Add EMA-VFI (CVPR 2023) frame interpolation support

Integrate EMA-VFI alongside existing BIM-VFI with three new ComfyUI nodes:
Load EMA-VFI Model, EMA-VFI Interpolate, and EMA-VFI Segment Interpolate.

Architecture files vendored from MCG-NJU/EMA-VFI with device-awareness
fixes (removed hardcoded .cuda() calls), warp cache management, and
relative imports. InputPadder extended to support EMA-VFI's replicate
center-symmetric padding. Auto-installs timm dependency on first load.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

.gitignore

@@ -0,0 +1,3 @@
+__pycache__/
+*.pyc
+*.pyo

README.md

@@ -1,10 +1,12 @@
-# ComfyUI BIM-VFI
+# ComfyUI BIM-VFI + EMA-VFI
 
-ComfyUI custom nodes for video frame interpolation using [BiM-VFI](https://github.com/KAIST-VICLab/BiM-VFI) (CVPR 2025). Designed for long videos with thousands of frames — processes them without running out of VRAM.
+ComfyUI custom nodes for video frame interpolation using [BiM-VFI](https://github.com/KAIST-VICLab/BiM-VFI) (CVPR 2025) and [EMA-VFI](https://github.com/MCG-NJU/EMA-VFI) (CVPR 2023). Designed for long videos with thousands of frames — processes them without running out of VRAM.
 
 ## Nodes
 
-### Load BIM-VFI Model
+### BIM-VFI
+
+#### Load BIM-VFI Model
 
 Loads the BiM-VFI checkpoint. Auto-downloads from Google Drive on first use to `ComfyUI/models/bim-vfi/`.
 
@@ -14,7 +16,7 @@ Loads the BiM-VFI checkpoint. Auto-downloads from Google Drive on first use to `
 | **auto_pyr_level** | Auto-select pyramid level by resolution (<540p=3, 540p=5, 1080p=6, 4K=7) |
 | **pyr_level** | Manual pyramid level (3-7), only used when auto is off |
 
-### BIM-VFI Interpolate
+#### BIM-VFI Interpolate
 
 Interpolates frames from an image batch.
 
@@ -24,12 +26,47 @@ Interpolates frames from an image batch.
 | **model** | Model from the loader node |
 | **multiplier** | 2x, 4x, or 8x frame rate (recursive 2x passes) |
 | **batch_size** | Frame pairs processed simultaneously (higher = faster, more VRAM) |
-| **chunk_size** | Process in segments of N input frames (0 = disabled). Bounds memory for very long videos. Result is identical to processing all at once |
+| **chunk_size** | Process in segments of N input frames (0 = disabled). Bounds VRAM for very long videos. Result is identical to processing all at once |
 | **keep_device** | Keep model on GPU between pairs (faster, ~200MB constant VRAM) |
 | **all_on_gpu** | Keep all intermediate frames on GPU (fast, needs large VRAM) |
 | **clear_cache_after_n_frames** | Clear CUDA cache every N pairs to prevent VRAM buildup |
 
-**Output frame count:** 2x = 2N-1, 4x = 4N-3, 8x = 8N-7
+#### BIM-VFI Segment Interpolate
+
+Same as Interpolate but processes a single segment of the input. Chain multiple instances with Save nodes between them to bound peak RAM. The model pass-through output forces sequential execution.
+
+#### BIM-VFI Concat Videos
+
+Concatenates segment video files into a single video using ffmpeg. Connect from the last Segment Interpolate's model output to ensure it runs after all segments are saved.
+
+### EMA-VFI
+
+#### Load EMA-VFI Model
+
+Loads an EMA-VFI checkpoint. Auto-downloads from Google Drive on first use to `ComfyUI/models/ema-vfi/`. Variant (large/small) and timestep support are auto-detected from the filename.
+
+| Input | Description |
+|-------|-------------|
+| **model_path** | Checkpoint file from `models/ema-vfi/` |
+| **tta** | Test-time augmentation: flip input and average with unflipped result (~2x slower, slightly better quality) |
+
+Available checkpoints:
+| Checkpoint | Variant | Params | Arbitrary timestep |
+|-----------|---------|--------|-------------------|
+| `ours_t.pkl` | Large | ~65M | Yes |
+| `ours.pkl` | Large | ~65M | No (fixed 0.5) |
+| `ours_small_t.pkl` | Small | ~14M | Yes |
+| `ours_small.pkl` | Small | ~14M | No (fixed 0.5) |
+
+#### EMA-VFI Interpolate
+
+Interpolates frames from an image batch. Same controls as BIM-VFI Interpolate.
+
+#### EMA-VFI Segment Interpolate
+
+Same as EMA-VFI Interpolate but processes a single segment. Same pattern as BIM-VFI Segment Interpolate.
+
+**Output frame count (both models):** 2x = 2N-1, 4x = 4N-3, 8x = 8N-7
 
 ## Installation
 
@@ -40,7 +77,7 @@ cd ComfyUI/custom_nodes
 git clone https://github.com/your-user/Comfyui-BIM-VFI.git
 ```
 
-Dependencies (`gdown`, `cupy`) are auto-installed on first load. The correct `cupy` variant is detected from your PyTorch CUDA version.
+Dependencies (`gdown`, `cupy`, `timm`) are auto-installed on first load. The correct `cupy` variant is detected from your PyTorch CUDA version.
 
 > **Warning:** `cupy` is a large package (~800MB) and compilation/installation can take several minutes. The first ComfyUI startup after installing this node may appear to hang while `cupy` installs in the background. Check the console log for progress. If auto-install fails (e.g. missing build tools in Docker), install manually with:
 > ```bash
@@ -57,7 +94,8 @@ python install.py
 ### Requirements
 
 - PyTorch with CUDA
-- `cupy` (matching your CUDA version)
+- `cupy` (matching your CUDA version, for BIM-VFI)
+- `timm` (for EMA-VFI)
 - `gdown` (for model auto-download)
 
 ## VRAM Guide
@@ -71,9 +109,9 @@ python install.py
 
 ## Acknowledgments
 
-This project wraps the official [BiM-VFI](https://github.com/KAIST-VICLab/BiM-VFI) implementation by the [KAIST VIC Lab](https://github.com/KAIST-VICLab). The model architecture files in `bim_vfi_arch/` are vendored from their repository with minimal modifications (relative imports, inference-only paths).
+This project wraps the official [BiM-VFI](https://github.com/KAIST-VICLab/BiM-VFI) implementation by the [KAIST VIC Lab](https://github.com/KAIST-VICLab) and the official [EMA-VFI](https://github.com/MCG-NJU/EMA-VFI) implementation by MCG-NJU. Architecture files in `bim_vfi_arch/` and `ema_vfi_arch/` are vendored from their respective repositories with minimal modifications (relative imports, device-awareness fixes, inference-only paths).
 
-**Paper:**
+**BiM-VFI:**
 > Wonyong Seo, Jihyong Oh, and Munchurl Kim.
 > "BiM-VFI: Bidirectional Motion Field-Guided Frame Interpolation for Video with Non-uniform Motions."
 > *IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)*, 2025.
@@ -88,6 +126,23 @@ This project wraps the official [BiM-VFI](https://github.com/KAIST-VICLab/BiM-VF
 }
 ```
 
+**EMA-VFI:**
+> Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang.
+> "Extracting Motion and Appearance via Inter-Frame Attention for Efficient Video Frame Interpolation."
+> *IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)*, 2023.
+> [[arXiv]](https://arxiv.org/abs/2303.00440) [[GitHub]](https://github.com/MCG-NJU/EMA-VFI)
+
+```bibtex
+@inproceedings{zhang2023emavfi,
+  title={Extracting Motion and Appearance via Inter-Frame Attention for Efficient Video Frame Interpolation},
+  author={Zhang, Guozhen and Zhu, Yuhan and Wang, Haonan and Chen, Youxin and Wu, Gangshan and Wang, Limin},
+  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
+  year={2023}
+}
+```
+
 ## License
 
 The BiM-VFI model weights and architecture code are provided by KAIST VIC Lab for **research and education purposes only**. Commercial use requires permission from the principal investigator (Prof. Munchurl Kim, mkimee@kaist.ac.kr). See the [original repository](https://github.com/KAIST-VICLab/BiM-VFI) for details.
+
+The EMA-VFI model weights and architecture code are released under the [Apache 2.0 License](https://github.com/MCG-NJU/EMA-VFI/blob/main/LICENSE). See the [original repository](https://github.com/MCG-NJU/EMA-VFI) for details.

__init__.py

@@ -14,6 +14,13 @@ def _auto_install_deps():
         logger.info("[BIM-VFI] Installing gdown...")
         subprocess.check_call([sys.executable, "-m", "pip", "install", "gdown"])
 
+    # timm (required for EMA-VFI's MotionFormer backbone)
+    try:
+        import timm  # noqa: F401
+    except ImportError:
+        logger.info("[BIM-VFI] Installing timm...")
+        subprocess.check_call([sys.executable, "-m", "pip", "install", "timm"])
+
     # cupy
     try:
         import cupy  # noqa: F401
@@ -30,13 +37,19 @@ def _auto_install_deps():
 
 _auto_install_deps()
 
-from .nodes import LoadBIMVFIModel, BIMVFIInterpolate, BIMVFISegmentInterpolate, BIMVFIConcatVideos
+from .nodes import (
+    LoadBIMVFIModel, BIMVFIInterpolate, BIMVFISegmentInterpolate, BIMVFIConcatVideos,
+    LoadEMAVFIModel, EMAVFIInterpolate, EMAVFISegmentInterpolate,
+)
 
 NODE_CLASS_MAPPINGS = {
     "LoadBIMVFIModel": LoadBIMVFIModel,
     "BIMVFIInterpolate": BIMVFIInterpolate,
     "BIMVFISegmentInterpolate": BIMVFISegmentInterpolate,
     "BIMVFIConcatVideos": BIMVFIConcatVideos,
+    "LoadEMAVFIModel": LoadEMAVFIModel,
+    "EMAVFIInterpolate": EMAVFIInterpolate,
+    "EMAVFISegmentInterpolate": EMAVFISegmentInterpolate,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
@@ -44,4 +57,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "BIMVFIInterpolate": "BIM-VFI Interpolate",
     "BIMVFISegmentInterpolate": "BIM-VFI Segment Interpolate",
     "BIMVFIConcatVideos": "BIM-VFI Concat Videos",
+    "LoadEMAVFIModel": "Load EMA-VFI Model",
+    "EMAVFIInterpolate": "EMA-VFI Interpolate",
+    "EMAVFISegmentInterpolate": "EMA-VFI Segment Interpolate",
 }

ema_vfi_arch/__init__.py

@@ -0,0 +1,5 @@
+from .feature_extractor import feature_extractor
+from .flow_estimation import MultiScaleFlow
+from .warplayer import clear_warp_cache
+
+__all__ = ['feature_extractor', 'MultiScaleFlow', 'clear_warp_cache']

ema_vfi_arch/feature_extractor.py

@@ -0,0 +1,515 @@
+import torch
+import torch.nn as nn
+import math
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+def window_partition(x, window_size):
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size[0], window_size[0], W // window_size[1], window_size[1], C)
+    windows = (
+        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size[0]*window_size[1], C)
+    )
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    nwB, N, C = windows.shape
+    windows = windows.view(-1, window_size[0], window_size[1], C)
+    B = int(nwB / (H * W / window_size[0] / window_size[1]))
+    x = windows.view(
+        B, H // window_size[0], W // window_size[1], window_size[0], window_size[1], -1
+    )
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+def pad_if_needed(x, size, window_size):
+    n, h, w, c = size
+    pad_h = math.ceil(h / window_size[0]) * window_size[0] - h
+    pad_w = math.ceil(w / window_size[1]) * window_size[1] - w
+    if pad_h > 0 or pad_w > 0:  # center-pad the feature on H and W axes
+        img_mask = torch.zeros((1, h+pad_h, w+pad_w, 1))  # 1 H W 1
+        h_slices = (
+            slice(0, pad_h//2),
+            slice(pad_h//2, h+pad_h//2),
+            slice(h+pad_h//2, None),
+        )
+        w_slices = (
+            slice(0, pad_w//2),
+            slice(pad_w//2, w+pad_w//2),
+            slice(w+pad_w//2, None),
+        )
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(
+            img_mask, window_size
+        )  # nW, window_size*window_size, 1
+        mask_windows = mask_windows.squeeze(-1)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(
+            attn_mask != 0, float(-100.0)
+        ).masked_fill(attn_mask == 0, float(0.0))
+        return nn.functional.pad(
+            x,
+            (0, 0, pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2),
+        ), attn_mask
+    return x, None
+
+
+def depad_if_needed(x, size, window_size):
+    n, h, w, c = size
+    pad_h = math.ceil(h / window_size[0]) * window_size[0] - h
+    pad_w = math.ceil(w / window_size[1]) * window_size[1] - w
+    if pad_h > 0 or pad_w > 0:  # remove the center-padding on feature
+        return x[:, pad_h // 2 : pad_h // 2 + h, pad_w // 2 : pad_w // 2 + w, :].contiguous()
+    return x
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.dwconv = DWConv(hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+        self.relu = nn.ReLU(inplace=True)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, H, W):
+        x = self.fc1(x)
+        x = self.dwconv(x, H, W)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class InterFrameAttention(nn.Module):
+    def __init__(self, dim, motion_dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
+        super().__init__()
+        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
+
+        self.dim = dim
+        self.motion_dim = motion_dim
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = nn.Linear(dim, dim, bias=qkv_bias)
+        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
+        self.cor_embed = nn.Linear(2, motion_dim, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.motion_proj = nn.Linear(motion_dim, motion_dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x1, x2, cor, H, W, mask=None):
+        B, N, C = x1.shape
+        B, N, C_c = cor.shape
+        q = self.q(x1).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+        kv = self.kv(x2).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        cor_embed_ = self.cor_embed(cor)
+        cor_embed = cor_embed_.reshape(B, N, self.num_heads, self.motion_dim // self.num_heads).permute(0, 2, 1, 3)
+        k, v = kv[0], kv[1]
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+
+        if mask is not None:
+            nW = mask.shape[0] # mask: nW, N, N
+            attn = attn.view(B // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
+                1
+            ).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = attn.softmax(dim=-1)
+        else:
+            attn = attn.softmax(dim=-1)
+
+        attn = self.attn_drop(attn)
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        c_reverse = (attn @ cor_embed).transpose(1, 2).reshape(B, N, -1)
+        motion = self.motion_proj(c_reverse-cor_embed_)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x, motion
+
+
+class MotionFormerBlock(nn.Module):
+    def __init__(self, dim, motion_dim, num_heads, window_size=0, shift_size=0, mlp_ratio=4., bidirectional=True, qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm,):
+        super().__init__()
+        self.window_size = window_size
+        if not isinstance(self.window_size, (tuple, list)):
+            self.window_size = to_2tuple(window_size)
+        self.shift_size = shift_size
+        if not isinstance(self.shift_size, (tuple, list)):
+            self.shift_size = to_2tuple(shift_size)
+        self.bidirectional = bidirectional
+        self.norm1 = norm_layer(dim)
+        self.attn = InterFrameAttention(
+            dim,
+            motion_dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop)
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x, cor, H, W, B):
+        x = x.view(2*B, H, W, -1)
+        x_pad, mask = pad_if_needed(x, x.size(), self.window_size)
+        cor_pad, _ = pad_if_needed(cor, cor.size(), self.window_size)
+
+        if self.shift_size[0] or self.shift_size[1]:
+            _, H_p, W_p, C = x_pad.shape
+            x_pad = torch.roll(x_pad, shifts=(-self.shift_size[0], -self.shift_size[1]), dims=(1, 2))
+            cor_pad = torch.roll(cor_pad, shifts=(-self.shift_size[0], -self.shift_size[1]), dims=(1, 2))
+
+            if hasattr(self, 'HW') and self.HW.item() == H_p * W_p:
+                shift_mask = self.attn_mask
+            else:
+                shift_mask = torch.zeros((1, H_p, W_p, 1))  # 1 H W 1
+                h_slices = (slice(0, -self.window_size[0]),
+                            slice(-self.window_size[0], -self.shift_size[0]),
+                            slice(-self.shift_size[0], None))
+                w_slices = (slice(0, -self.window_size[1]),
+                            slice(-self.window_size[1], -self.shift_size[1]),
+                            slice(-self.shift_size[1], None))
+                cnt = 0
+                for h in h_slices:
+                    for w in w_slices:
+                        shift_mask[:, h, w, :] = cnt
+                        cnt += 1
+
+                mask_windows = window_partition(shift_mask, self.window_size).squeeze(-1)
+                shift_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+                shift_mask = shift_mask.masked_fill(shift_mask != 0,
+                                float(-100.0)).masked_fill(shift_mask == 0,
+                                float(0.0))
+
+                if mask is not None:
+                    shift_mask = shift_mask.masked_fill(mask != 0,
+                                float(-100.0))
+                self.register_buffer("attn_mask", shift_mask)
+                self.register_buffer("HW", torch.Tensor([H_p*W_p]))
+        else:
+            shift_mask = mask
+
+        if shift_mask is not None:
+            shift_mask = shift_mask.to(x_pad.device)
+
+
+        _, Hw, Ww, C = x_pad.shape
+        x_win = window_partition(x_pad, self.window_size)
+        cor_win = window_partition(cor_pad, self.window_size)
+
+        nwB = x_win.shape[0]
+        x_norm = self.norm1(x_win)
+
+        x_reverse = torch.cat([x_norm[nwB//2:], x_norm[:nwB//2]])
+        x_appearence, x_motion = self.attn(x_norm, x_reverse, cor_win, H, W, shift_mask)
+        x_norm = x_norm + self.drop_path(x_appearence)
+
+        x_back = x_norm
+        x_back_win = window_reverse(x_back, self.window_size, Hw, Ww)
+        x_motion = window_reverse(x_motion, self.window_size, Hw, Ww)
+
+        if self.shift_size[0] or self.shift_size[1]:
+            x_back_win = torch.roll(x_back_win, shifts=(self.shift_size[0], self.shift_size[1]), dims=(1, 2))
+            x_motion = torch.roll(x_motion, shifts=(self.shift_size[0], self.shift_size[1]), dims=(1, 2))
+
+        x = depad_if_needed(x_back_win, x.size(), self.window_size).view(2*B, H * W, -1)
+        x_motion = depad_if_needed(x_motion, cor.size(), self.window_size).view(2*B, H * W, -1)
+
+        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
+        return x, x_motion
+
+
+class ConvBlock(nn.Module):
+    def __init__(self, in_dim, out_dim, depths=2,act_layer=nn.PReLU):
+        super().__init__()
+        layers = []
+        for i in range(depths):
+            if i == 0:
+                layers.append(nn.Conv2d(in_dim, out_dim, 3,1,1))
+            else:
+                layers.append(nn.Conv2d(out_dim, out_dim, 3,1,1))
+            layers.extend([
+                act_layer(out_dim),
+            ])
+        self.conv = nn.Sequential(*layers)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x):
+        x = self.conv(x)
+        return x
+
+
+class OverlapPatchEmbed(nn.Module):
+    def __init__(self, patch_size=7, stride=4, in_chans=3, embed_dim=768):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+
+        self.patch_size = patch_size
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=stride,
+                              padding=(patch_size[0] // 2, patch_size[1] // 2))
+        self.norm = nn.LayerNorm(embed_dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, x):
+        x = self.proj(x)
+        _, _, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+
+        return x, H, W
+
+
+class CrossScalePatchEmbed(nn.Module):
+    def __init__(self, in_dims=[16,32,64], embed_dim=768):
+        super().__init__()
+        base_dim = in_dims[0]
+
+        layers = []
+        for i in range(len(in_dims)):
+            for j in range(2 ** i):
+                layers.append(nn.Conv2d(in_dims[-1-i], base_dim, 3, 2**(i+1), 1+j, 1+j))
+        self.layers = nn.ModuleList(layers)
+        self.proj = nn.Conv2d(base_dim * len(layers), embed_dim, 1, 1)
+        self.norm = nn.LayerNorm(embed_dim)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, xs):
+        ys = []
+        k = 0
+        for i in range(len(xs)):
+            for _ in range(2 ** i):
+                ys.append(self.layers[k](xs[-1-i]))
+                k += 1
+        x = self.proj(torch.cat(ys,1))
+        _, _, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+
+        return x, H, W
+
+
+class MotionFormer(nn.Module):
+    def __init__(self, in_chans=3, embed_dims=[32, 64, 128, 256, 512], motion_dims=64, num_heads=[8, 16],
+                 mlp_ratios=[4, 4], qkv_bias=True, qk_scale=None, drop_rate=0.,
+                 attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
+                 depths=[2, 2, 2, 6, 2], window_sizes=[11, 11],**kwarg):
+        super().__init__()
+        self.depths = depths
+        self.num_stages = len(embed_dims)
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+        cur = 0
+
+        self.conv_stages = self.num_stages - len(num_heads)
+
+        for i in range(self.num_stages):
+            if i == 0:
+                block = ConvBlock(in_chans,embed_dims[i],depths[i])
+            else:
+                if i < self.conv_stages:
+                    patch_embed = nn.Sequential(
+                        nn.Conv2d(embed_dims[i-1], embed_dims[i], 3,2,1),
+                        nn.PReLU(embed_dims[i])
+                    )
+                    block = ConvBlock(embed_dims[i],embed_dims[i],depths[i])
+                else:
+                    if i == self.conv_stages:
+                        patch_embed = CrossScalePatchEmbed(embed_dims[:i],
+                                                        embed_dim=embed_dims[i])
+                    else:
+                        patch_embed = OverlapPatchEmbed(patch_size=3,
+                                                        stride=2,
+                                                        in_chans=embed_dims[i - 1],
+                                                        embed_dim=embed_dims[i])
+
+                    block = nn.ModuleList([MotionFormerBlock(
+                        dim=embed_dims[i], motion_dim=motion_dims[i], num_heads=num_heads[i-self.conv_stages], window_size=window_sizes[i-self.conv_stages],
+                        shift_size= 0 if (j % 2) == 0 else window_sizes[i-self.conv_stages] // 2,
+                        mlp_ratio=mlp_ratios[i-self.conv_stages], qkv_bias=qkv_bias, qk_scale=qk_scale,
+                        drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + j], norm_layer=norm_layer)
+                        for j in range(depths[i])])
+
+                    norm = norm_layer(embed_dims[i])
+                    setattr(self, f"norm{i + 1}", norm)
+                setattr(self, f"patch_embed{i + 1}", patch_embed)
+            cur += depths[i]
+
+            setattr(self, f"block{i + 1}", block)
+
+        self.cor = {}
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def get_cor(self, shape, device):
+        k = (str(shape), str(device))
+        if k not in self.cor:
+            tenHorizontal = torch.linspace(-1.0, 1.0, shape[2], device=device).view(
+                1, 1, 1, shape[2]).expand(shape[0], -1, shape[1], -1).permute(0, 2, 3, 1)
+            tenVertical = torch.linspace(-1.0, 1.0, shape[1], device=device).view(
+                1, 1, shape[1], 1).expand(shape[0], -1, -1, shape[2]).permute(0, 2, 3, 1)
+            self.cor[k] = torch.cat([tenHorizontal, tenVertical], -1).to(device)
+        return self.cor[k]
+
+    def forward(self, x1, x2):
+        B = x1.shape[0]
+        x = torch.cat([x1, x2], 0)
+        motion_features = []
+        appearence_features = []
+        xs = []
+        for i in range(self.num_stages):
+            motion_features.append([])
+            patch_embed = getattr(self, f"patch_embed{i + 1}",None)
+            block = getattr(self, f"block{i + 1}",None)
+            norm = getattr(self, f"norm{i + 1}",None)
+            if i < self.conv_stages:
+                if i > 0:
+                    x = patch_embed(x)
+                x = block(x)
+                xs.append(x)
+            else:
+                if i == self.conv_stages:
+                    x, H, W = patch_embed(xs)
+                else:
+                    x, H, W = patch_embed(x)
+                cor = self.get_cor((x.shape[0], H, W), x.device)
+                for blk in block:
+                    x, x_motion = blk(x, cor, H, W, B)
+                    motion_features[i].append(x_motion.reshape(2*B, H, W, -1).permute(0, 3, 1, 2).contiguous())
+                x = norm(x)
+                x = x.reshape(2*B, H, W, -1).permute(0, 3, 1, 2).contiguous()
+                motion_features[i] = torch.cat(motion_features[i], 1)
+            appearence_features.append(x)
+        return appearence_features, motion_features
+
+
+class DWConv(nn.Module):
+    def __init__(self, dim):
+        super(DWConv, self).__init__()
+        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
+
+    def forward(self, x, H, W):
+        B, N, C = x.shape
+        x = x.transpose(1, 2).reshape(B, C, H, W)
+        x = self.dwconv(x)
+        x = x.reshape(B, C, -1).transpose(1, 2)
+
+        return x
+
+
+def feature_extractor(**kargs):
+    model = MotionFormer(**kargs)
+    return model

ema_vfi_arch/flow_estimation.py

@@ -0,0 +1,141 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .warplayer import warp
+from .refine import *
+
+
+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)
+    )
+
+
+class Head(nn.Module):
+    def __init__(self, in_planes, scale, c, in_else=17):
+        super(Head, self).__init__()
+        self.upsample = nn.Sequential(nn.PixelShuffle(2), nn.PixelShuffle(2))
+        self.scale = scale
+        self.conv = nn.Sequential(
+                                  conv(in_planes*2 // (4*4) + in_else, c),
+                                  conv(c, c),
+                                  conv(c, 5),
+                                  )
+
+    def forward(self, motion_feature, x, flow): # /16 /8 /4
+        motion_feature = self.upsample(motion_feature) #/4 /2 /1
+        if self.scale != 4:
+            x = F.interpolate(x, scale_factor = 4. / self.scale, mode="bilinear", align_corners=False)
+        if flow != None:
+            if self.scale != 4:
+                flow = F.interpolate(flow, scale_factor = 4. / self.scale, mode="bilinear", align_corners=False) * 4. / self.scale
+            x = torch.cat((x, flow), 1)
+        x = self.conv(torch.cat([motion_feature, x], 1))
+        if self.scale != 4:
+            x = F.interpolate(x, scale_factor = self.scale // 4, mode="bilinear", align_corners=False)
+            flow = x[:, :4] * (self.scale // 4)
+        else:
+            flow = x[:, :4]
+        mask = x[:, 4:5]
+        return flow, mask
+
+
+class MultiScaleFlow(nn.Module):
+    def __init__(self, backbone, **kargs):
+        super(MultiScaleFlow, self).__init__()
+        self.flow_num_stage = len(kargs['hidden_dims'])
+        self.feature_bone = backbone
+        self.block = nn.ModuleList([Head( kargs['motion_dims'][-1-i] * kargs['depths'][-1-i] + kargs['embed_dims'][-1-i],
+                            kargs['scales'][-1-i],
+                            kargs['hidden_dims'][-1-i],
+                            6 if i==0 else 17)
+                            for i in range(self.flow_num_stage)])
+        self.unet = Unet(kargs['c'] * 2)
+
+    def warp_features(self, xs, flow):
+        y0 = []
+        y1 = []
+        B = xs[0].size(0) // 2
+        for x in xs:
+            y0.append(warp(x[:B], flow[:, 0:2]))
+            y1.append(warp(x[B:], flow[:, 2:4]))
+            flow = F.interpolate(flow, scale_factor=0.5, mode="bilinear", align_corners=False, recompute_scale_factor=False) * 0.5
+        return y0, y1
+
+    def calculate_flow(self, imgs, timestep, af=None, mf=None):
+        img0, img1 = imgs[:, :3], imgs[:, 3:6]
+        B = img0.size(0)
+        flow, mask = None, None
+        # appearence_features & motion_features
+        if (af is None) or (mf is None):
+            af, mf = self.feature_bone(img0, img1)
+        for i in range(self.flow_num_stage):
+            t = torch.full(mf[-1-i][:B].shape, timestep, dtype=torch.float, device=imgs.device)
+            if flow != None:
+                warped_img0 = warp(img0, flow[:, :2])
+                warped_img1 = warp(img1, flow[:, 2:4])
+                flow_, mask_ = self.block[i](
+                    torch.cat([t*mf[-1-i][:B],(1-t)*mf[-1-i][B:],af[-1-i][:B],af[-1-i][B:]],1),
+                    torch.cat((img0, img1, warped_img0, warped_img1, mask), 1),
+                    flow
+                    )
+                flow = flow + flow_
+                mask = mask + mask_
+            else:
+                flow, mask = self.block[i](
+                    torch.cat([t*mf[-1-i][:B],(1-t)*mf[-1-i][B:],af[-1-i][:B],af[-1-i][B:]],1),
+                    torch.cat((img0, img1), 1),
+                    None
+                    )
+
+        return flow, mask
+
+    def coraseWarp_and_Refine(self, imgs, af, flow, mask):
+        img0, img1 = imgs[:, :3], imgs[:, 3:6]
+        warped_img0 = warp(img0, flow[:, :2])
+        warped_img1 = warp(img1, flow[:, 2:4])
+        c0, c1 = self.warp_features(af, flow)
+        tmp = self.unet(img0, img1, warped_img0, warped_img1, mask, flow, c0, c1)
+        res = tmp[:, :3] * 2 - 1
+        mask_ = torch.sigmoid(mask)
+        merged = warped_img0 * mask_ + warped_img1 * (1 - mask_)
+        pred = torch.clamp(merged + res, 0, 1)
+        return pred
+
+
+    # Actually consist of 'calculate_flow' and 'coraseWarp_and_Refine'
+    def forward(self, x, timestep=0.5):
+        img0, img1 = x[:, :3], x[:, 3:6]
+        B = x.size(0)
+        flow_list = []
+        merged = []
+        mask_list = []
+        warped_img0 = img0
+        warped_img1 = img1
+        flow = None
+        # appearence_features & motion_features
+        af, mf = self.feature_bone(img0, img1)
+        for i in range(self.flow_num_stage):
+            t = torch.full(mf[-1-i][:B].shape, timestep, dtype=torch.float, device=x.device)
+            if flow != None:
+                flow_d, mask_d = self.block[i]( torch.cat([t*mf[-1-i][:B], (1-timestep)*mf[-1-i][B:],af[-1-i][:B],af[-1-i][B:]],1),
+                                                torch.cat((img0, img1, warped_img0, warped_img1, mask), 1), flow)
+                flow = flow + flow_d
+                mask = mask + mask_d
+            else:
+                flow, mask = self.block[i]( torch.cat([t*mf[-1-i][:B], (1-t)*mf[-1-i][B:],af[-1-i][:B],af[-1-i][B:]],1),
+                                            torch.cat((img0, img1), 1), None)
+            mask_list.append(torch.sigmoid(mask))
+            flow_list.append(flow)
+            warped_img0 = warp(img0, flow[:, :2])
+            warped_img1 = warp(img1, flow[:, 2:4])
+            merged.append(warped_img0 * mask_list[i] + warped_img1 * (1 - mask_list[i]))
+
+        c0, c1 = self.warp_features(af, flow)
+        tmp = self.unet(img0, img1, warped_img0, warped_img1, mask, flow, c0, c1)
+        res = tmp[:, :3] * 2 - 1
+        pred = torch.clamp(merged[-1] + res, 0, 1)
+        return flow_list, mask_list, merged, pred

ema_vfi_arch/refine.py

@@ -0,0 +1,70 @@
+import torch
+import torch.nn as nn
+import math
+from timm.models.layers import trunc_normal_
+
+
+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 deconv(in_planes, out_planes, kernel_size=4, stride=2, padding=1):
+    return nn.Sequential(
+        torch.nn.ConvTranspose2d(in_channels=in_planes, out_channels=out_planes, kernel_size=4, stride=2, padding=1, bias=True),
+        nn.PReLU(out_planes)
+        )
+
+class Conv2(nn.Module):
+    def __init__(self, in_planes, out_planes, stride=2):
+        super(Conv2, self).__init__()
+        self.conv1 = conv(in_planes, out_planes, 3, stride, 1)
+        self.conv2 = conv(out_planes, out_planes, 3, 1, 1)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.conv2(x)
+        return x
+
+class Unet(nn.Module):
+    def __init__(self, c, out=3):
+        super(Unet, self).__init__()
+        self.down0 = Conv2(17+c, 2*c)
+        self.down1 = Conv2(4*c, 4*c)
+        self.down2 = Conv2(8*c, 8*c)
+        self.down3 = Conv2(16*c, 16*c)
+        self.up0 = deconv(32*c, 8*c)
+        self.up1 = deconv(16*c, 4*c)
+        self.up2 = deconv(8*c, 2*c)
+        self.up3 = deconv(4*c, c)
+        self.conv = nn.Conv2d(c, out, 3, 1, 1)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+        elif isinstance(m, nn.Conv2d):
+            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+            fan_out //= m.groups
+            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
+            if m.bias is not None:
+                m.bias.data.zero_()
+
+    def forward(self, img0, img1, warped_img0, warped_img1, mask, flow, c0, c1):
+        s0 = self.down0(torch.cat((img0, img1, warped_img0, warped_img1, mask, flow,c0[0], c1[0]), 1))
+        s1 = self.down1(torch.cat((s0, c0[1], c1[1]), 1))
+        s2 = self.down2(torch.cat((s1, c0[2], c1[2]), 1))
+        s3 = self.down3(torch.cat((s2, c0[3], c1[3]), 1))
+        x = self.up0(torch.cat((s3, c0[4], c1[4]), 1))
+        x = self.up1(torch.cat((x, s2), 1))
+        x = self.up2(torch.cat((x, s1), 1))
+        x = self.up3(torch.cat((x, s0), 1))
+        x = self.conv(x)
+        return torch.sigmoid(x)

ema_vfi_arch/warplayer.py

@@ -0,0 +1,25 @@
+import torch
+
+backwarp_tenGrid = {}
+
+
+def clear_warp_cache():
+    """Free all cached grid tensors (call between frame pairs to reclaim VRAM)."""
+    backwarp_tenGrid.clear()
+
+
+def warp(tenInput, tenFlow):
+    k = (str(tenFlow.device), str(tenFlow.size()))
+    if k not in backwarp_tenGrid:
+        tenHorizontal = torch.linspace(-1.0, 1.0, tenFlow.shape[3], device=tenFlow.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=tenFlow.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(tenFlow.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)

inference.py

@@ -1,5 +1,15 @@
+import logging
+from functools import partial
+
 import torch
+import torch.nn as nn
+
 from .bim_vfi_arch import BiMVFI
+from .ema_vfi_arch import feature_extractor as ema_feature_extractor
+from .ema_vfi_arch import MultiScaleFlow as EMAMultiScaleFlow
+from .utils.padder import InputPadder
+
+logger = logging.getLogger("BIM-VFI")
 
 
 class BiMVFIModel:
@@ -112,3 +122,163 @@ class BiMVFIModel:
         interp = result_dict["imgt_pred"]
         interp = torch.clamp(interp, 0, 1)
         return interp
+
+
+# ---------------------------------------------------------------------------
+# EMA-VFI model wrapper
+# ---------------------------------------------------------------------------
+
+def _ema_init_model_config(F=32, W=7, depth=[2, 2, 2, 4, 4]):
+    """Build EMA-VFI model config dicts (backbone + multiscale)."""
+    return {
+        'embed_dims': [F, 2*F, 4*F, 8*F, 16*F],
+        'motion_dims': [0, 0, 0, 8*F//depth[-2], 16*F//depth[-1]],
+        'num_heads': [8*F//32, 16*F//32],
+        'mlp_ratios': [4, 4],
+        'qkv_bias': True,
+        'norm_layer': partial(nn.LayerNorm, eps=1e-6),
+        'depths': depth,
+        'window_sizes': [W, W]
+    }, {
+        'embed_dims': [F, 2*F, 4*F, 8*F, 16*F],
+        'motion_dims': [0, 0, 0, 8*F//depth[-2], 16*F//depth[-1]],
+        'depths': depth,
+        'num_heads': [8*F//32, 16*F//32],
+        'window_sizes': [W, W],
+        'scales': [4, 8, 16],
+        'hidden_dims': [4*F, 4*F],
+        'c': F
+    }
+
+
+def _ema_detect_variant(filename):
+    """Auto-detect model variant and timestep support from filename.
+
+    Returns (F, depth, supports_arbitrary_t).
+    """
+    name = filename.lower()
+    is_small = "small" in name
+    supports_t = "_t." in name or "_t_" in name or name.endswith("_t")
+
+    if is_small:
+        return 16, [2, 2, 2, 2, 2], supports_t
+    else:
+        return 32, [2, 2, 2, 4, 4], supports_t
+
+
+class EMAVFIModel:
+    """Clean inference wrapper around EMA-VFI for ComfyUI integration."""
+
+    def __init__(self, checkpoint_path, variant="auto", tta=False, device="cpu"):
+        import os
+        filename = os.path.basename(checkpoint_path)
+
+        if variant == "auto":
+            F_dim, depth, self.supports_arbitrary_t = _ema_detect_variant(filename)
+        elif variant == "small":
+            F_dim, depth = 16, [2, 2, 2, 2, 2]
+            self.supports_arbitrary_t = "_t." in filename.lower() or "_t_" in filename.lower()
+        else:  # large
+            F_dim, depth = 32, [2, 2, 2, 4, 4]
+            self.supports_arbitrary_t = "_t." in filename.lower() or "_t_" in filename.lower()
+
+        self.tta = tta
+        self.device = device
+        self.variant_name = "small" if F_dim == 16 else "large"
+
+        backbone_cfg, multiscale_cfg = _ema_init_model_config(F=F_dim, depth=depth)
+        backbone = ema_feature_extractor(**backbone_cfg)
+        self.model = EMAMultiScaleFlow(backbone, **multiscale_cfg)
+        self._load_checkpoint(checkpoint_path)
+        self.model.eval()
+        self.model.to(device)
+
+    def _load_checkpoint(self, checkpoint_path):
+        """Load checkpoint with module prefix stripping and buffer filtering."""
+        state_dict = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
+
+        # Handle wrapped checkpoint formats
+        if isinstance(state_dict, dict):
+            if "model" in state_dict:
+                state_dict = state_dict["model"]
+            elif "state_dict" in state_dict:
+                state_dict = state_dict["state_dict"]
+
+        # Strip "module." prefix and filter out attn_mask/HW buffers
+        cleaned = {}
+        for k, v in state_dict.items():
+            if "attn_mask" in k or k.endswith(".HW"):
+                continue
+            key = k
+            if key.startswith("module."):
+                key = key[len("module."):]
+            cleaned[key] = v
+
+        self.model.load_state_dict(cleaned)
+
+    def to(self, device):
+        """Move model to device (returns self for chaining)."""
+        self.device = device
+        self.model.to(device)
+        return self
+
+    @torch.no_grad()
+    def _inference(self, img0, img1, timestep=0.5):
+        """Run single inference pass. Inputs already padded, on device."""
+        B = img0.shape[0]
+        imgs = torch.cat((img0, img1), 1)
+
+        if self.tta:
+            imgs_ = imgs.flip(2).flip(3)
+            input_batch = torch.cat((imgs, imgs_), 0)
+            _, _, _, preds = self.model(input_batch, timestep=timestep)
+            return (preds[:B] + preds[B:].flip(2).flip(3)) / 2.
+        else:
+            _, _, _, pred = self.model(imgs, timestep=timestep)
+            return pred
+
+    @torch.no_grad()
+    def interpolate_pair(self, frame0, frame1, time_step=0.5):
+        """Interpolate a single frame between two input frames.
+
+        Args:
+            frame0: [1, C, H, W] tensor, float32, range [0, 1]
+            frame1: [1, C, H, W] tensor, float32, range [0, 1]
+            time_step: float in (0, 1)
+
+        Returns:
+            Interpolated frame as [1, C, H, W] tensor, float32, clamped to [0, 1]
+        """
+        device = next(self.model.parameters()).device
+        img0 = frame0.to(device)
+        img1 = frame1.to(device)
+
+        padder = InputPadder(img0.shape, divisor=32, mode='replicate', center=True)
+        img0, img1 = padder.pad(img0, img1)
+
+        pred = self._inference(img0, img1, timestep=time_step)
+        pred = padder.unpad(pred)
+        return torch.clamp(pred, 0, 1)
+
+    @torch.no_grad()
+    def interpolate_batch(self, frames0, frames1, time_step=0.5):
+        """Interpolate multiple frame pairs at once.
+
+        Args:
+            frames0: [B, C, H, W] tensor, float32, range [0, 1]
+            frames1: [B, C, H, W] tensor, float32, range [0, 1]
+            time_step: float in (0, 1)
+
+        Returns:
+            Interpolated frames as [B, C, H, W] tensor, float32, clamped to [0, 1]
+        """
+        device = next(self.model.parameters()).device
+        img0 = frames0.to(device)
+        img1 = frames1.to(device)
+
+        padder = InputPadder(img0.shape, divisor=32, mode='replicate', center=True)
+        img0, img1 = padder.pad(img0, img1)
+
+        pred = self._inference(img0, img1, timestep=time_step)
+        pred = padder.unpad(pred)
+        return torch.clamp(pred, 0, 1)

nodes.py

@@ -8,20 +8,29 @@ import torch
 import folder_paths
 from comfy.utils import ProgressBar
 
-from .inference import BiMVFIModel
+from .inference import BiMVFIModel, EMAVFIModel
 from .bim_vfi_arch import clear_backwarp_cache
+from .ema_vfi_arch import clear_warp_cache as clear_ema_warp_cache
 
 logger = logging.getLogger("BIM-VFI")
 
-# Google Drive file ID for the pretrained model
+# Google Drive file ID for the pretrained BIM-VFI model
 GDRIVE_FILE_ID = "18Wre7XyRtu_wtFRzcsit6oNfHiFRt9vC"
 MODEL_FILENAME = "bim_vfi.pth"
 
-# Register the model folder with ComfyUI
+# Google Drive folder ID for EMA-VFI pretrained models
+EMA_GDRIVE_FOLDER_ID = "16jUa3HkQ85Z5lb5gce1yoaWkP-rdCd0o"
+EMA_DEFAULT_MODEL = "ours_t.pkl"
+
+# Register model folders with ComfyUI
 MODEL_DIR = os.path.join(folder_paths.models_dir, "bim-vfi")
 if not os.path.exists(MODEL_DIR):
     os.makedirs(MODEL_DIR, exist_ok=True)
 
+EMA_MODEL_DIR = os.path.join(folder_paths.models_dir, "ema-vfi")
+if not os.path.exists(EMA_MODEL_DIR):
+    os.makedirs(EMA_MODEL_DIR, exist_ok=True)
+
 
 def get_available_models():
     """List available checkpoint files in the bim-vfi model directory."""
@@ -456,3 +465,305 @@ class BIMVFIConcatVideos:
                 os.remove(concat_list_path)
 
         return (output_path,)
+
+
+# ---------------------------------------------------------------------------
+# EMA-VFI nodes
+# ---------------------------------------------------------------------------
+
+def get_available_ema_models():
+    """List available checkpoint files in the ema-vfi model directory."""
+    models = []
+    if os.path.isdir(EMA_MODEL_DIR):
+        for f in os.listdir(EMA_MODEL_DIR):
+            if f.endswith((".pkl", ".pth", ".pt", ".ckpt", ".safetensors")):
+                models.append(f)
+    if not models:
+        models.append(EMA_DEFAULT_MODEL)  # Will trigger auto-download
+    return sorted(models)
+
+
+def download_ema_model_from_gdrive(folder_id, dest_path):
+    """Download EMA-VFI model from Google Drive folder using gdown."""
+    try:
+        import gdown
+    except ImportError:
+        raise RuntimeError(
+            "gdown is required to auto-download the EMA-VFI model. "
+            "Install it with: pip install gdown"
+        )
+    filename = os.path.basename(dest_path)
+    url = f"https://drive.google.com/drive/folders/{folder_id}"
+    logger.info(f"Downloading {filename} from Google Drive folder to {dest_path}...")
+    os.makedirs(os.path.dirname(dest_path), exist_ok=True)
+    gdown.download_folder(url, output=os.path.dirname(dest_path), quiet=False, remaining_ok=True)
+    if not os.path.exists(dest_path):
+        raise RuntimeError(
+            f"Failed to download {filename}. Please download manually from "
+            f"https://drive.google.com/drive/folders/{folder_id} "
+            f"and place it in {os.path.dirname(dest_path)}"
+        )
+    logger.info("Download complete.")
+
+
+class LoadEMAVFIModel:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "model_path": (get_available_ema_models(), {
+                    "default": EMA_DEFAULT_MODEL,
+                    "tooltip": "Checkpoint file from models/ema-vfi/. Auto-downloads on first use if missing. "
+                               "Variant (large/small) and timestep support are auto-detected from filename.",
+                }),
+                "tta": ("BOOLEAN", {
+                    "default": False,
+                    "tooltip": "Test-time augmentation: flip input and average with unflipped result. "
+                               "~2x slower but slightly better quality. Recommended for large model only.",
+                }),
+            }
+        }
+
+    RETURN_TYPES = ("EMA_VFI_MODEL",)
+    RETURN_NAMES = ("model",)
+    FUNCTION = "load_model"
+    CATEGORY = "video/EMA-VFI"
+
+    def load_model(self, model_path, tta):
+        full_path = os.path.join(EMA_MODEL_DIR, model_path)
+
+        if not os.path.exists(full_path):
+            logger.info(f"Model not found at {full_path}, attempting download...")
+            download_ema_model_from_gdrive(EMA_GDRIVE_FOLDER_ID, full_path)
+
+        wrapper = EMAVFIModel(
+            checkpoint_path=full_path,
+            variant="auto",
+            tta=tta,
+            device="cpu",
+        )
+
+        t_mode = "arbitrary" if wrapper.supports_arbitrary_t else "fixed (0.5)"
+        logger.info(f"EMA-VFI model loaded (variant={wrapper.variant_name}, timestep={t_mode}, tta={tta})")
+        return (wrapper,)
+
+
+class EMAVFIInterpolate:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "images": ("IMAGE", {
+                    "tooltip": "Input image batch. Output frame count: 2x=(2N-1), 4x=(4N-3), 8x=(8N-7).",
+                }),
+                "model": ("EMA_VFI_MODEL", {
+                    "tooltip": "EMA-VFI model from the Load EMA-VFI Model node.",
+                }),
+                "multiplier": ([2, 4, 8], {
+                    "default": 2,
+                    "tooltip": "Frame rate multiplier. 2x=one interpolation pass, 4x=two recursive passes, 8x=three. Higher = more frames but longer processing.",
+                }),
+                "clear_cache_after_n_frames": ("INT", {
+                    "default": 10, "min": 1, "max": 100, "step": 1,
+                    "tooltip": "Clear CUDA cache every N frame pairs to prevent VRAM buildup. Lower = less VRAM but slower. Ignored when all_on_gpu is enabled.",
+                }),
+                "keep_device": ("BOOLEAN", {
+                    "default": True,
+                    "tooltip": "Keep model on GPU between frame pairs. Faster but uses more VRAM constantly. Disable to free VRAM between pairs (slower due to CPU-GPU transfers).",
+                }),
+                "all_on_gpu": ("BOOLEAN", {
+                    "default": False,
+                    "tooltip": "Store all intermediate frames on GPU instead of CPU. Much faster (no transfers) but requires enough VRAM for all frames. Recommended for 48GB+ cards.",
+                }),
+                "batch_size": ("INT", {
+                    "default": 1, "min": 1, "max": 64, "step": 1,
+                    "tooltip": "Number of frame pairs to process simultaneously. Higher = faster but uses more VRAM. Start with 1, increase until VRAM is full.",
+                }),
+                "chunk_size": ("INT", {
+                    "default": 0, "min": 0, "max": 10000, "step": 1,
+                    "tooltip": "Process input frames in chunks of this size (0=disabled). Bounds VRAM usage during processing but the full output is still assembled in RAM. To bound RAM, use the Segment Interpolate node instead.",
+                }),
+            }
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    RETURN_NAMES = ("images",)
+    FUNCTION = "interpolate"
+    CATEGORY = "video/EMA-VFI"
+
+    def _interpolate_frames(self, frames, model, num_passes, batch_size,
+                            device, storage_device, keep_device, all_on_gpu,
+                            clear_cache_after_n_frames, pbar, step_ref):
+        """Run all interpolation passes on a chunk of frames."""
+        for pass_idx in range(num_passes):
+            new_frames = []
+            num_pairs = frames.shape[0] - 1
+            pairs_since_clear = 0
+
+            for i in range(0, num_pairs, batch_size):
+                batch_end = min(i + batch_size, num_pairs)
+                actual_batch = batch_end - i
+
+                frames0 = frames[i:batch_end]
+                frames1 = frames[i + 1:batch_end + 1]
+
+                if not keep_device:
+                    model.to(device)
+
+                mids = model.interpolate_batch(frames0, frames1, time_step=0.5)
+                mids = mids.to(storage_device)
+
+                if not keep_device:
+                    model.to("cpu")
+
+                for j in range(actual_batch):
+                    new_frames.append(frames[i + j:i + j + 1])
+                    new_frames.append(mids[j:j+1])
+
+                step_ref[0] += actual_batch
+                pbar.update_absolute(step_ref[0])
+
+                pairs_since_clear += actual_batch
+                if not all_on_gpu and pairs_since_clear >= clear_cache_after_n_frames and torch.cuda.is_available():
+                    clear_ema_warp_cache()
+                    torch.cuda.empty_cache()
+                    pairs_since_clear = 0
+
+            new_frames.append(frames[-1:])
+            frames = torch.cat(new_frames, dim=0)
+
+            if not all_on_gpu and torch.cuda.is_available():
+                clear_ema_warp_cache()
+                torch.cuda.empty_cache()
+
+        return frames
+
+    @staticmethod
+    def _count_steps(num_frames, num_passes):
+        """Count total interpolation steps for a given input frame count."""
+        n = num_frames
+        total = 0
+        for _ in range(num_passes):
+            total += n - 1
+            n = 2 * n - 1
+        return total
+
+    def interpolate(self, images, model, multiplier, clear_cache_after_n_frames,
+                    keep_device, all_on_gpu, batch_size, chunk_size):
+        if images.shape[0] < 2:
+            return (images,)
+
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        num_passes = {2: 1, 4: 2, 8: 3}[multiplier]
+
+        if all_on_gpu:
+            keep_device = True
+
+        storage_device = device if all_on_gpu else torch.device("cpu")
+
+        # Convert from ComfyUI [B, H, W, C] to model [B, C, H, W]
+        all_frames = images.permute(0, 3, 1, 2).to(storage_device)
+        total_input = all_frames.shape[0]
+
+        # Build chunk boundaries (1-frame overlap between consecutive chunks)
+        if chunk_size < 2 or chunk_size >= total_input:
+            chunks = [(0, total_input)]
+        else:
+            chunks = []
+            start = 0
+            while start < total_input - 1:
+                end = min(start + chunk_size, total_input)
+                chunks.append((start, end))
+                start = end - 1  # overlap by 1 frame
+                if end == total_input:
+                    break
+
+        # Calculate total progress steps across all chunks
+        total_steps = sum(self._count_steps(ce - cs, num_passes) for cs, ce in chunks)
+        pbar = ProgressBar(total_steps)
+        step_ref = [0]
+
+        if keep_device:
+            model.to(device)
+
+        result_chunks = []
+        for chunk_idx, (chunk_start, chunk_end) in enumerate(chunks):
+            chunk_frames = all_frames[chunk_start:chunk_end].clone()
+
+            chunk_result = self._interpolate_frames(
+                chunk_frames, model, num_passes, batch_size,
+                device, storage_device, keep_device, all_on_gpu,
+                clear_cache_after_n_frames, pbar, step_ref,
+            )
+
+            # Skip first frame of subsequent chunks (duplicate of previous chunk's last frame)
+            if chunk_idx > 0:
+                chunk_result = chunk_result[1:]
+
+            # Move completed chunk to CPU to bound memory when chunking
+            if len(chunks) > 1:
+                chunk_result = chunk_result.cpu()
+
+            result_chunks.append(chunk_result)
+
+        result = torch.cat(result_chunks, dim=0)
+        # Convert back to ComfyUI [B, H, W, C], on CPU
+        result = result.cpu().permute(0, 2, 3, 1)
+        return (result,)
+
+
+class EMAVFISegmentInterpolate(EMAVFIInterpolate):
+    """Process a numbered segment of the input batch for EMA-VFI.
+
+    Chain multiple instances with Save nodes between them to bound peak RAM.
+    The model pass-through output forces sequential execution so each segment
+    saves and frees from RAM before the next starts.
+    """
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        base = EMAVFIInterpolate.INPUT_TYPES()
+        base["required"]["segment_index"] = ("INT", {
+            "default": 0, "min": 0, "max": 10000, "step": 1,
+            "tooltip": "Which segment to process (0-based). Bounds RAM by only producing this segment's output frames, "
+                       "unlike chunk_size which bounds VRAM but still assembles the full output in RAM. "
+                       "Chain the model output to the next Segment Interpolate to force sequential execution.",
+        })
+        base["required"]["segment_size"] = ("INT", {
+            "default": 500, "min": 2, "max": 10000, "step": 1,
+            "tooltip": "Number of input frames per segment. Adjacent segments overlap by 1 frame for seamless stitching. "
+                       "Smaller = less peak RAM per segment. Save each segment's output to disk before the next runs.",
+        })
+        return base
+
+    RETURN_TYPES = ("IMAGE", "EMA_VFI_MODEL")
+    RETURN_NAMES = ("images", "model")
+    FUNCTION = "interpolate"
+    CATEGORY = "video/EMA-VFI"
+
+    def interpolate(self, images, model, multiplier, clear_cache_after_n_frames,
+                    keep_device, all_on_gpu, batch_size, chunk_size,
+                    segment_index, segment_size):
+        total_input = images.shape[0]
+
+        # Compute segment boundaries (1-frame overlap)
+        start = segment_index * (segment_size - 1)
+        end = min(start + segment_size, total_input)
+
+        if start >= total_input - 1:
+            # Past the end — return empty single frame + model
+            return (images[:1], model)
+
+        segment_images = images[start:end]
+        is_continuation = segment_index > 0
+
+        # Delegate to the parent interpolation logic
+        (result,) = super().interpolate(
+            segment_images, model, multiplier, clear_cache_after_n_frames,
+            keep_device, all_on_gpu, batch_size, chunk_size,
+        )
+
+        if is_continuation:
+            result = result[1:]  # skip duplicate boundary frame
+
+        return (result, model)

utils/padder.py

@@ -4,17 +4,22 @@ import torch.nn.functional as F
 class InputPadder:
     """ Pads images such that dimensions are divisible by divisor """
 
-    def __init__(self, dims, divisor=16):
+    def __init__(self, dims, divisor=16, mode='constant', center=False):
         self.ht, self.wd = dims[-2:]
+        self.mode = mode
         pad_ht = (((self.ht // divisor) + 1) * divisor - self.ht) % divisor
         pad_wd = (((self.wd // divisor) + 1) * divisor - self.wd) % divisor
-        self._pad = [0, pad_wd, 0, pad_ht]
+        if center:
+            self._pad = [pad_wd // 2, pad_wd - pad_wd // 2,
+                         pad_ht // 2, pad_ht - pad_ht // 2]
+        else:
+            self._pad = [0, pad_wd, 0, pad_ht]
 
     def pad(self, *inputs):
         if len(inputs) == 1:
-            return F.pad(inputs[0], self._pad, mode='constant')
+            return F.pad(inputs[0], self._pad, mode=self.mode)
         else:
-            return [F.pad(x, self._pad, mode='constant') for x in inputs]
+            return [F.pad(x, self._pad, mode=self.mode) for x in inputs]
 
     def unpad(self, *inputs):
         if len(inputs) == 1: