (modules/processing_scripts/sampler.py)
• reworks the list of samplers to no longer include schedulers
• writes schedule type as Schedule type infotext field (this was previously a setting)
• schedule type of Automatic chooses the default appropriate scheduler for the sampler

I tested selecting different samplers and schedulers, changing rho and other params in settings, infotext pasting from old and current versions, all seems to work. Comments welcome.

Screenshots/videos:

can we get the old sampler back ???

[twinkler83]

I never use schedulers like karras, exponential, etc because of their airbrush effect, especially on human skin.
Can we please have a "normal" option in "schedule type" selection?

[undeadx1]

I've been using the traditional approach of integrating the scheduler and sampler by implementing it myself. Can you please provide a unified UI like the old one?

[MoonRide303]

@undeadx1 Samplers and schedulers are two different things. The "traditional approach" presenting them as single entity was a design mistake. New samplers and new schedulers are being developed each year. If you have let's say 10 samplers and 5 schedulers then instead of having 2 short lists you'd have to maintain list of all 50 combinations. I understand some people were used to that, and that some software was built around how it previously worked, but this "traditional approach" should be just left in the history.

[drgrib]

How do you change this value in API calls?

[w-e-w]

How do you change this value in API calls?

you change the scheduler but changeing the scheduler payload, like just like any other value
for more read /docs and / or use API payload display wiki/api