Merge pull request #12818 from catboxanon/sgm

Add option to align with sgm repo's sampling implementation

modules/sd_samplers_kdiffusion.py

@@ -144,7 +144,13 @@ def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning,
         sigmas = self.get_sigmas(p, steps)
         sigma_sched = sigmas[steps - t_enc - 1:]
 
-        xi = x + noise * sigma_sched[0]
+        if opts.sgm_noise_multiplier:
+            p.extra_generation_params["SGM noise multiplier"] = True
+            noise_multiplier = torch.sqrt(1.0 + sigma_sched[0] ** 2.0)
+        else:
+            noise_multiplier = sigma_sched[0]
+
+        xi = x + noise * noise_multiplier
 
         if opts.img2img_extra_noise > 0:
             p.extra_generation_params["Extra noise"] = opts.img2img_extra_noise
@@ -197,7 +203,11 @@ def sample(self, p, x, conditioning, unconditional_conditioning, steps=None, ima
 
         sigmas = self.get_sigmas(p, steps)
 
-        x = x * sigmas[0]
+        if opts.sgm_noise_multiplier:
+            p.extra_generation_params["SGM noise multiplier"] = True
+            x = x * torch.sqrt(1.0 + sigmas[0] ** 2.0)
+        else:
+            x = x * sigmas[0]
 
         extra_params_kwargs = self.initialize(p)
         parameters = inspect.signature(self.func).parameters

modules/shared_options.py

@@ -309,6 +309,7 @@
     '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)"),
     'eta_noise_seed_delta': OptionInfo(0, "Eta noise seed delta", gr.Number, {"precision": 0}, infotext='ENSD').info("ENSD; does not improve anything, just produces different results for ancestral samplers - only useful for reproducing images"),
     'always_discard_next_to_last_sigma': OptionInfo(False, "Always discard next-to-last sigma", infotext='Discard penultimate sigma').link("PR", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/6044"),
+    'sgm_noise_multiplier': OptionInfo(False, "SGM noise multiplier", infotext='SGM noise multplier').link("PR", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/12818").info("Match initial noise to official SDXL implementation - only useful for reproducing images"),
     'uni_pc_variant': OptionInfo("bh1", "UniPC variant", gr.Radio, {"choices": ["bh1", "bh2", "vary_coeff"]}, infotext='UniPC variant'),
     'uni_pc_skip_type': OptionInfo("time_uniform", "UniPC skip type", gr.Radio, {"choices": ["time_uniform", "time_quadratic", "logSNR"]}, infotext='UniPC skip type'),
     'uni_pc_order': OptionInfo(3, "UniPC order", gr.Slider, {"minimum": 1, "maximum": 50, "step": 1}, infotext='UniPC order').info("must be < sampling steps"),

scripts/xyz_grid.py

@@ -265,6 +265,7 @@ def __init__(self, *args, **kwargs):
     AxisOption("Token merging ratio", float, apply_override('token_merging_ratio')),
     AxisOption("Token merging ratio high-res", float, apply_override('token_merging_ratio_hr')),
     AxisOption("Always discard next-to-last sigma", str, apply_override('always_discard_next_to_last_sigma', boolean=True), choices=boolean_choice(reverse=True)),
+    AxisOption("SGM noise multiplier", str, apply_override('sgm_noise_multiplier', boolean=True), choices=boolean_choice(reverse=True)),
     AxisOption("Refiner checkpoint", str, apply_field('refiner_checkpoint'), format_value=format_remove_path, confirm=confirm_checkpoints_or_none, cost=1.0, choices=lambda: ['None'] + sorted(sd_models.checkpoints_list, key=str.casefold)),
     AxisOption("Refiner switch at", float, apply_field('refiner_switch_at')),
     AxisOption("RNG source", str, apply_override("randn_source"), choices=lambda: ["GPU", "CPU", "NV"]),