digital-garbageandClaude Sonnet 5.5
0a8e16ab2b
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
93 lines
4.5 KiB
Python
93 lines
4.5 KiB
Python
"""Latent to pictures, and to sound where there is any.
|
|
|
|
Writing the file is FunUI Save Video's job (core's SaveVideo encoded on one
|
|
CPU thread); VHS is unnecessary either way.
|
|
|
|
What core cannot generalise is the decode itself for a model whose latent is not
|
|
one tensor. An AV latent carries a video branch and an audio branch that need
|
|
different VAEs, and how they are arranged is the model's business. So the same
|
|
rule as the empty latent applies: a model's own module may claim the decode, and
|
|
otherwise it is the ordinary single-tensor path.
|
|
|
|
The node takes the MODEL for one reason: so a provider can answer "is this mine"
|
|
by IDENTITY rather than by shape. Without it, the H3 module claimed any latent
|
|
that happened to have two parts -- correct today, because H3 is the only thing
|
|
producing one, and silently wrong the moment a second AV model exists. Deciding
|
|
from shape is how a model mismatch turns into an unrelated-looking fault instead
|
|
of an error.
|
|
"""
|
|
|
|
from comfy_api.latest import io
|
|
|
|
from ..._core import log, registry as registry_mod, vae_tiles
|
|
|
|
CAPABILITY = "decode"
|
|
|
|
|
|
def _plain_decode(vae, latent, tile_size):
|
|
return vae_tiles.tiled_decode(vae, latent, tile_size)
|
|
|
|
|
|
class FunUIDecode(io.ComfyNode):
|
|
@classmethod
|
|
def define_schema(cls) -> io.Schema:
|
|
return io.Schema(
|
|
node_id="FunUIDecode",
|
|
display_name="FunUI Decode",
|
|
category="FunUI/Output",
|
|
description="Decode a latent to images, and to audio when the model has any.",
|
|
inputs=[
|
|
io.Latent.Input("samples"),
|
|
io.Vae.Input("vae", tooltip="The picture VAE."),
|
|
io.Model.Input("model", optional=True,
|
|
tooltip="The model this latent came from. Needed only when a "
|
|
"model has its own way of being decoded."),
|
|
io.Vae.Input("audio_vae", optional=True,
|
|
tooltip="Only for models that generate sound alongside the video."),
|
|
io.Int.Input("tile_size", default=0, min=0, max=4096, step=64, optional=True,
|
|
tooltip="Decodes in tiles of this many pixels to fit less VRAM. 0 uses the model's own way. For MiniMax H3 keep 0, since bigger tiles show a grid."),
|
|
],
|
|
outputs=[
|
|
io.Image.Output(display_name="images"),
|
|
io.Audio.Output(display_name="audio"),
|
|
io.String.Output(display_name="status"),
|
|
],
|
|
)
|
|
|
|
@classmethod
|
|
def execute(cls, samples, vae, model=None, audio_vae=None, tile_size=0) -> io.NodeOutput:
|
|
latent = samples["samples"]
|
|
vae = getattr(vae, "funui_decode_vae", None) or vae # a decode-only VAE named in the VAE loader
|
|
|
|
for spec, decode in registry_mod.current().providers(CAPABILITY):
|
|
try:
|
|
claimed = decode(latent, model=model, vae=vae, audio_vae=audio_vae, tile_size=tile_size)
|
|
except Exception as exc: # noqa: BLE001
|
|
# Same protocol as the empty latent: returning None means "not my
|
|
# model", so getting this far means it WAS and it broke. Falling
|
|
# back would decode a nested latent as if it were one tensor.
|
|
log.broke(f"{spec.id}.{CAPABILITY}", exc, "decoding this model's latent")
|
|
raise RuntimeError(
|
|
f"{spec.id} handles this model's decode and failed: "
|
|
f"{type(exc).__name__}: {exc}. Refusing to decode it as a plain "
|
|
f"latent, which would not be the same picture."
|
|
) from exc
|
|
if claimed is not None:
|
|
images, audio = claimed
|
|
return io.NodeOutput(images, audio, f"{spec.id} decoded this")
|
|
|
|
if getattr(latent, "is_nested", False):
|
|
# Nothing claimed it and it is not one tensor: vae.decode would either
|
|
# raise somewhere confusing or quietly decode the wrong branch.
|
|
raise RuntimeError(
|
|
"This latent has more than one part and no installed module knows "
|
|
"how to decode it."
|
|
+ ("" if model is not None else
|
|
" The model input is not wired, so no module could recognise it.")
|
|
+ " The module for this model is missing.")
|
|
|
|
images = _plain_decode(vae, latent, tile_size)
|
|
if len(images.shape) == 5: # a batch of clips -> a strip of frames
|
|
images = images.reshape(-1, *images.shape[-3:])
|
|
return io.NodeOutput(images, None, "decoded as a single latent")
|