do not work well for turbo model

[amberfay]

for xl model it works well but for turbo model ，the image is as use rong SAMPLER so i want to know where could i set SAMPLER

[haofanwang]

What turbo model are you using? From the community, dreamshaperXL_turboDpmppSDE is a good choice and can generate decent results, may you take a try?

[nosiu]

@amberfay there is an example in the readme in the LCM Lora section. List of schedulers https://huggingface.co/docs/diffusers/en/api/schedulers/overview

[Dolfik1]

What turbo model are you using? From the community, dreamshaperXL_turboDpmppSDE is a good choice and can generate decent results, may you take a try?

The dreamshaperXL model generally performs well, but for some reason, the images it generates appear somewhat unrealistic and contain visual artifacts. Here's the command I used:

python3 gradio_demo/app.py --enable_LCM false --pretrained_model_name_or_path models/dreamshaperXL_turboDpmppSDE.safetensors

Here are some examples of the output. As you can see, the images have some noticeable distortions:

Here's another example with the guidance scale set to 2:

In addition to the previous attempts, I also tried modifying the scheduler's parameters. Here's the command I used::

scheduler = diffusers.EulerDiscreteScheduler.from_config({
                "_class_name": "EulerDiscreteScheduler",
                "_diffusers_version": "0.22.0.dev0",
                "beta_end": 0.012,
                "beta_schedule": "scaled_linear",
                "beta_start": 0.00085,
                "clip_sample": False,
                "interpolation_type": "linear",
                "num_train_timesteps": 1000,
                "prediction_type": "epsilon",
                "sample_max_value": 1.0,
                "set_alpha_to_one": False,
                "skip_prk_steps": True,
                "steps_offset": 1,
                "timestep_spacing": "leading",
                "trained_betas": None,
                "use_karras_sigmas": True
                })

The results look better with these changes, but the images still lack detail: