File size: 7,468 Bytes
51a7201
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c70af80
51a7201
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c70af80
 
 
51a7201
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
import os
import math
import glob
import urllib.request

import torch

try:
    import folder_paths
    CKPT_DIR = os.path.join(folder_paths.models_dir, "gapfiller")
except Exception:
    CKPT_DIR = os.path.join(os.path.dirname(__file__), "checkpoints")
os.makedirs(CKPT_DIR, exist_ok=True)

from .gapfiller_model import GapFillerNet

DEFAULT_CKPT = "gapfiller_v1.pt"
WEIGHT_URLS = [
    "https://github.com/KernelPanicKitten/ComfyUI-GapFiller/releases/download/v0.1.0/gapfiller_v1.pt",
]

CFG = {
    "current": ((128, 96, 72), (8, 6, 4)),
    "bigger": ((192, 144, 96), (10, 8, 6)),
    "large": ((256, 192, 128), (12, 10, 8)),
}
_CACHE = {}


def _load_safetensors(path, device):
    """Load safetensors format checkpoint."""
    try:
        from safetensors.torch import load_file
        return load_file(path)
    except ImportError:
        print("[GapFiller] safetensors not installed, falling back to torch.load")
        return torch.load(path, map_location=device)


def _load(name, device):
    key = (name, str(device))
    if key not in _CACHE:
        path = os.path.join(CKPT_DIR, name)
        if not os.path.exists(path):
            raise FileNotFoundError(f"Checkpoint not found: {path}")
        
        # Check file extension to determine loading method
        if path.endswith('.safetensors'):
            ck = _load_safetensors(path, device)
        else:
            ck = torch.load(path, map_location=device)
        
        chans, nres = CFG.get(ck.get("size") if isinstance(ck, dict) else None, CFG["large"])
        model = GapFillerNet(chans=chans, nres=nres).to(device).eval()
        
        # Handle both safetensors dict and torch dict formats
        if isinstance(ck, dict):
            if "model" in ck:
                model.load_state_dict(ck["model"])
            else:
                model.load_state_dict(ck)
        else:
            model.load_state_dict(ck)
            
        _CACHE[key] = model
    return _CACHE[key]


def _ckpts():
    found = sorted(os.path.basename(p) for p in glob.glob(os.path.join(CKPT_DIR, "*.pt")))
    found += sorted(os.path.basename(p) for p in glob.glob(os.path.join(CKPT_DIR, "*.safetensors")))
    return found or [DEFAULT_CKPT]


KNOBS = {
    "sharpness": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 2.5, "step": 0.05,
                            "tooltip": "Scales the learned detail. Above 1 is crisper, below 1 softer."}),
    "blend_bias": ("FLOAT", {"default": 0.0, "min": -3.0, "max": 3.0, "step": 0.1,
                             "tooltip": "Biases the blend toward the previous (+) or next (-) frame."}),
    "flow_scale": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.5, "step": 0.05,
                             "tooltip": "Damps estimated motion. Below 1 is safer on very large motion."}),
    "scene_thresh": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01,
                               "tooltip": "Above this frame difference, cut instead of morphing. 0 disables."}),
    "ensemble": ("BOOLEAN", {"default": True,
                             "tooltip": "Average both temporal directions. Roughly 2x compute, more accurate."}),
    "scale_factor": ([0.0, 0.25, 0.5, 1.0], {"default": 0.0,
                     "tooltip": "Flow resolution. 0 is automatic (coarse flow on large motion)."}),
}


def _run(model, a, b, t, device, knobs):
    with torch.no_grad(), torch.autocast("cuda", torch.bfloat16, enabled=(device.type == "cuda")):
        out = model(a[None].to(device), b[None].to(device), t=t, **knobs)
    return out[0].float().clamp(0, 1).cpu()


class GapFillerInterpolate:
    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "images": ("IMAGE",),
                "multiplier": ("INT", {"default": 2, "min": 2, "max": 16}),
                "ckpt_name": (_ckpts(),),
                **KNOBS,
            },
            "optional": {
                "loop": ("BOOLEAN", {"default": False,
                                     "tooltip": "Also interpolate last to first for a seamless loop."}),
            },
        }

    RETURN_TYPES = ("IMAGE",)
    FUNCTION = "interpolate"
    CATEGORY = "GapFiller"

    @torch.no_grad()
    def interpolate(self, images, multiplier, ckpt_name, sharpness, blend_bias,

                    flow_scale, scene_thresh, ensemble, scale_factor, loop=False):
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        model = _load(ckpt_name, device)
        knobs = dict(sharpness=sharpness, blend_bias=blend_bias, flow_scale=flow_scale,
                     scene_thresh=scene_thresh, ensemble=ensemble, scale_factor=scale_factor)
        seq = [images[i].permute(2, 0, 1).contiguous() for i in range(images.shape[0])]
        if loop and len(seq) > 1:
            seq = seq + [seq[0]]
        if len(seq) < 2:
            return (images,)
        out = []
        for i in range(len(seq) - 1):
            out.append(seq[i])
            for j in range(1, multiplier):
                out.append(_run(model, seq[i], seq[i + 1], j / multiplier, device, knobs))
        if not loop:
            out.append(seq[-1])
        return (torch.stack([f.permute(1, 2, 0) for f in out]).clamp(0, 1),)


class GapFillerRetime:
    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "images": ("IMAGE",),
                "src_fps": ("FLOAT", {"default": 24.0, "min": 1.0, "max": 240.0, "step": 0.01}),
                "target_fps": ("FLOAT", {"default": 60.0, "min": 1.0, "max": 240.0, "step": 0.01}),
                "ckpt_name": (_ckpts(),),
                **KNOBS,
            }
        }

    RETURN_TYPES = ("IMAGE",)
    FUNCTION = "retime"
    CATEGORY = "GapFiller"

    @torch.no_grad()
    def retime(self, images, src_fps, target_fps, ckpt_name, sharpness, blend_bias,

               flow_scale, scene_thresh, ensemble, scale_factor):
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        model = _load(ckpt_name, device)
        knobs = dict(sharpness=sharpness, blend_bias=blend_bias, flow_scale=flow_scale,
                     scene_thresh=scene_thresh, ensemble=ensemble, scale_factor=scale_factor)
        seq = [images[i].permute(2, 0, 1).contiguous() for i in range(images.shape[0])]
        n = len(seq)
        if n < 2:
            return (images,)
        nout = max(1, int(round((n - 1) / src_fps * target_fps)) + 1)
        out = []
        for i in range(nout):
            ts = (i / target_fps) * src_fps
            j = int(math.floor(ts))
            t = ts - j
            if j >= n - 1:
                out.append(seq[-1])
            elif t < 1e-4:
                out.append(seq[j])          # lands on a real frame, no synthesis needed
            else:
                out.append(_run(model, seq[j], seq[j + 1], t, device, knobs))
        return (torch.stack([f.permute(1, 2, 0) for f in out]).clamp(0, 1),)


NODE_CLASS_MAPPINGS = {
    "GapFillerInterpolate": GapFillerInterpolate,
    "GapFillerRetime": GapFillerRetime,
}
NODE_DISPLAY_NAME_MAPPINGS = {
    "GapFillerInterpolate": "GapFiller (frame interpolate)",
    "GapFillerRetime": "GapFiller (retime to fps)",
}
__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]