expose flow_shift, add vid2vid

hyvideo/diffusion/pipelines/pipeline_hunyuan_video.py

@@ -226,6 +226,7 @@ class HunyuanVideoPipeline(DiffusionPipeline):
         else:
             latents = latents.to(device)
             timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, denoise_strength, device)
+            latent_timestep = timesteps[:1]
             frames_needed = noise.shape[1]
             current_frames = latents.shape[1]
             
@@ -236,8 +237,9 @@ class HunyuanVideoPipeline(DiffusionPipeline):
                 self.additional_frames = repeat_factor
             elif frames_needed < current_frames:
                 latents = latents[:, :frames_needed, :, :, :]
+            print(timesteps)
 
-            latents = noise.to(device) *0.5 + latents * 0.5
+            latents = latents * (1 - latent_timestep / 1000) + latent_timestep / 1000 * noise
 
         # Check existence to make it compatible with FlowMatchEulerDiscreteScheduler
         if hasattr(self.scheduler, "init_noise_sigma"):

hyvideo/diffusion/schedulers/scheduling_flow_match_discrete.py

@@ -77,9 +77,10 @@ class FlowMatchDiscreteScheduler(SchedulerMixin, ConfigMixin):
         n_tokens: Optional[int] = None,
     ):
         sigmas = torch.linspace(1, 0, num_train_timesteps + 1)
-
+        print("Scheduler config:", self.config)
         if not reverse:
             sigmas = sigmas.flip(0)
+        self.shift = shift
 
         self.sigmas = sigmas
         # the value fed to model
@@ -183,7 +184,7 @@ class FlowMatchDiscreteScheduler(SchedulerMixin, ConfigMixin):
         return sample
 
     def sd3_time_shift(self, t: torch.Tensor):
-        return (self.config.shift * t) / (1 + (self.config.shift - 1) * t)
+        return (self.shift * t) / (1 + (self.shift - 1) * t)
 
     def step(
         self,

nodes.py

@@ -572,6 +572,7 @@ class HyVideoSampler:
                 "num_frames": ("INT", {"default": 49, "min": 1, "max": 1024, "step": 4}),
                 "steps": ("INT", {"default": 30, "min": 1}),
                 "guidance_scale": ("FLOAT", {"default": 6.0, "min": 0.0, "max": 30.0, "step": 0.01}),
+                "flow_shift": ("FLOAT", {"default": 9.0, "min": 0.0, "max": 30.0, "step": 0.01}),
                 "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
                 "force_offload": ("BOOLEAN", {"default": True}),
                 
@@ -587,7 +588,7 @@ class HyVideoSampler:
     FUNCTION = "process"
     CATEGORY = "HunyuanVideoWrapper"
 
-    def process(self, model, hyvid_embeds, steps, guidance_scale, seed, width, height, num_frames, samples=None, denoise_strength=1.0, force_offload=True):
+    def process(self, model, hyvid_embeds, flow_shift, steps, guidance_scale, seed, width, height, num_frames, samples=None, denoise_strength=1.0, force_offload=True):
         mm.unload_all_models()
         mm.soft_empty_cache()
 
@@ -622,6 +623,8 @@ class HyVideoSampler:
             model["pipe"].transformer, num_frames, target_height, target_width
         )
         n_tokens = freqs_cos.shape[0]
+
+        model["pipe"].scheduler.shift = flow_shift
   
         # autocast_context = torch.autocast(
         #     mm.get_autocast_device(device), dtype=dtype