159 lines
6.6 KiB
Python
159 lines
6.6 KiB
Python
import torch
|
|
import math
|
|
from .helper_functions import (
|
|
generate_dim_variables,
|
|
getIndexTensorAlongDim,
|
|
parse_expr,
|
|
commonLazy,
|
|
normalize_to_common_shape,
|
|
as_tensor,
|
|
prepare_inputs
|
|
)
|
|
|
|
from .Parser.UnifiedMathVisitor import UnifiedMathVisitor
|
|
from comfy_api.latest import io
|
|
from comfy_api.latest import VideoComponents, VideoFromComponents
|
|
|
|
|
|
class VideoMathNode(io.ComfyNode):
|
|
"""
|
|
Enables math expressions on Videos (Images + Audio).
|
|
Video object is expected to have 'images' and 'audio' components.
|
|
Expressions for 'Images' and 'Audio' are handled separately.
|
|
"""
|
|
|
|
@classmethod
|
|
def define_schema(cls) -> io.Schema:
|
|
return io.Schema(
|
|
node_id="mrmth_VideoMathNode",
|
|
category="More math",
|
|
display_name="Video math",
|
|
inputs=[
|
|
io.Video.Input(id="a"),
|
|
io.Video.Input(id="b", optional=True, lazy=True),
|
|
io.Video.Input(id="c", optional=True, lazy=True),
|
|
io.Video.Input(id="d", optional=True, lazy=True),
|
|
io.Float.Input(id="w", default=0.0, optional=True, lazy=True, force_input=True),
|
|
io.Float.Input(id="x", default=0.0, optional=True, lazy=True, force_input=True),
|
|
io.Float.Input(id="y", default=0.0, optional=True, lazy=True, force_input=True),
|
|
io.Float.Input(id="z", default=0.0, optional=True, lazy=True, force_input=True),
|
|
io.String.Input(id="Images", default="a*(1-w)+b*w", tooltip="Expression for the image frames"),
|
|
io.String.Input(id="Audio", default="a*(1-w)+b*w", tooltip="Expression for the audio component"),
|
|
io.Combo.Input(
|
|
id="length_mismatch",
|
|
options=["tile", "error", "pad"],
|
|
default="error",
|
|
tooltip="How to handle mismatched frame/sample counts. tile: repeat shorter inputs; error: raise error on mismatch; pad: treat missing as zero."
|
|
)
|
|
],
|
|
outputs=[
|
|
io.Video.Output(),
|
|
],
|
|
)
|
|
|
|
@classmethod
|
|
def check_lazy_status(cls, Images, Audio, a, b=[], c=[], d=[], w=0, x=0, y=0, z=0, length_mismatch="tile"):
|
|
return commonLazy(Images, a, b, c, d, w, x, y, z) + commonLazy(Audio, a, b, c, d, w, x, y, z)
|
|
|
|
@classmethod
|
|
def execute(cls, Images, Audio, a, b=None, c=None, d=None, w=0.0, x=0.0, y=0.0, z=0.0, length_mismatch="tile") -> io.NodeOutput:
|
|
|
|
ac, bc, cc, dc = prepare_inputs(a,b,c,d)
|
|
imgs_ae = ac.images
|
|
imgs_be = bc.images
|
|
imgs_ce = cc.images
|
|
imgs_de = dc.images
|
|
imgs_ae, imgs_be, imgs_ce, imgs_de = normalize_to_common_shape(imgs_ae, imgs_be, imgs_ce, imgs_de, mode=length_mismatch)
|
|
|
|
img_vars = {
|
|
"a": imgs_ae, "b": imgs_be, "c": imgs_ce, "d": imgs_de,
|
|
"w": w, "x": x, "y": y, "z": z,
|
|
"B": getIndexTensorAlongDim(imgs_ae, 0),
|
|
"frame": getIndexTensorAlongDim(imgs_ae, 0),
|
|
"C": getIndexTensorAlongDim(imgs_ae, 1),
|
|
"channel": getIndexTensorAlongDim(imgs_ae, 1),
|
|
"H": imgs_ae.shape[2],
|
|
"height": imgs_ae.shape[2],
|
|
"W": imgs_ae.shape[3],
|
|
"width": imgs_ae.shape[3],
|
|
"T": imgs_ae.shape[0],
|
|
"frame_count": imgs_ae.shape[0],
|
|
"R": float(ac.frame_rate),
|
|
"frame_rate": float(ac.frame_rate),
|
|
"N": imgs_ae.shape[1],
|
|
"channel_count": imgs_ae.shape[1],
|
|
} | generate_dim_variables(imgs_ae)
|
|
|
|
tree = parse_expr(Images);
|
|
visitor = UnifiedMathVisitor(img_vars, imgs_ae.shape)
|
|
result_tensor = visitor.visit(tree)
|
|
result_imgs = as_tensor(result_tensor, imgs_ae.shape)
|
|
|
|
# --- Process Audio ---
|
|
wav_a = ac.audio["waveform"]
|
|
sample_rate = ac.audio["sample_rate"]
|
|
wav_b = bc.audio["waveform"] if bc is not None else None
|
|
wav_c = cc.audio["waveform"] if cc is not None else None
|
|
wav_d = dc.audio["waveform"] if dc is not None else None
|
|
|
|
# Determine target_wav_len as max of all provided audio
|
|
wav_lengths = [wav_a.shape[2]]
|
|
if bc is not None: wav_lengths.append(wav_b.shape[2])
|
|
if cc is not None: wav_lengths.append(wav_c.shape[2])
|
|
if dc is not None: wav_lengths.append(wav_d.shape[2])
|
|
target_wav_len = max(wav_lengths)
|
|
|
|
def resolve_wav(tensor, comp, name):
|
|
if comp is None:
|
|
return 0.0
|
|
|
|
curr_len = tensor.shape[2]
|
|
if curr_len == target_wav_len:
|
|
return tensor
|
|
|
|
if length_mismatch == "tile":
|
|
return tensor.repeat(1, 1, math.ceil(target_wav_len / curr_len))[:, :, :target_wav_len]
|
|
elif length_mismatch == "pad":
|
|
out = torch.zeros((tensor.shape[0], tensor.shape[1], target_wav_len), device=tensor.device, dtype=tensor.dtype)
|
|
out[:, :, :curr_len] = tensor
|
|
return out
|
|
else: # error
|
|
raise ValueError(f"Audio samples mismatch in {name}: expected {target_wav_len} to match longest input, got {curr_len}. Set length_mismatch to 'broadcast' or 'pad' to handle this.")
|
|
|
|
wav_ae = resolve_wav(wav_a, True, "input a")
|
|
wav_be = resolve_wav(wav_b, bc, "input b")
|
|
wav_ce = resolve_wav(wav_c, cc, "input c")
|
|
wav_de = resolve_wav(wav_d, dc, "input d")
|
|
|
|
wav_ae, wav_be, wav_ce, wav_de = normalize_to_common_shape(wav_ae, wav_be, wav_ce, wav_de, mode=length_mismatch)
|
|
|
|
wav_vars = {
|
|
"a": wav_ae, "b": wav_be, "c": wav_ce, "d": wav_de,
|
|
"w": w, "x": x, "y": y, "z": z,
|
|
"B": getIndexTensorAlongDim(wav_ae, 0),
|
|
"batch": getIndexTensorAlongDim(wav_ae, 0),
|
|
"C": getIndexTensorAlongDim(wav_ae, 1),
|
|
"channel": getIndexTensorAlongDim(wav_ae, 1),
|
|
"S": getIndexTensorAlongDim(wav_ae, 2),
|
|
"sample": getIndexTensorAlongDim(wav_ae, 2),
|
|
"R": sample_rate,
|
|
"sample_rate": sample_rate,
|
|
"T": wav_ae.shape[2],
|
|
"sample_count": wav_ae.shape[2],
|
|
"N": wav_ae.shape[1],
|
|
"channel_count": wav_ae.shape[1],
|
|
} | generate_dim_variables(wav_ae)
|
|
|
|
tree = parse_expr(Audio);
|
|
visitor = UnifiedMathVisitor(wav_vars, wav_ae.shape)
|
|
result_tensor = visitor.visit(tree)
|
|
result_audio_wav = as_tensor(result_tensor, wav_ae.shape)
|
|
|
|
output = VideoFromComponents(
|
|
VideoComponents(
|
|
images=result_imgs, audio={"waveform": result_audio_wav, "sample_rate": sample_rate}, frame_rate=ac.frame_rate
|
|
)
|
|
)
|
|
|
|
return (output,)
|