Merge a7e8008fdc into 6592bffc60

* Add phi_2 solver type to seeds_2

* Add sampler node of seeds_2

comfy/k_diffusion/sampling.py

@@ -1557,10 +1557,13 @@ def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None
 
 
 @torch.no_grad()
-def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5):
+def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5, solver_type="phi_1"):
     """SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VP Data Prediction) stage 2.
     arXiv: https://arxiv.org/abs/2305.14267 (NeurIPS 2023)
     """
+    if solver_type not in {"phi_1", "phi_2"}:
+        raise ValueError("solver_type must be 'phi_1' or 'phi_2'")
+
     extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
@@ -1600,8 +1603,14 @@ def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=Non
         denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
 
         # Step 2
-        denoised_d = torch.lerp(denoised, denoised_2, fac)
-        x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d
+        if solver_type == "phi_1":
+            denoised_d = torch.lerp(denoised, denoised_2, fac)
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * ei_h_phi_1(-h_eta) * denoised_d
+        elif solver_type == "phi_2":
+            b2 = ei_h_phi_2(-h_eta) / r
+            b1 = ei_h_phi_1(-h_eta) - b2
+            x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x - alpha_t * (b1 * denoised + b2 * denoised_2)
+
         if inject_noise:
             segment_factor = (r - 1) * h * eta
             sde_noise = sde_noise * segment_factor.exp()

comfy/ops.py

@@ -592,7 +592,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                     quant_conf = {"format": self.quant_format}
                     if self._full_precision_mm:
                         quant_conf["full_precision_matrix_mult"] = True
-                    sd["{}comfy_quant".format(prefix)] = torch.frombuffer(json.dumps(quant_conf).encode('utf-8'), dtype=torch.uint8)
+                    sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8)
                 return sd
 
             def _forward(self, input, weight, bias):

comfy/utils.py

@@ -1262,6 +1262,6 @@ def convert_old_quants(state_dict, model_prefix="", metadata={}):
     if quant_metadata is not None:
         layers = quant_metadata["layers"]
         for k, v in layers.items():
-            state_dict["{}.comfy_quant".format(k)] = torch.frombuffer(json.dumps(v).encode('utf-8'), dtype=torch.uint8)
+            state_dict["{}.comfy_quant".format(k)] = torch.tensor(list(json.dumps(v).encode('utf-8')), dtype=torch.uint8)
 
     return state_dict, metadata

comfy_extras/nodes_custom_sampler.py

@@ -659,6 +659,31 @@ class SamplerSASolver(io.ComfyNode):
     get_sampler = execute
 
 
+class SamplerSEEDS2(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="SamplerSEEDS2",
+            category="sampling/custom_sampling/samplers",
+            inputs=[
+                io.Combo.Input("solver_type", options=["phi_1", "phi_2"]),
+                io.Float.Input("eta", default=1.0, min=0.0, max=100.0, step=0.01, round=False, tooltip="Stochastic strength"),
+                io.Float.Input("s_noise", default=1.0, min=0.0, max=100.0, step=0.01, round=False, tooltip="SDE noise multiplier"),
+                io.Float.Input("r", default=0.5, min=0.01, max=1.0, step=0.01, round=False, tooltip="Relative step size for the intermediate stage (c2 node)"),
+            ],
+            outputs=[io.Sampler.Output()]
+        )
+
+    @classmethod
+    def execute(cls, solver_type, eta, s_noise, r) -> io.NodeOutput:
+        sampler_name = "seeds_2"
+        sampler = comfy.samplers.ksampler(
+            sampler_name,
+            {"eta": eta, "s_noise": s_noise, "r": r, "solver_type": solver_type},
+        )
+        return io.NodeOutput(sampler)
+
+
 class Noise_EmptyNoise:
     def __init__(self):
         self.seed = 0
@@ -996,6 +1021,7 @@ class CustomSamplersExtension(ComfyExtension):
             SamplerDPMAdaptative,
             SamplerER_SDE,
             SamplerSASolver,
+            SamplerSEEDS2,
             SplitSigmas,
             SplitSigmasDenoise,
             FlipSigmas,

comfy_extras/nodes_hunyuan.py

@@ -124,6 +124,87 @@ class HunyuanVideo15ImageToVideo(io.ComfyNode):
         return io.NodeOutput(positive, negative, out_latent)
 
 
+class HunyuanVideo15FirstLastFrameToVideo(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="HunyuanVideo15FirstLastFrameToVideo",
+            category="conditioning/video_models",
+            is_experimental=True,
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=848, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=33, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_start_image", optional=True),
+                io.ClipVisionOutput.Input("clip_vision_end_image", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.Image.Input("end_image", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_start_image=None, clip_vision_end_image=None) -> io.NodeOutput:
+
+        latent = torch.zeros([batch_size, 32, ((length - 1) // 4) + 1, height // 16, width // 16],
+                            device=comfy.model_management.intermediate_device())
+
+        concat_latent_image = torch.zeros((batch_size, 32, latent.shape[2], latent.shape[3], latent.shape[4]),
+                                         device=comfy.model_management.intermediate_device())
+
+        mask = torch.ones((1, 1, latent.shape[2], latent.shape[3], latent.shape[4]),
+                         device=comfy.model_management.intermediate_device())
+
+        if start_image is not None:
+            start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+
+            encoded_start = vae.encode(start_image[:, :, :, :3])
+
+            concat_latent_image[:, :, :encoded_start.shape[2], :, :] = encoded_start
+
+            start_frames_in_latent = ((start_image.shape[0] - 1) // 4) + 1
+            mask[:, :, :start_frames_in_latent] = 0.0
+
+        if end_image is not None:
+            end_image = comfy.utils.common_upscale(end_image[-length:].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+
+            encoded_end = vae.encode(end_image[:, :, :, :3])
+
+            end_frames_in_latent = ((end_image.shape[0] - 1) // 4) + 1
+            concat_latent_image[:, :, -end_frames_in_latent:, :, :] = encoded_end[:, :, -end_frames_in_latent:, :, :]
+
+            mask[:, :, -end_frames_in_latent:] = 0.0
+
+        positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
+        negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
+
+        clip_vision_output = None
+        if clip_vision_start_image is not None:
+            clip_vision_output = clip_vision_start_image
+
+        if clip_vision_end_image is not None:
+            if clip_vision_output is not None:
+                pass # Use only one embedding for now
+            else:
+                clip_vision_output = clip_vision_end_image
+
+        if clip_vision_output is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output})
+            negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output})
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return io.NodeOutput(positive, negative, out_latent)
+
+
 class HunyuanVideo15SuperResolution(io.ComfyNode):
     @classmethod
     def define_schema(cls):
@@ -406,6 +487,7 @@ class HunyuanExtension(ComfyExtension):
             EmptyHunyuanLatentVideo,
             EmptyHunyuanVideo15Latent,
             HunyuanVideo15ImageToVideo,
+            HunyuanVideo15FirstLastFrameToVideo,
             HunyuanVideo15SuperResolution,
             HunyuanVideo15LatentUpscaleWithModel,
             LatentUpscaleModelLoader,