[Schedulers] Add SGMUniform

[asomoza]

Thanks to @rollingcookies, we can see in this issue that this schedulers works great with the Hyper and probably also Lighting loras/unets.

It'd be fantastic if someone can contribute this scheduler to diffusers.

Please let me know if someone is willing to do this.

[ighoshsubho]

@asomoza can I pick up this after #9148 gets resolved if this is not picked till then

[asomoza]

@ighoshsubho thanks for picking this up, go ahead when you have time. Isn't that issue resolved though? We have the inpainting and img2img pipelines merged by now.

I found that you probably mean #9402.

[ighoshsubho]

@ighoshsubho thanks for picking this up, go ahead when you have time. Isn't that issue resolved though? We have the inpainting and img2img pipelines merged by now.

I found that you probably mean #9402.

my bad ya its #9402

[hlky]

SGM Uniform implementation in WebUI, same in Forge

In ByteDance/SDXL-Lightning demo:

pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", prediction_type="sample" if num_inference_steps==1 else "epsilon")

Reproduction and comparison:

from diffusers import EulerDiscreteScheduler
import torch

beta_start = 0.00085
beta_end = 0.012
num_train_timesteps = 1000
betas = (
    torch.linspace(
        beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32
    )
    ** 2
)
alphas = 1.0 - betas
alphas_cumprod = torch.cumprod(alphas, dim=0)
# not flipped, contrary to diffusers
sigmas = ((1 - alphas_cumprod) / alphas_cumprod) ** 0.5
log_sigmas = sigmas.log()
discard_next_to_last_sigma = False


def sgm_uniform(n: int, sigma_min: float, sigma_max: float):
    start = sigma_to_t(torch.tensor(sigma_max))
    end = sigma_to_t(torch.tensor(sigma_min))
    sigs = [t_to_sigma(ts) for ts in torch.linspace(start, end, n)[:-1]]
    sigs += [0.0]
    return torch.FloatTensor(sigs)


def sigma_to_t(sigma: torch.Tensor):
    log_sigma = sigma.log()
    dists = log_sigma - log_sigmas[:, None]
    low_idx = dists.ge(0).cumsum(dim=0).argmax(dim=0).clamp(max=log_sigmas.shape[0] - 2)
    high_idx = low_idx + 1
    low, high = log_sigmas[low_idx], log_sigmas[high_idx]
    w = (low - log_sigma) / (low - high)
    w = w.clamp(0, 1)
    t = (1 - w) * low_idx + w * high_idx
    return t.view(sigma.shape)


def t_to_sigma(t: torch.Tensor):
    t = t.float()
    low_idx, high_idx, w = t.floor().long(), t.ceil().long(), t.frac()
    log_sigma = (1 - w) * log_sigmas[low_idx] + w * log_sigmas[high_idx]
    return log_sigma.exp()


m_sigma_min, m_sigma_max = (sigmas[0].item(), sigmas[-1].item())

num_inference_steps = 4

scheduler: EulerDiscreteScheduler = EulerDiscreteScheduler.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    subfolder="scheduler",
    timestep_spacing="trailing",
    prediction_type="sample" if num_inference_steps == 1 else "epsilon",
)
scheduler.set_timesteps(num_inference_steps=num_inference_steps)

print(scheduler.sigmas)
# >>> tensor([14.6146,  4.0817,  1.6129,  0.6932,  0.0000])

sgm_uniform_sigmas = sgm_uniform(
    n=num_inference_steps + (1 if not discard_next_to_last_sigma else 0),
    sigma_min=m_sigma_min,
    sigma_max=m_sigma_max,
)

print(sgm_uniform_sigmas)
# >>> tensor([14.6146,  4.0861,  1.6156,  0.6952,  0.0000])

EulerDiscreteScheduler with trailing timesteps:

tensor([14.6146,  4.0817,  1.6129,  0.6932,  0.0000])

WebUI euler with sgm_uniform:

tensor([14.6146,  4.0861,  1.6156,  0.6952,  0.0000])

The tiny differences are likely due to numpy in Diffusers vs torch.

tl;dr SGM Uniform is already supported in Diffusers with timestep_spacing="trailing".

[asomoza]

@hlky thanks a lot for your findings, I haven't had the time to investigate the samplers + schedulers in the webuis so I'm really grateful you did this.

@rollingcookies does this resolve your issue then? you just need to change the timestep_spacing in the scheduler to have the equivalent of SGM Uniform.

cc: @yiyixuxu for awareness.

[hlky]

@asomoza If there's interest in supporting the other schedulers from the webuis I can look into existing equivalents and implement them as required. Also for this case the documentation could be updated.

[asomoza]

yes, if the schedulers are good and popular we can add them but we need to evaluate them first.

For example there was an issue opened about the DPM++ 3M SDE Karras. I did some tests and some online digging and found it not that used (at least for images I consider popular or good) and it also wasn't very good in my testings, since no one else answered with a good use case, we didn't add it.

But we're very open to suggestions and opinions about other schedulers/samplers.

I agree about the documentation, ccing @stevhliu to know what would be the best way add it.

[hlky]

I've done a preliminary popularity check with one of my datasets where the "Schedule type" field is present in metadata so images from A1111/Forge:

{
    "Karras": 329096,
    "Automatic": 148458,
    "Simple": 56595,
    "Exponential": 51861,
    "SGM Uniform": 21927,
    "Beta": 8815,
    "Polyexponential": 6807,
    "Align Your Steps": 5119,
    "Uniform": 4342,
    "Normal": 2772,
    "DDIM": 1583,
    "KL Optimal": 1500,
    "Turbo": 465,
    "Align Your Steps 32": 429,
    "Align Your Steps GITS": 169,
    "DPM++ 3M SDE Karras": 88,
}

Total: 640026

"Automatic" here means the default.

I'll check other datasets as the metadata in that one is limited although I imagine the distribution will be roughly the same.

[rollingcookies]

@asomoza Thank you so much for your efforts and help, I have forwarded this information to the folks at Invoke, hopefully this will be enough to add SGMUniform Scheduler to Invoke

[asomoza]

@rollingcookies no problem but this was all thanks to @hlky Wish you luck, I don't see why they wouldn't add it since it's a really small change.

[asomoza]

@hlky about these:

"Simple": 56595,
"Exponential": 51861,
"Beta": 8815,
"Polyexponential": 6807,

Any of these could work if we don't have them right now, I've seen good generations with these.