Merge commit 'bf8f76505fdd25ea99770f262eb9371509bf226d' as 'vendor/dumas_h3_continuum'
This commit is contained in:
+381
@@ -0,0 +1,381 @@
|
||||
"""Persistent MiniMax H3 video-reference conditioning for the V3.4 sampler."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from dataclasses import dataclass
|
||||
import hashlib
|
||||
import json
|
||||
import math
|
||||
from typing import Any
|
||||
|
||||
import torch
|
||||
|
||||
from .temporal import align_frame_count_up
|
||||
|
||||
|
||||
REFERENCE_VIDEO_CONTRACT_VERSION = 1
|
||||
REFERENCE_VIDEO_PREPROCESS_VERSION = 3
|
||||
REFERENCE_VIDEO_SIZE_EFFICIENT = "Efficient - 0.4 MP"
|
||||
REFERENCE_VIDEO_SIZE_BALANCED = "Balanced - 0.6 MP"
|
||||
REFERENCE_VIDEO_SIZE_MATCH_OUTPUT = "Match Output"
|
||||
REFERENCE_VIDEO_SIZE_OPTIONS = (
|
||||
REFERENCE_VIDEO_SIZE_EFFICIENT,
|
||||
REFERENCE_VIDEO_SIZE_BALANCED,
|
||||
REFERENCE_VIDEO_SIZE_MATCH_OUTPUT,
|
||||
)
|
||||
_CANVAS_MULTIPLE = 32
|
||||
_EFFICIENT_AREA = 400_000
|
||||
_BALANCED_AREA = 600_000
|
||||
_SOURCE_FPS = 24
|
||||
_HASH_CHUNK_FRAMES = 8
|
||||
|
||||
|
||||
class ReferenceVideoError(ValueError):
|
||||
pass
|
||||
|
||||
|
||||
def _canonical_hash(value: Any) -> str:
|
||||
payload = json.dumps(
|
||||
value, sort_keys=True, separators=(",", ":"), ensure_ascii=True
|
||||
).encode("utf-8")
|
||||
return hashlib.sha256(payload).hexdigest()
|
||||
|
||||
|
||||
def _tensor_hash(value: torch.Tensor) -> str:
|
||||
tensor = value.detach()
|
||||
digest = hashlib.sha256()
|
||||
digest.update(str(tuple(int(v) for v in tensor.shape)).encode("ascii"))
|
||||
digest.update(str(tensor.dtype).encode("ascii"))
|
||||
for start in range(0, int(tensor.shape[0]), _HASH_CHUNK_FRAMES):
|
||||
chunk = tensor[start : start + _HASH_CHUNK_FRAMES].to(
|
||||
device="cpu"
|
||||
).contiguous()
|
||||
digest.update(chunk.numpy().tobytes(order="C"))
|
||||
return digest.hexdigest()
|
||||
|
||||
|
||||
def _round_canvas(value: float) -> int:
|
||||
return max(
|
||||
_CANVAS_MULTIPLE,
|
||||
int(round(float(value) / _CANVAS_MULTIPLE)) * _CANVAS_MULTIPLE,
|
||||
)
|
||||
|
||||
|
||||
def resolve_reference_video_size(
|
||||
source_width: int,
|
||||
source_height: int,
|
||||
*,
|
||||
output_width: int,
|
||||
output_height: int,
|
||||
size_mode: str,
|
||||
) -> tuple[int, int]:
|
||||
if size_mode == REFERENCE_VIDEO_SIZE_MATCH_OUTPUT:
|
||||
target_area = int(output_width) * int(output_height)
|
||||
elif size_mode == REFERENCE_VIDEO_SIZE_BALANCED:
|
||||
target_area = _BALANCED_AREA
|
||||
else:
|
||||
target_area = _EFFICIENT_AREA
|
||||
source_area = int(source_width) * int(source_height)
|
||||
scale = min(1.0, math.sqrt(float(target_area) / float(source_area)))
|
||||
target_width = _round_canvas(int(source_width) * scale)
|
||||
target_height = _round_canvas(int(source_height) * scale)
|
||||
if source_width >= _CANVAS_MULTIPLE:
|
||||
target_width = min(
|
||||
target_width,
|
||||
max(_CANVAS_MULTIPLE, (int(source_width) // _CANVAS_MULTIPLE) * _CANVAS_MULTIPLE),
|
||||
)
|
||||
if source_height >= _CANVAS_MULTIPLE:
|
||||
target_height = min(
|
||||
target_height,
|
||||
max(_CANVAS_MULTIPLE, (int(source_height) // _CANVAS_MULTIPLE) * _CANVAS_MULTIPLE),
|
||||
)
|
||||
return target_width, target_height
|
||||
|
||||
|
||||
# Preserve the established private helper for existing internal callers/tests.
|
||||
_resolved_size = resolve_reference_video_size
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class ReferenceVideoSource:
|
||||
frames: torch.Tensor
|
||||
source_shape: tuple[int, ...]
|
||||
source_dtype: str
|
||||
source_sha256: str
|
||||
size_mode: str
|
||||
target_width: int
|
||||
target_height: int
|
||||
frame_count: int
|
||||
combined_hash: str
|
||||
|
||||
@property
|
||||
def contract(self) -> dict[str, Any]:
|
||||
return {
|
||||
"reference_video_contract_version": REFERENCE_VIDEO_CONTRACT_VERSION,
|
||||
"source_shape": list(self.source_shape),
|
||||
"source_dtype": self.source_dtype,
|
||||
"source_sha256": self.source_sha256,
|
||||
"source_fps": _SOURCE_FPS,
|
||||
"size_mode": self.size_mode,
|
||||
"target_width": self.target_width,
|
||||
"target_height": self.target_height,
|
||||
"frame_count": self.frame_count,
|
||||
"preprocess_version": REFERENCE_VIDEO_PREPROCESS_VERSION,
|
||||
"combined_hash": self.combined_hash,
|
||||
}
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class ReferenceVideoAssets:
|
||||
source: ReferenceVideoSource
|
||||
item: dict[str, Any]
|
||||
block: dict[str, Any]
|
||||
|
||||
|
||||
def resolve_reference_video_frame_count(
|
||||
source_frames: int,
|
||||
target_frames: int,
|
||||
) -> int:
|
||||
"""Resolve the used Video Guide frame count without touching frame data."""
|
||||
|
||||
source_count = int(source_frames)
|
||||
target_count = int(target_frames)
|
||||
if source_count < 5 or target_count < 5:
|
||||
raise ReferenceVideoError("Video Reference requires at least 5 frames")
|
||||
target_cap = align_frame_count_up(target_count)
|
||||
available_count = min(source_count, target_cap)
|
||||
return min(align_frame_count_up(available_count), target_cap)
|
||||
|
||||
|
||||
# Preserve the established private name while exposing one shared pure resolver.
|
||||
_resolve_reference_frame_count = resolve_reference_video_frame_count
|
||||
|
||||
|
||||
def _fit_reference_frames(frames: torch.Tensor, frame_count: int) -> torch.Tensor:
|
||||
requested_count = int(frame_count)
|
||||
current_count = int(frames.shape[0])
|
||||
if current_count >= requested_count:
|
||||
return frames[:requested_count].contiguous()
|
||||
if current_count < 1:
|
||||
raise ReferenceVideoError("Video Reference contains no frames")
|
||||
padding = frames[-1:].expand(requested_count - current_count, -1, -1, -1)
|
||||
return torch.cat((frames, padding), dim=0).contiguous()
|
||||
|
||||
|
||||
def _prepare_reference_frames(
|
||||
value: torch.Tensor,
|
||||
*,
|
||||
frame_count: int,
|
||||
) -> tuple[torch.Tensor, tuple[int, ...], str, str]:
|
||||
"""Hash the full normalized source while retaining only its used prefix."""
|
||||
|
||||
source_shape = (
|
||||
int(value.shape[0]),
|
||||
int(value.shape[1]),
|
||||
int(value.shape[2]),
|
||||
3,
|
||||
)
|
||||
source_dtype = str(torch.float32)
|
||||
retained_count = min(int(source_shape[0]), int(frame_count))
|
||||
retained = torch.empty(
|
||||
(retained_count, *source_shape[1:]),
|
||||
dtype=torch.float32,
|
||||
device="cpu",
|
||||
)
|
||||
digest = hashlib.sha256()
|
||||
digest.update(str(source_shape).encode("ascii"))
|
||||
digest.update(source_dtype.encode("ascii"))
|
||||
|
||||
for start in range(0, source_shape[0], _HASH_CHUNK_FRAMES):
|
||||
stop = min(start + _HASH_CHUNK_FRAMES, source_shape[0])
|
||||
chunk = value[start:stop, :, :, :3].detach().to(
|
||||
device="cpu", dtype=torch.float32
|
||||
).contiguous()
|
||||
if not bool(torch.isfinite(chunk).all().item()):
|
||||
raise ReferenceVideoError("Video Reference contains NaN or infinity")
|
||||
digest.update(chunk.numpy().tobytes(order="C"))
|
||||
if start < retained_count:
|
||||
retained_stop = min(stop, retained_count)
|
||||
retained[start:retained_stop].copy_(chunk[: retained_stop - start])
|
||||
|
||||
return retained, source_shape, source_dtype, digest.hexdigest()
|
||||
|
||||
|
||||
def prepare_reference_video_source(
|
||||
reference_video_1: Any,
|
||||
*,
|
||||
target_frames: int,
|
||||
output_width: int,
|
||||
output_height: int,
|
||||
size_mode: str = REFERENCE_VIDEO_SIZE_EFFICIENT,
|
||||
) -> ReferenceVideoSource | None:
|
||||
if reference_video_1 is None:
|
||||
return None
|
||||
if not torch.is_tensor(reference_video_1) or reference_video_1.ndim != 4:
|
||||
raise ReferenceVideoError(
|
||||
"Video Reference must be a ComfyUI IMAGE batch with shape [T, H, W, C]"
|
||||
)
|
||||
if int(reference_video_1.shape[-1]) < 3:
|
||||
raise ReferenceVideoError("Video Reference must contain RGB channels")
|
||||
frame_count = resolve_reference_video_frame_count(
|
||||
int(reference_video_1.shape[0]),
|
||||
int(target_frames),
|
||||
)
|
||||
frames, source_shape, source_dtype, source_sha256 = _prepare_reference_frames(
|
||||
reference_video_1,
|
||||
frame_count=frame_count,
|
||||
)
|
||||
normalized_size_mode = (
|
||||
str(size_mode)
|
||||
if str(size_mode) in REFERENCE_VIDEO_SIZE_OPTIONS
|
||||
else REFERENCE_VIDEO_SIZE_EFFICIENT
|
||||
)
|
||||
target_width, target_height = resolve_reference_video_size(
|
||||
int(source_shape[2]),
|
||||
int(source_shape[1]),
|
||||
output_width=int(output_width),
|
||||
output_height=int(output_height),
|
||||
size_mode=normalized_size_mode,
|
||||
)
|
||||
contract_base = {
|
||||
"reference_video_contract_version": REFERENCE_VIDEO_CONTRACT_VERSION,
|
||||
"source_shape": list(source_shape),
|
||||
"source_dtype": source_dtype,
|
||||
"source_sha256": source_sha256,
|
||||
"source_fps": _SOURCE_FPS,
|
||||
"size_mode": normalized_size_mode,
|
||||
"target_width": target_width,
|
||||
"target_height": target_height,
|
||||
"frame_count": frame_count,
|
||||
"preprocess_version": REFERENCE_VIDEO_PREPROCESS_VERSION,
|
||||
}
|
||||
return ReferenceVideoSource(
|
||||
frames=frames,
|
||||
source_shape=source_shape,
|
||||
source_dtype=source_dtype,
|
||||
source_sha256=source_sha256,
|
||||
size_mode=normalized_size_mode,
|
||||
target_width=target_width,
|
||||
target_height=target_height,
|
||||
frame_count=frame_count,
|
||||
combined_hash=_canonical_hash(contract_base),
|
||||
)
|
||||
|
||||
|
||||
def _prepare_reference_video_encode_input(
|
||||
source: ReferenceVideoSource,
|
||||
) -> tuple[torch.Tensor, dict[str, Any]]:
|
||||
available_count = min(int(source.frames.shape[0]), int(source.frame_count))
|
||||
frames = source.frames[:available_count]
|
||||
if (
|
||||
int(frames.shape[2]) != source.target_width
|
||||
or int(frames.shape[1]) != source.target_height
|
||||
):
|
||||
try:
|
||||
import comfy.utils
|
||||
except Exception as exc:
|
||||
raise ReferenceVideoError(
|
||||
"ComfyUI Core image resize support is unavailable"
|
||||
) from exc
|
||||
frames = comfy.utils.common_upscale(
|
||||
frames.movedim(-1, 1),
|
||||
source.target_width,
|
||||
source.target_height,
|
||||
"lanczos",
|
||||
"disabled",
|
||||
).movedim(1, -1).contiguous()
|
||||
frames = _fit_reference_frames(frames, source.frame_count)
|
||||
qwen_frames = frames[::_SOURCE_FPS // 2].contiguous()
|
||||
item = {
|
||||
"type": "video",
|
||||
"data": qwen_frames,
|
||||
"timestamps": [index / 2.0 for index in range(int(qwen_frames.shape[0]))],
|
||||
}
|
||||
return frames, item
|
||||
|
||||
|
||||
def _build_reference_video_assets(
|
||||
source: ReferenceVideoSource,
|
||||
item: dict[str, Any],
|
||||
latent: torch.Tensor,
|
||||
) -> ReferenceVideoAssets:
|
||||
block = {
|
||||
"kind": "video",
|
||||
"latent_t": int(latent.shape[2]),
|
||||
"latent_h": source.target_height // 16,
|
||||
"latent_w": source.target_width // 16,
|
||||
"ref_audio_t": 0,
|
||||
"latent": latent,
|
||||
"audio_latent": None,
|
||||
}
|
||||
return ReferenceVideoAssets(source=source, item=item, block=block)
|
||||
|
||||
|
||||
def encode_reference_video(
|
||||
video_vae: Any,
|
||||
source: ReferenceVideoSource,
|
||||
) -> ReferenceVideoAssets:
|
||||
frames, item = _prepare_reference_video_encode_input(source)
|
||||
try:
|
||||
latent = video_vae.encode(frames).detach().to("cpu").contiguous()
|
||||
except Exception as exc:
|
||||
raise ReferenceVideoError("Video Reference VAE Encode failed") from exc
|
||||
return _build_reference_video_assets(source, item, latent)
|
||||
|
||||
|
||||
def encode_reference_video_cached(
|
||||
video_vae: Any,
|
||||
source: ReferenceVideoSource,
|
||||
*,
|
||||
cache_event=None,
|
||||
) -> ReferenceVideoAssets:
|
||||
"""Encode only the V3.8 Video Guide latent through the shared cache."""
|
||||
|
||||
from .v3.ref_encode_cache import (
|
||||
get_ref_encode_cache,
|
||||
make_ref_encode_cache_key,
|
||||
)
|
||||
|
||||
frames, item = _prepare_reference_video_encode_input(source)
|
||||
cache = get_ref_encode_cache()
|
||||
key = make_ref_encode_cache_key(
|
||||
"video_guide",
|
||||
REFERENCE_VIDEO_PREPROCESS_VERSION,
|
||||
source.combined_hash,
|
||||
)
|
||||
latent = cache.lookup(video_vae, key, event_sink=cache_event)
|
||||
if latent is None:
|
||||
try:
|
||||
latent = video_vae.encode(frames).detach().to("cpu").contiguous()
|
||||
except Exception as exc:
|
||||
raise ReferenceVideoError("Video Reference VAE Encode failed") from exc
|
||||
if cache.supports_vae(video_vae):
|
||||
cache.store(video_vae, key, latent, event_sink=cache_event)
|
||||
return _build_reference_video_assets(source, item, latent)
|
||||
|
||||
|
||||
def combine_reference_video_identity(
|
||||
visual_identity_hash: str,
|
||||
source: ReferenceVideoSource | None,
|
||||
) -> str:
|
||||
if source is None:
|
||||
return str(visual_identity_hash)
|
||||
return _canonical_hash(
|
||||
{
|
||||
"reference_video_identity_version": 1,
|
||||
"visual_identity_hash": str(visual_identity_hash),
|
||||
"reference_video_hash": source.combined_hash,
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
def validate_reference_video_prompts(
|
||||
prompts: list[str], source: ReferenceVideoSource | None
|
||||
) -> str:
|
||||
if source is None or any("<Video 1>" in str(prompt) for prompt in prompts):
|
||||
return ""
|
||||
return (
|
||||
"H3C-P103 Warning: Video Reference is connected but the prompt contains no "
|
||||
"<Video 1> tag; video still conditions generation, but an explicit tag is "
|
||||
"recommended."
|
||||
)
|
||||
Reference in New Issue
Block a user