Merge branch 'master' into beta

Author: jn-jairo

comfy/k_diffusion/sampling.py

@@ -750,3 +750,61 @@ def sample_lcm(model, x, sigmas, extra_args=None, callback=None, disable=None, n
         if sigmas[i + 1] > 0:
             x += sigmas[i + 1] * noise_sampler(sigmas[i], sigmas[i + 1])
     return x
+
+
+
+@torch.no_grad()
+def sample_heunpp2(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1.):
+    # From MIT licensed: https://github.com/Carzit/sd-webui-samplers-scheduler/
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    s_end = sigmas[-1]
+    for i in trange(len(sigmas) - 1, disable=disable):
+        gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.
+        eps = torch.randn_like(x) * s_noise
+        sigma_hat = sigmas[i] * (gamma + 1)
+        if gamma > 0:
+            x = x + eps * (sigma_hat ** 2 - sigmas[i] ** 2) ** 0.5
+        denoised = model(x, sigma_hat * s_in, **extra_args)
+        d = to_d(x, sigma_hat, denoised)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
+        dt = sigmas[i + 1] - sigma_hat
+        if sigmas[i + 1] == s_end:
+            # Euler method
+            x = x + d * dt
+        elif sigmas[i + 2] == s_end:
+
+            # Heun's method
+            x_2 = x + d * dt
+            denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
+            d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
+
+            w = 2 * sigmas[0]
+            w2 = sigmas[i+1]/w
+            w1 = 1 - w2
+
+            d_prime = d * w1 + d_2 * w2
+
+
+            x = x + d_prime * dt
+
+        else:
+            # Heun++
+            x_2 = x + d * dt
+            denoised_2 = model(x_2, sigmas[i + 1] * s_in, **extra_args)
+            d_2 = to_d(x_2, sigmas[i + 1], denoised_2)
+            dt_2 = sigmas[i + 2] - sigmas[i + 1]
+
+            x_3 = x_2 + d_2 * dt_2
+            denoised_3 = model(x_3, sigmas[i + 2] * s_in, **extra_args)
+            d_3 = to_d(x_3, sigmas[i + 2], denoised_3)
+
+            w = 3 * sigmas[0]
+            w2 = sigmas[i + 1] / w
+            w3 = sigmas[i + 2] / w
+            w1 = 1 - w2 - w3
+
+            d_prime = w1 * d + w2 * d_2 + w3 * d_3
+            x = x + d_prime * dt
+    return x

comfy/ldm/modules/diffusionmodules/openaimodel.py

@@ -624,6 +624,11 @@ def forward(self, x, timesteps=None, context=None, y=None, control=None, transfo
             transformer_options["block"] = ("input", id)
             h = forward_timestep_embed(module, h, emb, context, transformer_options)
             h = apply_control(h, control, 'input')
+            if "input_block_patch" in transformer_patches:
+                patch = transformer_patches["input_block_patch"]
+                for p in patch:
+                    h = p(h, transformer_options)
+
             hs.append(h)
 
         transformer_options["block"] = ("middle", 0)

comfy/model_patcher.py

@@ -96,6 +96,9 @@ def set_model_attn1_output_patch(self, patch):
     def set_model_attn2_output_patch(self, patch):
         self.set_model_patch(patch, "attn2_output_patch")
 
+    def set_model_input_block_patch(self, patch):
+        self.set_model_patch(patch, "input_block_patch")
+
     def set_model_output_block_patch(self, patch):
         self.set_model_patch(patch, "output_block_patch")
 

comfy/samplers.py

@@ -518,46 +518,63 @@ class UNIPCBH2(Sampler):
     def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
         return uni_pc.sample_unipc(model_wrap, noise, latent_image, sigmas, max_denoise=self.max_denoise(model_wrap, sigmas), extra_args=extra_args, noise_mask=denoise_mask, callback=callback, variant='bh2', disable=disable_pbar)
 
-KSAMPLER_NAMES = ["euler", "euler_ancestral", "heun", "dpm_2", "dpm_2_ancestral",
+KSAMPLER_NAMES = ["euler", "euler_ancestral", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_sde", "dpmpp_sde_gpu",
                   "dpmpp_2m", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm"]
 
-def ksampler(sampler_name, extra_options={}, inpaint_options={}):
-    class KSAMPLER(Sampler):
-        def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
-            extra_args["denoise_mask"] = denoise_mask
-            model_k = KSamplerX0Inpaint(model_wrap)
-            model_k.latent_image = latent_image
-            if inpaint_options.get("random", False): #TODO: Should this be the default?
-                generator = torch.manual_seed(extra_args.get("seed", 41) + 1)
-                model_k.noise = torch.randn(noise.shape, generator=generator, device="cpu").to(noise.dtype).to(noise.device)
-            else:
-                model_k.noise = noise
+class KSAMPLER(Sampler):
+    def __init__(self, sampler_function, extra_options={}, inpaint_options={}):
+        self.sampler_function = sampler_function
+        self.extra_options = extra_options
+        self.inpaint_options = inpaint_options
 
-            if self.max_denoise(model_wrap, sigmas):
-                noise = noise * torch.sqrt(1.0 + sigmas[0] ** 2.0)
-            else:
-                noise = noise * sigmas[0]
+    def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
+        extra_args["denoise_mask"] = denoise_mask
+        model_k = KSamplerX0Inpaint(model_wrap)
+        model_k.latent_image = latent_image
+        if self.inpaint_options.get("random", False): #TODO: Should this be the default?
+            generator = torch.manual_seed(extra_args.get("seed", 41) + 1)
+            model_k.noise = torch.randn(noise.shape, generator=generator, device="cpu").to(noise.dtype).to(noise.device)
+        else:
+            model_k.noise = noise
 
-            k_callback = None
-            total_steps = len(sigmas) - 1
-            if callback is not None:
-                k_callback = lambda x: callback(x["i"], x["denoised"], x["x"], total_steps)
+        if self.max_denoise(model_wrap, sigmas):
+            noise = noise * torch.sqrt(1.0 + sigmas[0] ** 2.0)
+        else:
+            noise = noise * sigmas[0]
+
+        k_callback = None
+        total_steps = len(sigmas) - 1
+        if callback is not None:
+            k_callback = lambda x: callback(x["i"], x["denoised"], x["x"], total_steps)
+
+        if latent_image is not None:
+            noise += latent_image
 
+        samples = self.sampler_function(model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar, **self.extra_options)
+        return samples
+
+
+def ksampler(sampler_name, extra_options={}, inpaint_options={}):
+    if sampler_name == "dpm_fast":
+        def dpm_fast_function(model, noise, sigmas, extra_args, callback, disable):
             sigma_min = sigmas[-1]
             if sigma_min == 0:
                 sigma_min = sigmas[-2]
+            total_steps = len(sigmas) - 1
+            return k_diffusion_sampling.sample_dpm_fast(model, noise, sigma_min, sigmas[0], total_steps, extra_args=extra_args, callback=callback, disable=disable)
+        sampler_function = dpm_fast_function
+    elif sampler_name == "dpm_adaptive":
+        def dpm_adaptive_function(model, noise, sigmas, extra_args, callback, disable):
+            sigma_min = sigmas[-1]
+            if sigma_min == 0:
+                sigma_min = sigmas[-2]
+            return k_diffusion_sampling.sample_dpm_adaptive(model, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=callback, disable=disable)
+        sampler_function = dpm_adaptive_function
+    else:
+        sampler_function = getattr(k_diffusion_sampling, "sample_{}".format(sampler_name))
 
-            if latent_image is not None:
-                noise += latent_image
-            if sampler_name == "dpm_fast":
-                samples = k_diffusion_sampling.sample_dpm_fast(model_k, noise, sigma_min, sigmas[0], total_steps, extra_args=extra_args, callback=k_callback, disable=disable_pbar)
-            elif sampler_name == "dpm_adaptive":
-                samples = k_diffusion_sampling.sample_dpm_adaptive(model_k, noise, sigma_min, sigmas[0], extra_args=extra_args, callback=k_callback, disable=disable_pbar)
-            else:
-                samples = getattr(k_diffusion_sampling, "sample_{}".format(sampler_name))(model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar, **extra_options)
-            return samples
-    return KSAMPLER
+    return KSAMPLER(sampler_function, extra_options, inpaint_options)
 
 def wrap_model(model):
     model_denoise = CFGNoisePredictor(model)
@@ -618,11 +635,11 @@ def calculate_sigmas_scheduler(model, scheduler_name, steps):
         print("error invalid scheduler", self.scheduler)
     return sigmas
 
-def sampler_class(name):
+def sampler_object(name):
     if name == "uni_pc":
-        sampler = UNIPC
+        sampler = UNIPC()
     elif name == "uni_pc_bh2":
-        sampler = UNIPCBH2
+        sampler = UNIPCBH2()
     elif name == "ddim":
         sampler = ksampler("euler", inpaint_options={"random": True})
     else:
@@ -687,6 +704,6 @@ def sample(self, noise, positive, negative, cfg, latent_image=None, start_step=N
                 else:
                     return torch.zeros_like(noise)
 
-        sampler = sampler_class(self.sampler)
+        sampler = sampler_object(self.sampler)
 
-        return sample(self.model, noise, positive, negative, cfg, self.device, sampler(), sigmas, self.model_options, latent_image=latent_image, denoise_mask=denoise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)
+        return sample(self.model, noise, positive, negative, cfg, self.device, sampler, sigmas, self.model_options, latent_image=latent_image, denoise_mask=denoise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)

comfy/sd1_clip.py

@@ -173,9 +173,9 @@ def forward(self, tokens):
         if getattr(self.transformer, self.inner_name).final_layer_norm.weight.dtype != torch.float32:
             precision_scope = torch.autocast
         else:
-            precision_scope = lambda a, b: contextlib.nullcontext(a)
+            precision_scope = lambda a, dtype: contextlib.nullcontext(a)
 
-        with precision_scope(model_management.get_autocast_device(device), torch.float32):
+        with precision_scope(model_management.get_autocast_device(device), dtype=torch.float32):
             attention_mask = None
             if self.enable_attention_masks:
                 attention_mask = torch.zeros_like(tokens)

comfy/utils.py

@@ -333,13 +333,13 @@ def slerp(b1, b2, r):
         res[dot < 1e-5 - 1] = (b1 * (1.0-r) + b2 * r)[dot < 1e-5 - 1]
         return res
 
-    def generate_bilinear_data(length_old, length_new):
-        coords_1 = torch.arange(length_old).reshape((1,1,1,-1)).to(torch.float32)
+    def generate_bilinear_data(length_old, length_new, device):
+        coords_1 = torch.arange(length_old, dtype=torch.float32, device=device).reshape((1,1,1,-1))
         coords_1 = torch.nn.functional.interpolate(coords_1, size=(1, length_new), mode="bilinear")
         ratios = coords_1 - coords_1.floor()
         coords_1 = coords_1.to(torch.int64)
 
-        coords_2 = torch.arange(length_old).reshape((1,1,1,-1)).to(torch.float32) + 1
+        coords_2 = torch.arange(length_old, dtype=torch.float32, device=device).reshape((1,1,1,-1)) + 1
         coords_2[:,:,:,-1] -= 1
         coords_2 = torch.nn.functional.interpolate(coords_2, size=(1, length_new), mode="bilinear")
         coords_2 = coords_2.to(torch.int64)
@@ -349,7 +349,7 @@ def generate_bilinear_data(length_old, length_new):
     h_new, w_new = (height, width)
 
     #linear w
-    ratios, coords_1, coords_2 = generate_bilinear_data(w, w_new)
+    ratios, coords_1, coords_2 = generate_bilinear_data(w, w_new, samples.device)
     coords_1 = coords_1.expand((n, c, h, -1))
     coords_2 = coords_2.expand((n, c, h, -1))
     ratios = ratios.expand((n, 1, h, -1))
@@ -362,7 +362,7 @@ def generate_bilinear_data(length_old, length_new):
     result = result.reshape(n, h, w_new, c).movedim(-1, 1)
 
     #linear h
-    ratios, coords_1, coords_2 = generate_bilinear_data(h, h_new)
+    ratios, coords_1, coords_2 = generate_bilinear_data(h, h_new, samples.device)
     coords_1 = coords_1.reshape((1,1,-1,1)).expand((n, c, -1, w_new))
     coords_2 = coords_2.reshape((1,1,-1,1)).expand((n, c, -1, w_new))
     ratios = ratios.reshape((1,1,-1,1)).expand((n, 1, -1, w_new))

comfy_extras/nodes_custom_sampler.py

@@ -149,7 +149,7 @@ def INPUT_TYPES(s):
     FUNCTION = "get_sampler"
 
     def get_sampler(self, sampler_name):
-        sampler = comfy.samplers.sampler_class(sampler_name)()
+        sampler = comfy.samplers.sampler_object(sampler_name)
         return (sampler, )
 
 class SamplerDPMPP_2M_SDE:
@@ -172,7 +172,7 @@ def get_sampler(self, solver_type, eta, s_noise, noise_device):
             sampler_name = "dpmpp_2m_sde"
         else:
             sampler_name = "dpmpp_2m_sde_gpu"
-        sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "solver_type": solver_type})()
+        sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "solver_type": solver_type})
         return (sampler, )
 
 
@@ -196,7 +196,7 @@ def get_sampler(self, eta, s_noise, r, noise_device):
             sampler_name = "dpmpp_sde"
         else:
             sampler_name = "dpmpp_sde_gpu"
-        sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "r": r})()
+        sampler = comfy.samplers.ksampler(sampler_name, {"eta": eta, "s_noise": s_noise, "r": r})
         return (sampler, )
 
 class SamplerCustom: