Added Rife Temporal upsampling and Lanczos spatial upsampling

gradio_server.py

@@ -97,6 +97,8 @@ def process_prompt_and_add_tasks(
     image_to_end,
     video_to_continue,
     max_frames,
+    temporal_upsampling,
+    spatial_upsampling,    
     RIFLEx_setting,
     slg_switch,
     slg_layers,
@@ -230,6 +232,8 @@ def process_prompt_and_add_tasks(
             "image_to_end" : image_end,
             "video_to_continue" : video_to_continue ,
             "max_frames" : max_frames,
+            "temporal_upsampling" : temporal_upsampling,
+            "spatial_upsampling" : spatial_upsampling,
             "RIFLEx_setting" : RIFLEx_setting,
             "slg_switch" : slg_switch,
             "slg_layers" : slg_layers,
@@ -852,48 +856,63 @@ model_filename = ""
 # compile = "transformer"
 
 def preprocess_loras(sd):
+    if wan_model == None:
+        return sd
+    model_filename = wan_model._model_file_name
+
     first = next(iter(sd), None)
     if first == None:
         return sd
-    if not first.startswith("lora_unet_"):
-        return sd
-    new_sd = {}
-    print("Converting Lora Safetensors format to Lora Diffusers format")
-    alphas = {}
-    repl_list = ["cross_attn", "self_attn", "ffn"]
-    src_list = ["_" + k + "_" for k in repl_list]
-    tgt_list = ["." + k + "." for k in repl_list]
     
-    for k,v in sd.items():
-        k = k.replace("lora_unet_blocks_","diffusion_model.blocks.")
+    if first.startswith("lora_unet_"):
+        new_sd = {}
+        print("Converting Lora Safetensors format to Lora Diffusers format")
+        alphas = {}
+        repl_list = ["cross_attn", "self_attn", "ffn"]
+        src_list = ["_" + k + "_" for k in repl_list]
+        tgt_list = ["." + k + "." for k in repl_list]
 
-        for s,t in zip(src_list, tgt_list):
-            k = k.replace(s,t)
+        for k,v in sd.items():
+            k = k.replace("lora_unet_blocks_","diffusion_model.blocks.")
 
-        k = k.replace("lora_up","lora_B")
-        k = k.replace("lora_down","lora_A")
+            for s,t in zip(src_list, tgt_list):
+                k = k.replace(s,t)
 
-        if "alpha" in k:
-            alphas[k] = v
-        else:
+            k = k.replace("lora_up","lora_B")
+            k = k.replace("lora_down","lora_A")
+
+            if "alpha" in k:
+                alphas[k] = v
+            else:
+                new_sd[k] = v
+
+        new_alphas = {}
+        for k,v in new_sd.items():
+            if "lora_B" in k:
+                dim = v.shape[1]
+            elif "lora_A" in k:
+                dim = v.shape[0]
+            else:
+                continue
+            alpha_key = k[:-len("lora_X.weight")] +"alpha"
+            if alpha_key in alphas:
+                scale = alphas[alpha_key] / dim
+                new_alphas[alpha_key] = scale
+            else:
+                print(f"Lora alpha'{alpha_key}' is missing")
+        new_sd.update(new_alphas)
+        sd = new_sd
+
+    if "text2video" in model_filename:
+        new_sd = {}
+        # convert loras for i2v to t2v
+        for k,v in sd.items():
+            if  any(layer in k for layer in ["cross_attn.k_img", "cross_attn.v_img"]):
+                continue
             new_sd[k] = v
+        sd = new_sd
 
-    new_alphas = {}
-    for k,v in new_sd.items():
-        if "lora_B" in k:
-            dim = v.shape[1]
-        elif "lora_A" in k:
-            dim = v.shape[0]
-        else:
-            continue
-        alpha_key = k[:-len("lora_X.weight")] +"alpha"
-        if alpha_key in alphas:
-            scale = alphas[alpha_key] / dim
-            new_alphas[alpha_key] = scale
-        else:
-            print(f"Lora alpha'{alpha_key}' is missing")
-    new_sd.update(new_alphas)
-    return new_sd
+    return sd
 
 
 def download_models(transformer_filename, text_encoder_filename):
@@ -905,7 +924,7 @@ def download_models(transformer_filename, text_encoder_filename):
     from huggingface_hub import hf_hub_download, snapshot_download    
     repoId = "DeepBeepMeep/Wan2.1" 
     sourceFolderList = ["xlm-roberta-large", "",  ]
-    fileList = [ [], ["Wan2.1_VAE_bf16.safetensors", "models_clip_open-clip-xlm-roberta-large-vit-huge-14-bf16.safetensors" ] + computeList(text_encoder_filename) + computeList(transformer_filename) ]   
+    fileList = [ [], ["Wan2.1_VAE_bf16.safetensors", "models_clip_open-clip-xlm-roberta-large-vit-huge-14-bf16.safetensors", "flownet.pkl" ] + computeList(text_encoder_filename) + computeList(transformer_filename) ]   
     targetRoot = "ckpts/" 
     for sourceFolder, files in zip(sourceFolderList,fileList ):
         if len(files)==0:
@@ -1094,6 +1113,7 @@ def load_models(i2v):
         wan_model, pipe = load_i2v_model(model_filename, "720P" if res720P else "480P")
     else:
         wan_model, pipe = load_t2v_model(model_filename, "")
+    wan_model._model_file_name = model_filename
     kwargs = { "extraModelsToQuantize": None}
     if profile == 2 or profile == 4:
         kwargs["budgets"] = { "transformer" : 100 if preload  == 0 else preload, "text_encoder" : 100, "*" : 1000 }
@@ -1441,6 +1461,8 @@ def generate_video(
     image_to_end,
     video_to_continue,
     max_frames,
+    temporal_upsampling,
+    spatial_upsampling,
     RIFLEx_setting,
     slg_switch,
     slg_layers,    
@@ -1693,6 +1715,7 @@ def generate_video(
                     cfg_star_switch = cfg_star_switch,
                     cfg_zero_step = cfg_zero_step,
                 )
+            # samples = torch.empty( (1,2)) #for testing
         except Exception as e:
             if temp_filename!= None and  os.path.isfile(temp_filename):
                 os.remove(temp_filename)
@@ -1717,8 +1740,6 @@ def generate_video(
                         VRAM_crash = True
                         break
 
-            _ , exc_value, exc_traceback = sys.exc_info()
-
             state["prompt"] = ""
             if VRAM_crash:
                 new_error = "The generation of the video has encountered an error: it is likely that you have unsufficient VRAM and you should therefore reduce the video resolution or its number of frames."
@@ -1760,16 +1781,60 @@ def generate_video(
             else:
                 file_name = f"{time_flag}_seed{seed}_{sanitize_file_name(prompt[:100]).strip()}.mp4"
             video_path = os.path.join(save_path, file_name)
+            # if False: # for testing
+            #     torch.save(sample, "ouput.pt")
+            # else:
+            #     sample =torch.load("ouput.pt")
+            exp = 0
+            fps = 16
+
+            if len(temporal_upsampling) > 0 or len(spatial_upsampling) > 0:                
+                progress_args = [0, status + " - Upsampling"]
+                progress(*progress_args )   
+                gen["progress_args"] = progress_args
+
+            if temporal_upsampling == "rife2":
+                exp = 1
+            elif temporal_upsampling == "rife4":
+                exp = 2
+            
+            if exp > 0: 
+                from rife.inference import temporal_interpolation
+                sample = temporal_interpolation( os.path.join("ckpts", "flownet.pkl"), sample, exp, device="cuda")
+                fps = fps * 2**exp
+
+            if len(spatial_upsampling) > 0:
+                from wan.utils.utils import resize_lanczos
+                if spatial_upsampling == "lanczos1.5":
+                    scale = 1.5
+                else:
+                    scale = 2
+                sample = (sample + 1) / 2
+                h, w = sample.shape[-2:]
+                h *= scale
+                w *= scale
+                new_frames =[]
+                for i in range( sample.shape[1] ):
+                    frame = sample[:, i]
+                    frame = resize_lanczos(frame, h, w)
+                    frame = frame.unsqueeze(1)
+                    new_frames.append(frame)
+                sample = torch.cat(new_frames, dim=1)
+                new_frames = None
+                sample = sample * 2 - 1
+
+
             cache_video(
                 tensor=sample[None],
                 save_file=video_path,
-                fps=16,
+                fps=fps,
                 nrow=1,
                 normalize=True,
                 value_range=(-1, 1))
     
+
             configs = get_settings_dict(state, image2video, prompt, 0 if image_to_end == None else 1 , video_length, resolution, num_inference_steps, seed, repeat_generation, multi_images_gen_type, guidance_scale, flow_shift, negative_prompt, loras_choices, 
-                  loras_mult_choices, tea_cache , tea_cache_start_step_perc, RIFLEx_setting, slg_switch, slg_layers, slg_start, slg_end, cfg_star_switch, cfg_zero_step)
+                  loras_mult_choices, tea_cache , tea_cache_start_step_perc, temporal_upsampling, spatial_upsampling, RIFLEx_setting, slg_switch, slg_layers, slg_start, slg_end, cfg_star_switch, cfg_zero_step)
 
             metadata_choice = server_config.get("metadata_choice","metadata")
             if metadata_choice == "json":
@@ -2231,7 +2296,7 @@ def switch_advanced(state, new_advanced, lset_name):
 
 
 def get_settings_dict(state, i2v, prompt, image_prompt_type, video_length, resolution, num_inference_steps, seed, repeat_generation, multi_images_gen_type, guidance_scale, flow_shift, negative_prompt, loras_choices, 
-                      loras_mult_choices, tea_cache_setting, tea_cache_start_step_perc, RIFLEx_setting, slg_switch, slg_layers, slg_start_perc, slg_end_perc, cfg_star_switch, cfg_zero_step):
+                      loras_mult_choices, tea_cache_setting, tea_cache_start_step_perc, temporal_upsampling, spatial_upsampling, RIFLEx_setting, slg_switch, slg_layers, slg_start_perc, slg_end_perc, cfg_star_switch, cfg_zero_step):
 
     loras = state["loras"]
     activated_loras = [Path( loras[int(no)]).parts[-1]  for no in loras_choices ]
@@ -2251,6 +2316,8 @@ def get_settings_dict(state, i2v, prompt, image_prompt_type, video_length, resol
         "loras_multipliers": loras_mult_choices,
         "tea_cache": tea_cache_setting,
         "tea_cache_start_step_perc": tea_cache_start_step_perc,
+        "temporal_upsampling" : temporal_upsampling, 
+        "spatial_upsampling" : spatial_upsampling,
         "RIFLEx_setting": RIFLEx_setting,
         "slg_switch": slg_switch,
         "slg_layers": slg_layers,
@@ -2269,14 +2336,14 @@ def get_settings_dict(state, i2v, prompt, image_prompt_type, video_length, resol
     return ui_settings
 
 def save_settings(state, prompt, image_prompt_type, video_length, resolution, num_inference_steps, seed, repeat_generation, multi_images_gen_type, guidance_scale, flow_shift, negative_prompt, loras_choices, 
-                      loras_mult_choices, tea_cache_setting, tea_cache_start_step_perc, RIFLEx_setting, slg_switch, slg_layers, slg_start_perc, slg_end_perc, cfg_star_switch, cfg_zero_step):
+                      loras_mult_choices, tea_cache_setting, tea_cache_start_step_perc, temporal_upsampling, spatial_upsampling, RIFLEx_setting, slg_switch, slg_layers, slg_start_perc, slg_end_perc, cfg_star_switch, cfg_zero_step):
 
     if state.get("validate_success",0) != 1:
         return
 
     image2video = state["image2video"]
     ui_defaults = get_settings_dict(state, image2video, prompt, image_prompt_type, video_length, resolution, num_inference_steps, seed, repeat_generation, multi_images_gen_type, guidance_scale, flow_shift, negative_prompt, loras_choices, 
-                      loras_mult_choices, tea_cache_setting, tea_cache_start_step_perc, RIFLEx_setting, slg_switch, slg_layers, slg_start_perc, slg_end_perc, cfg_star_switch, cfg_zero_step)
+                      loras_mult_choices, tea_cache_setting, tea_cache_start_step_perc, temporal_upsampling, spatial_upsampling, RIFLEx_setting, slg_switch, slg_layers, slg_start_perc, slg_end_perc, cfg_star_switch, cfg_zero_step)
 
     defaults_filename = get_settings_file_name(image2video)
 
@@ -2538,6 +2605,32 @@ def generate_video_tab(image2video=False):
                         )
                         tea_cache_start_step_perc = gr.Slider(0, 100, value=ui_defaults["tea_cache_start_step_perc"], step=1, label="Tea Cache starting moment in % of generation") 
 
+                    with gr.Row():
+                        gr.Markdown("<B>Upsampling</B>")
+                    with gr.Row():
+                        temporal_upsampling_choice = gr.Dropdown(
+                            choices=[
+                                ("Disabled", ""),
+                                ("Rife x2 (32 frames/s)", "rife2"), 
+                                ("Rife x4 (64 frames/s)", "rife4"), 
+                            ],
+                            value=ui_defaults.get("temporal_upsampling", ""),
+                            visible=True,
+                            scale = 1,
+                            label="Temporal Upsampling"
+                        )
+                        spatial_upsampling_choice = gr.Dropdown(
+                            choices=[
+                                ("Disabled", ""),
+                                ("Lanczos x1.5", "lanczos1.5"), 
+                                ("Lanczos x2.0", "lanczos2"), 
+                            ],
+                            value=ui_defaults.get("spatial_upsampling", ""),
+                            visible=True,
+                            scale = 1,
+                            label="Spatial Upsampling"
+                        )
+
                     gr.Markdown("<B>With Riflex you can generate videos longer than 5s which is the default duration of videos used to train the model</B>")
                     RIFLEx_setting = gr.Dropdown(
                         choices=[
@@ -2699,7 +2792,7 @@ def generate_video_tab(image2video=False):
                 )
         save_settings_btn.click( fn=validate_wizard_prompt, inputs =[state, wizard_prompt_activated_var, wizard_variables_var,  prompt, wizard_prompt, *prompt_vars] , outputs= [prompt]).then(
             save_settings, inputs = [state, prompt, image_prompt_type_radio, video_length, resolution, num_inference_steps, seed, repeat_generation, multi_images_gen_type, guidance_scale, flow_shift, negative_prompt, 
-                                                         loras_choices, loras_mult_choices, tea_cache_setting, tea_cache_start_step_perc, RIFLEx_setting, slg_switch, slg_layers,
+                                                         loras_choices, loras_mult_choices, tea_cache_setting, tea_cache_start_step_perc, temporal_upsampling_choice, spatial_upsampling_choice, RIFLEx_setting, slg_switch, slg_layers,
                                                          slg_start_perc, slg_end_perc, cfg_star_switch, cfg_zero_step  ], outputs = [])
         save_lset_btn.click(validate_save_lset, inputs=[lset_name], outputs=[apply_lset_btn, refresh_lora_btn, delete_lset_btn, save_lset_btn,confirm_save_lset_btn, cancel_lset_btn, save_lset_prompt_drop])
         confirm_save_lset_btn.click(fn=validate_wizard_prompt, inputs =[state, wizard_prompt_activated_var, wizard_variables_var, prompt, wizard_prompt, *prompt_vars] , outputs= [prompt]).then(
@@ -2758,6 +2851,8 @@ def generate_video_tab(image2video=False):
             image_to_end,
             video_to_continue,
             max_frames,
+            temporal_upsampling_choice,
+            spatial_upsampling_choice,
             RIFLEx_setting,
             slg_switch, 
             slg_layers,

rife/IFNet_HDv3.py

@@ -0,0 +1,133 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+# from ..model.warplayer import warp
+
+# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+backwarp_tenGrid = {}
+
+def warp(tenInput, tenFlow, device):
+    k = (str(tenFlow.device), str(tenFlow.size()))
+    if k not in backwarp_tenGrid:
+        tenHorizontal = torch.linspace(-1.0, 1.0, tenFlow.shape[3], device=device).view(
+            1, 1, 1, tenFlow.shape[3]).expand(tenFlow.shape[0], -1, tenFlow.shape[2], -1)
+        tenVertical = torch.linspace(-1.0, 1.0, tenFlow.shape[2], device=device).view(
+            1, 1, tenFlow.shape[2], 1).expand(tenFlow.shape[0], -1, -1, tenFlow.shape[3])
+        backwarp_tenGrid[k] = torch.cat(
+            [tenHorizontal, tenVertical], 1).to(device)
+
+    tenFlow = torch.cat([tenFlow[:, 0:1, :, :] / ((tenInput.shape[3] - 1.0) / 2.0),
+                         tenFlow[:, 1:2, :, :] / ((tenInput.shape[2] - 1.0) / 2.0)], 1)
+
+    g = (backwarp_tenGrid[k] + tenFlow).permute(0, 2, 3, 1)
+    return torch.nn.functional.grid_sample(input=tenInput, grid=g, mode='bilinear', padding_mode='border', align_corners=True)
+
+def conv(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=True),        
+        nn.PReLU(out_planes)
+    )
+
+def conv_bn(in_planes, out_planes, kernel_size=3, stride=1, padding=1, dilation=1):
+    return nn.Sequential(
+        nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size, stride=stride,
+                  padding=padding, dilation=dilation, bias=False),
+        nn.BatchNorm2d(out_planes),
+        nn.PReLU(out_planes)
+    )
+
+class IFBlock(nn.Module):
+    def __init__(self, in_planes, c=64):
+        super(IFBlock, self).__init__()
+        self.conv0 = nn.Sequential(
+            conv(in_planes, c//2, 3, 2, 1),
+            conv(c//2, c, 3, 2, 1),
+            )
+        self.convblock0 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.convblock1 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.convblock2 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.convblock3 = nn.Sequential(
+            conv(c, c),
+            conv(c, c)
+        )
+        self.conv1 = nn.Sequential(
+            nn.ConvTranspose2d(c, c//2, 4, 2, 1),
+            nn.PReLU(c//2),
+            nn.ConvTranspose2d(c//2, 4, 4, 2, 1),
+        )
+        self.conv2 = nn.Sequential(
+            nn.ConvTranspose2d(c, c//2, 4, 2, 1),
+            nn.PReLU(c//2),
+            nn.ConvTranspose2d(c//2, 1, 4, 2, 1),
+        )
+
+    def forward(self, x, flow, scale=1):
+        x = F.interpolate(x, scale_factor= 1. / scale, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+        flow = F.interpolate(flow, scale_factor= 1. / scale, mode="bilinear", align_corners=False, recompute_scale_factor=False) * 1. / scale
+        feat = self.conv0(torch.cat((x, flow), 1))
+        feat = self.convblock0(feat) + feat
+        feat = self.convblock1(feat) + feat
+        feat = self.convblock2(feat) + feat
+        feat = self.convblock3(feat) + feat        
+        flow = self.conv1(feat)
+        mask = self.conv2(feat)
+        flow = F.interpolate(flow, scale_factor=scale, mode="bilinear", align_corners=False, recompute_scale_factor=False) * scale
+        mask = F.interpolate(mask, scale_factor=scale, mode="bilinear", align_corners=False, recompute_scale_factor=False)
+        return flow, mask
+        
+class IFNet(nn.Module):
+    def __init__(self):
+        super(IFNet, self).__init__()
+        self.block0 = IFBlock(7+4, c=90)
+        self.block1 = IFBlock(7+4, c=90)
+        self.block2 = IFBlock(7+4, c=90)
+        self.block_tea = IFBlock(10+4, c=90)
+        # self.contextnet = Contextnet()
+        # self.unet = Unet()
+
+    def forward(self, x, scale_list=[4, 2, 1], training=False):
+        if training == False:
+            channel = x.shape[1] // 2
+            img0 = x[:, :channel]
+            img1 = x[:, channel:]
+        flow_list = []
+        merged = []
+        mask_list = []
+        warped_img0 = img0
+        warped_img1 = img1
+        flow = (x[:, :4]).detach() * 0
+        mask = (x[:, :1]).detach() * 0
+        loss_cons = 0
+        block = [self.block0, self.block1, self.block2]
+        for i in range(3):
+            f0, m0 = block[i](torch.cat((warped_img0[:, :3], warped_img1[:, :3], mask), 1), flow, scale=scale_list[i])
+            f1, m1 = block[i](torch.cat((warped_img1[:, :3], warped_img0[:, :3], -mask), 1), torch.cat((flow[:, 2:4], flow[:, :2]), 1), scale=scale_list[i])
+            flow = flow + (f0 + torch.cat((f1[:, 2:4], f1[:, :2]), 1)) / 2
+            mask = mask + (m0 + (-m1)) / 2
+            mask_list.append(mask)
+            flow_list.append(flow)
+            warped_img0 = warp(img0, flow[:, :2], device= flow.device)
+            warped_img1 = warp(img1, flow[:, 2:4], device= flow.device)
+            merged.append((warped_img0, warped_img1))
+        '''
+        c0 = self.contextnet(img0, flow[:, :2])
+        c1 = self.contextnet(img1, flow[:, 2:4])
+        tmp = self.unet(img0, img1, warped_img0, warped_img1, mask, flow, c0, c1)
+        res = tmp[:, 1:4] * 2 - 1
+        '''
+        for i in range(3):
+            mask_list[i] = torch.sigmoid(mask_list[i])
+            merged[i] = merged[i][0] * mask_list[i] + merged[i][1] * (1 - mask_list[i])
+            # merged[i] = torch.clamp(merged[i] + res, 0, 1)        
+        return flow_list, mask_list[2], merged

rife/RIFE_HDv3.py

@@ -0,0 +1,84 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from torch.optim import AdamW
+import torch.optim as optim
+import itertools
+from torch.nn.parallel import DistributedDataParallel as DDP
+from .IFNet_HDv3 import *
+import torch.nn.functional as F
+# from ..model.loss import *
+
+# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    
+class Model:
+    def __init__(self, local_rank=-1):
+        self.flownet = IFNet()
+        # self.device()
+        # self.optimG = AdamW(self.flownet.parameters(), lr=1e-6, weight_decay=1e-4)
+        # self.epe = EPE()
+        # self.vgg = VGGPerceptualLoss().to(device)
+        # self.sobel = SOBEL()
+        if local_rank != -1:
+            self.flownet = DDP(self.flownet, device_ids=[local_rank], output_device=local_rank)
+
+    def train(self):
+        self.flownet.train()
+
+    def eval(self):
+        self.flownet.eval()
+
+    def to(self, device):
+        self.flownet.to(device)
+
+    def load_model(self, path, rank=0, device = "cuda"):
+        self.device = device 
+        def convert(param):
+            if rank == -1:
+                return {
+                    k.replace("module.", ""): v
+                    for k, v in param.items()
+                    if "module." in k
+                }
+            else:
+                return param
+        self.flownet.load_state_dict(convert(torch.load(path, map_location=device)))
+        
+    def save_model(self, path, rank=0):
+        if rank == 0:
+            torch.save(self.flownet.state_dict(),'{}/flownet.pkl'.format(path))
+
+    def inference(self, img0, img1, scale=1.0):
+        imgs = torch.cat((img0, img1), 1)
+        scale_list = [4/scale, 2/scale, 1/scale]
+        flow, mask, merged = self.flownet(imgs, scale_list)
+        return merged[2]
+    
+    def update(self, imgs, gt, learning_rate=0, mul=1, training=True, flow_gt=None):
+        for param_group in self.optimG.param_groups:
+            param_group['lr'] = learning_rate
+        img0 = imgs[:, :3]
+        img1 = imgs[:, 3:]
+        if training:
+            self.train()
+        else:
+            self.eval()
+        scale = [4, 2, 1]
+        flow, mask, merged = self.flownet(torch.cat((imgs, gt), 1), scale=scale, training=training)
+        loss_l1 = (merged[2] - gt).abs().mean()
+        loss_smooth = self.sobel(flow[2], flow[2]*0).mean()
+        # loss_vgg = self.vgg(merged[2], gt)
+        if training:
+            self.optimG.zero_grad()
+            loss_G = loss_cons + loss_smooth * 0.1
+            loss_G.backward()
+            self.optimG.step()
+        else:
+            flow_teacher = flow[2]
+        return merged[2], {
+            'mask': mask,
+            'flow': flow[2][:, :2],
+            'loss_l1': loss_l1,
+            'loss_cons': loss_cons,
+            'loss_smooth': loss_smooth,
+            }

rife/inference.py

@@ -0,0 +1,119 @@
+import os
+import torch
+from torch.nn import functional as F
+# from .model.pytorch_msssim import ssim_matlab
+from .ssim import ssim_matlab
+
+from .RIFE_HDv3 import Model
+
+def get_frame(frames, frame_no):
+    if frame_no >= frames.shape[1]:
+        return None
+    frame = (frames[:, frame_no] + 1) /2
+    frame = frame.clip(0., 1.)
+    return frame
+
+def add_frame(frames, frame, h, w):
+    frame = (frame * 2) - 1
+    frame = frame.clip(-1., 1.)    
+    frame = frame.squeeze(0)
+    frame = frame[:, :h, :w]
+    frame = frame.unsqueeze(1)
+    frames.append(frame.cpu())
+
+def process_frames(model, device, frames, exp):
+    pos = 0
+    output_frames = []
+
+    lastframe = get_frame(frames, 0)
+    _,  h, w = lastframe.shape
+    scale = 1
+    fp16 = False
+
+    def make_inference(I0, I1, n):
+        middle = model.inference(I0, I1, scale)
+        if n == 1:
+            return [middle]
+        first_half = make_inference(I0, middle, n=n//2)
+        second_half = make_inference(middle, I1, n=n//2)
+        if n%2:
+            return [*first_half, middle, *second_half]
+        else:
+            return [*first_half, *second_half]
+
+    tmp = max(32, int(32 / scale))
+    ph = ((h - 1) // tmp + 1) * tmp
+    pw = ((w - 1) // tmp + 1) * tmp
+    padding = (0, pw - w, 0, ph - h)
+
+    def pad_image(img):
+        if(fp16):
+            return F.pad(img, padding).half()
+        else:
+            return F.pad(img, padding)
+
+    I1 = lastframe.to(device, non_blocking=True).unsqueeze(0)
+    I1 = pad_image(I1)
+    temp = None # save lastframe when processing static frame
+
+    while True:
+        if temp is not None:
+            frame = temp
+            temp = None
+        else:
+            pos += 1
+            frame = get_frame(frames, pos)
+        if frame is None:
+            break
+        I0 = I1
+        I1 = frame.to(device, non_blocking=True).unsqueeze(0)
+        I1 = pad_image(I1)
+        I0_small = F.interpolate(I0, (32, 32), mode='bilinear', align_corners=False)
+        I1_small = F.interpolate(I1, (32, 32), mode='bilinear', align_corners=False)
+        ssim = ssim_matlab(I0_small[:, :3], I1_small[:, :3])
+
+        break_flag = False
+        if ssim > 0.996:        
+            pos += 1
+            frame = get_frame(frames, pos)
+            if frame is None:
+                break_flag = True
+                frame = lastframe
+            else:
+                temp = frame
+            I1 = frame.to(device, non_blocking=True).unsqueeze(0)
+            I1 = pad_image(I1)
+            I1 = model.inference(I0, I1, scale)
+            I1_small = F.interpolate(I1, (32, 32), mode='bilinear', align_corners=False)
+            ssim = ssim_matlab(I0_small[:, :3], I1_small[:, :3])
+            frame = I1[0]
+        
+        if ssim < 0.2:
+            output = []
+            for _ in range((2 ** exp) - 1):
+                output.append(I0)
+        else:
+            output = make_inference(I0, I1, 2**exp-1) if exp else []
+
+        add_frame(output_frames, lastframe, h, w)
+        for mid in output:
+            add_frame(output_frames, mid, h, w)
+        lastframe = frame
+        if break_flag:
+            break
+
+    add_frame(output_frames, lastframe, h, w)
+    return torch.cat( output_frames, dim=1)
+
+def temporal_interpolation(model_path, frames, exp, device ="cuda"):
+
+    model = Model()
+    model.load_model(model_path, -1, device=device)
+
+    model.eval()
+    model.to(device=device)
+
+    with torch.no_grad():    
+        output = process_frames(model, device, frames, exp)
+
+    return output

rife/ssim.py

@@ -0,0 +1,200 @@
+import torch
+import torch.nn.functional as F
+from math import exp
+import numpy as np
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+def gaussian(window_size, sigma):
+    gauss = torch.Tensor([exp(-(x - window_size//2)**2/float(2*sigma**2)) for x in range(window_size)])
+    return gauss/gauss.sum()
+
+
+def create_window(window_size, channel=1):
+    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
+    _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0).to(device)
+    window = _2D_window.expand(channel, 1, window_size, window_size).contiguous()
+    return window
+
+def create_window_3d(window_size, channel=1):
+    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
+    _2D_window = _1D_window.mm(_1D_window.t())
+    _3D_window = _2D_window.unsqueeze(2) @ (_1D_window.t())
+    window = _3D_window.expand(1, channel, window_size, window_size, window_size).contiguous().to(device)
+    return window
+
+
+def ssim(img1, img2, window_size=11, window=None, size_average=True, full=False, val_range=None):
+    # Value range can be different from 255. Other common ranges are 1 (sigmoid) and 2 (tanh).
+    if val_range is None:
+        if torch.max(img1) > 128:
+            max_val = 255
+        else:
+            max_val = 1
+
+        if torch.min(img1) < -0.5:
+            min_val = -1
+        else:
+            min_val = 0
+        L = max_val - min_val
+    else:
+        L = val_range
+
+    padd = 0
+    (_, channel, height, width) = img1.size()
+    if window is None:
+        real_size = min(window_size, height, width)
+        window = create_window(real_size, channel=channel).to(img1.device)
+    
+    # mu1 = F.conv2d(img1, window, padding=padd, groups=channel)
+    # mu2 = F.conv2d(img2, window, padding=padd, groups=channel)
+    mu1 = F.conv2d(F.pad(img1, (5, 5, 5, 5), mode='replicate'), window, padding=padd, groups=channel)
+    mu2 = F.conv2d(F.pad(img2, (5, 5, 5, 5), mode='replicate'), window, padding=padd, groups=channel)
+
+    mu1_sq = mu1.pow(2)
+    mu2_sq = mu2.pow(2)
+    mu1_mu2 = mu1 * mu2
+
+    sigma1_sq = F.conv2d(F.pad(img1 * img1, (5, 5, 5, 5), 'replicate'), window, padding=padd, groups=channel) - mu1_sq
+    sigma2_sq = F.conv2d(F.pad(img2 * img2, (5, 5, 5, 5), 'replicate'), window, padding=padd, groups=channel) - mu2_sq
+    sigma12 = F.conv2d(F.pad(img1 * img2, (5, 5, 5, 5), 'replicate'), window, padding=padd, groups=channel) - mu1_mu2
+
+    C1 = (0.01 * L) ** 2
+    C2 = (0.03 * L) ** 2
+
+    v1 = 2.0 * sigma12 + C2
+    v2 = sigma1_sq + sigma2_sq + C2
+    cs = torch.mean(v1 / v2)  # contrast sensitivity
+
+    ssim_map = ((2 * mu1_mu2 + C1) * v1) / ((mu1_sq + mu2_sq + C1) * v2)
+
+    if size_average:
+        ret = ssim_map.mean()
+    else:
+        ret = ssim_map.mean(1).mean(1).mean(1)
+
+    if full:
+        return ret, cs
+    return ret
+
+
+def ssim_matlab(img1, img2, window_size=11, window=None, size_average=True, full=False, val_range=None):
+    # Value range can be different from 255. Other common ranges are 1 (sigmoid) and 2 (tanh).
+    if val_range is None:
+        if torch.max(img1) > 128:
+            max_val = 255
+        else:
+            max_val = 1
+
+        if torch.min(img1) < -0.5:
+            min_val = -1
+        else:
+            min_val = 0
+        L = max_val - min_val
+    else:
+        L = val_range
+
+    padd = 0
+    (_, _, height, width) = img1.size()
+    if window is None:
+        real_size = min(window_size, height, width)
+        window = create_window_3d(real_size, channel=1).to(img1.device)
+        # Channel is set to 1 since we consider color images as volumetric images
+
+    img1 = img1.unsqueeze(1)
+    img2 = img2.unsqueeze(1)
+
+    mu1 = F.conv3d(F.pad(img1, (5, 5, 5, 5, 5, 5), mode='replicate'), window, padding=padd, groups=1)
+    mu2 = F.conv3d(F.pad(img2, (5, 5, 5, 5, 5, 5), mode='replicate'), window, padding=padd, groups=1)
+
+    mu1_sq = mu1.pow(2)
+    mu2_sq = mu2.pow(2)
+    mu1_mu2 = mu1 * mu2
+
+    sigma1_sq = F.conv3d(F.pad(img1 * img1, (5, 5, 5, 5, 5, 5), 'replicate'), window, padding=padd, groups=1) - mu1_sq
+    sigma2_sq = F.conv3d(F.pad(img2 * img2, (5, 5, 5, 5, 5, 5), 'replicate'), window, padding=padd, groups=1) - mu2_sq
+    sigma12 = F.conv3d(F.pad(img1 * img2, (5, 5, 5, 5, 5, 5), 'replicate'), window, padding=padd, groups=1) - mu1_mu2
+
+    C1 = (0.01 * L) ** 2
+    C2 = (0.03 * L) ** 2
+
+    v1 = 2.0 * sigma12 + C2
+    v2 = sigma1_sq + sigma2_sq + C2
+    cs = torch.mean(v1 / v2)  # contrast sensitivity
+
+    ssim_map = ((2 * mu1_mu2 + C1) * v1) / ((mu1_sq + mu2_sq + C1) * v2)
+
+    if size_average:
+        ret = ssim_map.mean()
+    else:
+        ret = ssim_map.mean(1).mean(1).mean(1)
+
+    if full:
+        return ret, cs
+    return ret
+
+
+def msssim(img1, img2, window_size=11, size_average=True, val_range=None, normalize=False):
+    device = img1.device
+    weights = torch.FloatTensor([0.0448, 0.2856, 0.3001, 0.2363, 0.1333]).to(device)
+    levels = weights.size()[0]
+    mssim = []
+    mcs = []
+    for _ in range(levels):
+        sim, cs = ssim(img1, img2, window_size=window_size, size_average=size_average, full=True, val_range=val_range)
+        mssim.append(sim)
+        mcs.append(cs)
+
+        img1 = F.avg_pool2d(img1, (2, 2))
+        img2 = F.avg_pool2d(img2, (2, 2))
+
+    mssim = torch.stack(mssim)
+    mcs = torch.stack(mcs)
+
+    # Normalize (to avoid NaNs during training unstable models, not compliant with original definition)
+    if normalize:
+        mssim = (mssim + 1) / 2
+        mcs = (mcs + 1) / 2
+
+    pow1 = mcs ** weights
+    pow2 = mssim ** weights
+    # From Matlab implementation https://ece.uwaterloo.ca/~z70wang/research/iwssim/
+    output = torch.prod(pow1[:-1] * pow2[-1])
+    return output
+
+
+# Classes to re-use window
+class SSIM(torch.nn.Module):
+    def __init__(self, window_size=11, size_average=True, val_range=None):
+        super(SSIM, self).__init__()
+        self.window_size = window_size
+        self.size_average = size_average
+        self.val_range = val_range
+
+        # Assume 3 channel for SSIM
+        self.channel = 3
+        self.window = create_window(window_size, channel=self.channel)
+
+    def forward(self, img1, img2):
+        (_, channel, _, _) = img1.size()
+
+        if channel == self.channel and self.window.dtype == img1.dtype:
+            window = self.window
+        else:
+            window = create_window(self.window_size, channel).to(img1.device).type(img1.dtype)
+            self.window = window
+            self.channel = channel
+
+        _ssim = ssim(img1, img2, window=window, window_size=self.window_size, size_average=self.size_average)
+        dssim = (1 - _ssim) / 2
+        return dssim
+
+class MSSSIM(torch.nn.Module):
+    def __init__(self, window_size=11, size_average=True, channel=3):
+        super(MSSSIM, self).__init__()
+        self.window_size = window_size
+        self.size_average = size_average
+        self.channel = channel
+
+    def forward(self, img1, img2):
+        return msssim(img1, img2, window_size=self.window_size, size_average=self.size_average)

wan/image2video.py

@@ -25,8 +25,7 @@ from .utils.fm_solvers import (FlowDPMSolverMultistepScheduler,
                                get_sampling_sigmas, retrieve_timesteps)
 from .utils.fm_solvers_unipc import FlowUniPCMultistepScheduler
 from wan.modules.posemb_layers import get_rotary_pos_embed
-
-from PIL import Image
+from wan.utils.utils import resize_lanczos
 
 def optimized_scale(positive_flat, negative_flat):
 
@@ -41,10 +40,6 @@ def optimized_scale(positive_flat, negative_flat):
     
     return st_star
 
-def resize_lanczos(img, h, w):
-    img = Image.fromarray(np.clip(255. * img.movedim(0, -1).cpu().numpy(), 0, 255).astype(np.uint8))
-    img = img.resize((w,h), resample=Image.Resampling.LANCZOS) 
-    return torch.from_numpy(np.array(img).astype(np.float32) / 255.0).movedim(-1, 0)
 
 
 class WanI2V:
@@ -285,21 +280,6 @@ class WanI2V:
             self.clip.model.cpu()
 
         from mmgp import offload
-
-
-        # img_interpolated.save('aaa.png')        
-
-        # img_interpolated = torch.from_numpy(np.array(img_interpolated).astype(np.float32) / 255.0).movedim(-1, 0)
-
-        # img_interpolated = torch.nn.functional.interpolate(img[None].cpu(), size=(h, w), mode='lanczos')
-        # img_interpolated = img_interpolated.squeeze(0).transpose(0,2).transpose(1,0)
-        # img_interpolated = img_interpolated.clamp(-1, 1) 
-        # img_interpolated = (img_interpolated + 1)/2
-        # img_interpolated = (img_interpolated*255).type(torch.uint8) 
-        # img_interpolated = img_interpolated.cpu().numpy()
-        # xxx = Image.fromarray(img_interpolated, 'RGB')
-        # xxx.save('my.png')        
-
         offload.last_offload_obj.unload_all()
         if any_end_frame:
             mean2 = 0

wan/utils/utils.py

@@ -7,9 +7,16 @@ import os.path as osp
 import imageio
 import torch
 import torchvision
+from PIL import Image
+import numpy as np
 
 __all__ = ['cache_video', 'cache_image', 'str2bool']
 
+def resize_lanczos(img, h, w):
+    img = Image.fromarray(np.clip(255. * img.movedim(0, -1).cpu().numpy(), 0, 255).astype(np.uint8))
+    img = img.resize((w,h), resample=Image.Resampling.LANCZOS) 
+    return torch.from_numpy(np.array(img).astype(np.float32) / 255.0).movedim(-1, 0)
+
 
 def rand_name(length=8, suffix=''):
     name = binascii.b2a_hex(os.urandom(length)).decode('utf-8')