Merge pull request #15325 from AUTOMATIC1111/sgm_uniform

Sgm uniform scheduler for SDXL-Lightning models

modules/sd_samplers_custom_schedulers.py

@@ -0,0 +1,12 @@
+import torch
+
+
+def sgm_uniform(n, sigma_min, sigma_max, inner_model, device):
+    start = inner_model.sigma_to_t(torch.tensor(sigma_max))
+    end = inner_model.sigma_to_t(torch.tensor(sigma_min))
+    sigs = [
+        inner_model.t_to_sigma(ts)
+        for ts in torch.linspace(start, end, n)[:-1]
+    ]
+    sigs += [0.0]
+    return torch.FloatTensor(sigs).to(device)

modules/sd_samplers_kdiffusion.py

@@ -3,6 +3,7 @@
 import k_diffusion.sampling
 from modules import sd_samplers_common, sd_samplers_extra, sd_samplers_cfg_denoiser
 from modules.sd_samplers_cfg_denoiser import CFGDenoiser  # noqa: F401
+from modules.sd_samplers_custom_schedulers import sgm_uniform
 from modules.script_callbacks import ExtraNoiseParams, extra_noise_callback
 
 from modules.shared import opts
@@ -62,7 +63,8 @@
     'Automatic': None,
     'karras': k_diffusion.sampling.get_sigmas_karras,
     'exponential': k_diffusion.sampling.get_sigmas_exponential,
-    'polyexponential': k_diffusion.sampling.get_sigmas_polyexponential
+    'polyexponential': k_diffusion.sampling.get_sigmas_polyexponential,
+    'sgm_uniform' : sgm_uniform,
 }
 
 
@@ -121,6 +123,11 @@ def get_sigmas(self, p, steps):
             if opts.k_sched_type != 'exponential' and opts.rho != 0 and opts.rho != default_rho:
                 sigmas_kwargs['rho'] = opts.rho
                 p.extra_generation_params["Schedule rho"] = opts.rho
+            if opts.k_sched_type == 'sgm_uniform':
+                # Ensure the "step" will be target step + 1
+                steps += 1 if not discard_next_to_last_sigma else 0
+                sigmas_kwargs['inner_model'] = self.model_wrap
+                sigmas_kwargs.pop('rho', None)
 
             sigmas = sigmas_func(n=steps, **sigmas_kwargs, device=shared.device)
         elif self.config is not None and self.config.options.get('scheduler', None) == 'karras':

modules/shared_options.py

@@ -368,7 +368,7 @@
     's_tmin':  OptionInfo(0.0, "sigma tmin",  gr.Slider, {"minimum": 0.0, "maximum": 10.0, "step": 0.01}, infotext='Sigma tmin').info('enable stochasticity; start value of the sigma range; only applies to Euler, Heun, and DPM2'),
     's_tmax':  OptionInfo(0.0, "sigma tmax",  gr.Slider, {"minimum": 0.0, "maximum": 999.0, "step": 0.01}, infotext='Sigma tmax').info("0 = inf; end value of the sigma range; only applies to Euler, Heun, and DPM2"),
     's_noise': OptionInfo(1.0, "sigma noise", gr.Slider, {"minimum": 0.0, "maximum": 1.1, "step": 0.001}, infotext='Sigma noise').info('amount of additional noise to counteract loss of detail during sampling'),
-    'k_sched_type':  OptionInfo("Automatic", "Scheduler type", gr.Dropdown, {"choices": ["Automatic", "karras", "exponential", "polyexponential"]}, infotext='Schedule type').info("lets you override the noise schedule for k-diffusion samplers; choosing Automatic disables the three parameters below"),
+    'k_sched_type':  OptionInfo("Automatic", "Scheduler type", gr.Dropdown, {"choices": ["Automatic", "karras", "exponential", "polyexponential", "sgm_uniform"]}, infotext='Schedule type').info("lets you override the noise schedule for k-diffusion samplers; choosing Automatic disables the three parameters below"),
     'sigma_min': OptionInfo(0.0, "sigma min", gr.Number, infotext='Schedule min sigma').info("0 = default (~0.03); minimum noise strength for k-diffusion noise scheduler"),
     'sigma_max': OptionInfo(0.0, "sigma max", gr.Number, infotext='Schedule max sigma').info("0 = default (~14.6); maximum noise strength for k-diffusion noise scheduler"),
     'rho':  OptionInfo(0.0, "rho", gr.Number, infotext='Schedule rho').info("0 = default (7 for karras, 1 for polyexponential); higher values result in a steeper noise schedule (decreases faster)"),