speedup, second order correction, intensity control

Author: arrmansa

scripts/img2imgalt.py

@@ -13,7 +13,7 @@
 
 # Debugging notes - the original method apply_model is being called for sd1.5 is in modules.sd_hijack_utils and is ldm.models.diffusion.ddpm.LatentDiffusion
 # For sdxl - OpenAIWrapper will be called, which will call the underlying diffusion_model
-
+# When controlnet is enabled, the underlying model is not available to use, therefore we skip
 
 def find_noise_for_image(p, cond, uncond, cfg_scale, steps):
     x = p.init_latent
@@ -78,11 +78,11 @@ def find_noise_for_image(p, cond, uncond, cfg_scale, steps):
     return x / x.std()
 
 
-Cached = namedtuple("Cached", ["noise", "cfg_scale", "steps", "latent", "original_prompt", "original_negative_prompt", "sigma_adjustment"])
+Cached = namedtuple("Cached", ["noise", "cfg_scale", "steps", "latent", "original_prompt", "original_negative_prompt", "sigma_adjustment", "second_order_correction", "noise_sigma_intensity"])
 
 
 # Based on changes suggested by briansemrau in https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/736
-def find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg_scale, steps):
+def find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg_scale, steps, correction_factor, sigma_intensity):
     x = p.init_latent
 
     s_in = x.new_ones([x.shape[0]])
@@ -98,11 +98,7 @@ def find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg_scale, steps):
 
     for i in trange(1, len(sigmas)):
         shared.state.sampling_step += 1
-
-        x_in = torch.cat([x] * 2)
         sigma_in = torch.cat([sigmas[i - 1] * s_in] * 2)
-
-
         if shared.sd_model.is_sdxl:
             cond_tensor = cond['crossattn']
             uncond_tensor = uncond['crossattn']
@@ -113,46 +109,69 @@ def find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg_scale, steps):
         image_conditioning = torch.cat([p.image_conditioning] * 2)
         cond_in = {"c_concat": [image_conditioning], "c_crossattn": [cond_in]}
 
-        c_out, c_in = [K.utils.append_dims(k, x_in.ndim) for k in dnw.get_scalings(sigma_in)[skip:]]
-
         if i == 1:
             t = dnw.sigma_to_t(torch.cat([sigmas[i] * s_in] * 2))
+            dt = (sigmas[i] - sigmas[i - 1]) / (2 * sigmas[i])
         else:
             t = dnw.sigma_to_t(sigma_in)
+            dt = (sigmas[i] - sigmas[i - 1]) / sigmas[i - 1]
+
+        noise = noise_from_model(x, t, dt, sigma_in, cond_in, cfg_scale, dnw, skip)
+
+        if correction_factor > 0:
+            recalculated_noise = noise_from_model(x + noise, t, dt, sigma_in, cond_in, cfg_scale, dnw, skip)
+            noise = recalculated_noise * correction_factor + noise * (1 - correction_factor)
+
+        x += noise
+
+        sd_samplers_common.store_latent(x)
+
+        # This shouldn't be necessary, but solved some VRAM issues
+        #del x_in, sigma_in, cond_in, c_out, c_in, t
+        #del eps, denoised_uncond, denoised_cond, denoised, dt
+
+    shared.state.nextjob()
+
+    return x / (x.std()*(1 - sigma_intensity) + sigmas[-1]*sigma_intensity)
 
+def noise_from_model(x, t, dt, sigma_in, cond_in, cfg_scale, dnw, skip):
 
+    if cfg_scale == 1:  # Case where denoised_uncond should not be calculated - 50% speedup, also good for sdxl in experiments
+        x_in = x
+        sigma_in = sigma_in[1:2]
+        c_out, c_in = [K.utils.append_dims(k, x_in.ndim) for k in dnw.get_scalings(sigma_in)[skip:]]
+        cond_in = {"c_concat":[cond_in["c_concat"][0][1:2]], "c_crossattn": [cond_in["c_crossattn"][0][1:2]]}
         if shared.sd_model.is_sdxl:
             num_classes_hack = shared.sd_model.model.diffusion_model.num_classes
             shared.sd_model.model.diffusion_model.num_classes = None
             try:
-                eps = shared.sd_model.model(x_in * c_in, t, {"crossattn": cond_in["c_crossattn"][0]} )
+                eps = shared.sd_model.model(x_in * c_in, t[1:2], {"crossattn": cond_in["c_crossattn"][0]})
             finally:
                 shared.sd_model.model.diffusion_model.num_classes = num_classes_hack
         else:
-            eps = shared.sd_model.apply_model(x_in * c_in, t, cond=cond_in)
+            eps = shared.sd_model.apply_model(x_in * c_in, t[1:2], cond=cond_in)
 
-        denoised_uncond, denoised_cond = (x_in + eps * c_out).chunk(2)
+        return -eps * c_out* dt
+    else :
+        x_in = torch.cat([x] * 2)
 
-        denoised = denoised_uncond + (denoised_cond - denoised_uncond) * cfg_scale
+        c_out, c_in = [K.utils.append_dims(k, x_in.ndim) for k in dnw.get_scalings(sigma_in)[skip:]]
 
-        if i == 1:
-            d = (x - denoised) / (2 * sigmas[i])
+        if shared.sd_model.is_sdxl:
+            num_classes_hack = shared.sd_model.model.diffusion_model.num_classes
+            shared.sd_model.model.diffusion_model.num_classes = None
+            try:
+                eps = shared.sd_model.model(x_in * c_in, t, {"crossattn": cond_in["c_crossattn"][0]} )
+            finally:
+                shared.sd_model.model.diffusion_model.num_classes = num_classes_hack
         else:
-            d = (x - denoised) / sigmas[i - 1]
-
-        dt = sigmas[i] - sigmas[i - 1]
-        x = x + d * dt
-
-        sd_samplers_common.store_latent(x)
-
-        # This shouldn't be necessary, but solved some VRAM issues
-        del x_in, sigma_in, cond_in, c_out, c_in, t,
-        del eps, denoised_uncond, denoised_cond, denoised, d, dt
+            eps = shared.sd_model.apply_model(x_in * c_in, t, cond=cond_in)
 
-    shared.state.nextjob()
+        denoised_uncond, denoised_cond = (eps * c_out).chunk(2)
 
-    return x / sigmas[-1]
+        denoised = denoised_uncond + (denoised_cond - denoised_uncond) * cfg_scale
 
+        return -denoised * dt
 
 class Script(scripts.Script):
     def __init__(self):
@@ -183,17 +202,20 @@ def ui(self, is_img2img):
         cfg = gr.Slider(label="Decode CFG scale", minimum=0.0, maximum=15.0, step=0.1, value=1.0, elem_id=self.elem_id("cfg"))
         randomness = gr.Slider(label="Randomness", minimum=0.0, maximum=1.0, step=0.01, value=0.0, elem_id=self.elem_id("randomness"))
         sigma_adjustment = gr.Checkbox(label="Sigma adjustment for finding noise for image", value=False, elem_id=self.elem_id("sigma_adjustment"))
+        second_order_correction = gr.Slider(label="Correct noise by running model again", minimum=0.0, maximum=1.0, step=0.01, value=0.5, elem_id=self.elem_id("second_order_correction"))
+        noise_sigma_intensity = gr.Slider(label="Weight scaling std vs sigma based", minimum=-1.0, maximum=2.0, step=0.01, value=0.5, elem_id=self.elem_id("noise_sigma_intensity"))
 
         return [
             info,
             override_sampler,
             override_prompt, original_prompt, original_negative_prompt,
             override_steps, st,
             override_strength,
-            cfg, randomness, sigma_adjustment,
+            cfg, randomness, sigma_adjustment, second_order_correction,
+            noise_sigma_intensity
         ]
 
-    def run(self, p, _, override_sampler, override_prompt, original_prompt, original_negative_prompt, override_steps, st, override_strength, cfg, randomness, sigma_adjustment):
+    def run(self, p, _, override_sampler, override_prompt, original_prompt, original_negative_prompt, override_steps, st, override_strength, cfg, randomness, sigma_adjustment, second_order_correction, noise_sigma_intensity):
         # Override
         if override_sampler:
             p.sampler_name = "Euler"
@@ -211,7 +233,9 @@ def sample_extra(conditioning, unconditional_conditioning, seeds, subseeds, subs
             same_params = self.cache is not None and self.cache.cfg_scale == cfg and self.cache.steps == st \
                                 and self.cache.original_prompt == original_prompt \
                                 and self.cache.original_negative_prompt == original_negative_prompt \
-                                and self.cache.sigma_adjustment == sigma_adjustment
+                                and self.cache.sigma_adjustment == sigma_adjustment \
+                                and self.cache.second_order_correction == second_order_correction \
+                                and self.cache.noise_sigma_intensity == noise_sigma_intensity
             same_everything = same_params and self.cache.latent.shape == lat.shape and np.abs(self.cache.latent-lat).sum() < 100
 
             rand_noise = processing.create_random_tensors(p.init_latent.shape[1:], seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength, seed_resize_from_h=p.seed_resize_from_h, seed_resize_from_w=p.seed_resize_from_w, p=p)
@@ -231,10 +255,10 @@ def sample_extra(conditioning, unconditional_conditioning, seeds, subseeds, subs
                 cond = p.sd_model.get_learned_conditioning(p.batch_size * [original_prompt])
                 uncond = p.sd_model.get_learned_conditioning(p.batch_size * [original_negative_prompt])
                 if sigma_adjustment:
-                    rec_noise = find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg, st)
+                    rec_noise = find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg, st, second_order_correction, noise_sigma_intensity)
                 else:
                     rec_noise = find_noise_for_image(p, cond, uncond, cfg, st)
-                self.cache = Cached(rec_noise, cfg, st, lat, original_prompt, original_negative_prompt, sigma_adjustment)
+                self.cache = Cached(rec_noise, cfg, st, lat, original_prompt, original_negative_prompt, sigma_adjustment, second_order_correction, noise_sigma_intensity)
 
             combined_noise = ((1 - randomness) * rec_noise + randomness * rand_noise) / ((randomness**2 + (1-randomness)**2) ** 0.5)