Remove Extreme Speed Preset (LCM) - Yes or No?

[mashb1t]

Pros:

• prevent users from accidentally generating images in lower quality as LCM is quite outdated
• reduce legacy code and promote more recent and higher quality methods like Hyper-SD

Cons:

• users can't regenerate images from image log anymore, which have been generated with Extreme Speed

What do you think? Happy to hear your opinion.

[mashb1t]

Alternatives:

• replace performance option group with select element (dropdown)
• use different performance with same label "Extreme Speed" (might lead to confusion, see #3090)

[MetatronR]

I would add a specific component for the performane-optimised options.
It would be selectable as a radio option, but also having a dropdown for selecting the desired method.

The best algorithm/model would be selected by default and also having the '(recommended)' suffix.
Other options could have suffixes like: fast, fastest, etc.

[wisdombox44]

I tried Hyper SD 12 steps Lora i got good results. Hyper SD 12 can be an option too.

[HelloLaohe]

I initially hoped to replace it with PCM, but even after 16 steps, the PCM still didn't achieve the desired results. After further testing, I felt that the Hyper-SD 12-step method, tested various sampling techniques, and provided comparisons. I hope to see better results by switching LCM to the Hyper-SD 12-step method and changing Lightning to 8 steps.

       print("Enter Hyper-SD(12) mode.")
        progressbar(async_task, 1, "Downloading Hyper-SD(12) components ...")
        performance_loras += [(modules.config.downloading_sdxl_lcm_lora(), 1.0)]
        if refiner_model_name != "None":
          print(f"Refiner disabled in Hyper-SD(12)  mode.")
        refiner_model_name = "None"
        sampler_name = "euler_ancestral"
        scheduler_name = "ddim_uniform"
        sharpness = 1.0
        guidance_scale = 3.0
        adaptive_cfg = 8.0
        refiner_switch = 2.0
        adm_scaler_positive = 1.0
        adm_scaler_negative = 1.0
        adm_scaler_end = 0.0

[HelloLaohe]

The following are the images generated by using Hyper-SD 12-step method for reference.：

[wisdombox44]

That is great for reference.
Lightning 8 step results look nice too. Sharp, vivid, clear

[HelloLaohe]

I hope to be able to add hpysd-sd 12 steps soon. Lightning should switch to 8 steps, otherwise, it's very practical to manually modify it every day's update.

[maolei133]

@HelloLaohe 你好，你的页面中那些其他功能是安装了那些插件？能说说吗？特别是中文显示插件。谢谢了

[jaimeastorga2000]

Remove LCM, change the Lightning LoRA from 4-step to 8-step.

Fooocus is supposed to be easy to use; too many options in the UI is confusing. Maybe hide it behind an advanced/debug tab.

Speaking of, do we really need both Lightning and Hyper-SD?

[HelloLaohe]

1, not everyone's graphics card is powerful, so it is necessary to have a way to generate images quickly. I am using a 3080ti, and some people around me are using a 2060 graphics card.
2, the modified hyper-sd (12 steps) is my default setting, and it also has very good effects. I hope to share it with more friends. Of course, I am also open to adding it to the advanced settings, and I can make the adjustments myself (modify the code a bit), but it's a pity that more people are not using it.

Here are my test images, using different content as much as possible.

[HelloLaohe]

----------------Modify the async_worker.py file----------------

        refiner_model_name = "None"
        sampler_name = "euler_ancestral"
        scheduler_name = "sgm_uniform"
        sharpness = 1.2
        guidance_scale = 3.0
        adaptive_cfg = 7.5
        refiner_switch = 2.0
        adm_scaler_positive = 1.0
        adm_scaler_negative = 1.0
        adm_scaler_end = 0.0

----------------Modify the flags.py file----------------

class PerformanceLoRA(Enum):
    QUALITY = None
    SPEED = None
    EXTREME_SPEED = "sdxl_lcm_lora.safetensors"   # (The model name remains the same, but the actual file is the hyper-sd 12-step version.) 
    LIGHTNING = "sdxl_lightning_4step_lora.safetensors"
    HYPER_SD = "sdxl_hyper_sd_4step_lora.safetensors"

class Steps(IntEnum):
    QUALITY = 60
    SPEED = 30
    EXTREME_SPEED = 12
    LIGHTNING = 8
    HYPER_SD = 4

[Zsural]

Hello!Can you tell me where to download hyper-sd 12-step version?I didn't find downloadlinks in modules.config.py,Thanks alot!

[mashb1t]

@Zsural here you go: https://huggingface.co/ByteDance/Hyper-SD/blob/main/Hyper-SDXL-12steps-CFG-lora.safetensors
Keep in mind that this is the CFG version and you have to adjust the guidance scale and weight accordingly

[Zsural]

@mashb1t Thank you!

[steppres]

This could be a chance to make some more substantial changes to the performance settings as whole. At the moment for a novice user the names for each setting are somewhat misleading confusing. Maybe we could rename the settings to something like this:

"Quality", "Normal", "Fast", "Very Fast", "Ultra Fast"

These would correspond to 60-step, 30-step, 12-step (Hyper-SD), 8-step (Lightning) and 4-step (Hyper-SD).

It could also allow for the legacy LCM to to still be regenerated from old image logs, just not displayed in the UI. Just an idea.