Files
aigc-home/docs/hel-892/director-replace-reference/t8__preflight.py
T
2026-11-01 10:50:03 +08:00

118 lines
4.5 KiB
Python

from __future__ import annotations
import json
import torch
from comfy.ldm.minimax.model import MiniMaxH3Model
from .core import (
MAX_PIXELS,
MAX_TRAINED_FRAMES,
MIN_TRAINED_FRAMES,
VRAM_CAUTION_PIXELS,
align_frame_count,
classify_h3_vae,
sorted_autogrow_values,
validate_audio,
)
def run_preflight(
width: int,
height: int,
length: int,
audio_mode: str,
model=None,
video_vae=None,
audio_vae=None,
drive_audio=None,
ref_images=None,
ref_videos=None,
ref_audios=None,
):
errors: list[str] = []
warnings: list[str] = []
facts: dict[str, object] = {}
if width <= 0 or height <= 0 or width % 32 or height % 32:
errors.append("width and height must be positive multiples of 32")
pixels = width * height
facts["pixels"] = pixels
facts["max_pixel_area"] = MAX_PIXELS
if pixels > MAX_PIXELS:
errors.append(
f"canvas has {pixels:,} pixels; configured H3 2.0MP cap is "
f"{MAX_PIXELS:,} pixels (1920x1088)"
)
elif pixels > VRAM_CAUTION_PIXELS:
warnings.append(
f"high-resolution canvas has {pixels:,} pixels; this is supported up to "
"1920x1088 but may require substantially more VRAM"
)
aligned = align_frame_count(length)
facts["aligned_frames"] = aligned
if aligned != length:
warnings.append(f"length {length} will snap up to {aligned} on the 17n+5 grid")
if not MIN_TRAINED_FRAMES <= aligned <= MAX_TRAINED_FRAMES:
warnings.append(
f"{aligned} frames is outside the approximately trained {MIN_TRAINED_FRAMES}-{MAX_TRAINED_FRAMES} range"
)
if model is not None:
diffusion_model = getattr(getattr(model, "model", None), "diffusion_model", None)
if not isinstance(diffusion_model, MiniMaxH3Model):
errors.append("model is not a native ComfyUI MiniMax H3 diffusion model")
if video_vae is not None:
video_vae_kind = classify_h3_vae(video_vae)
facts["video_vae_kind"] = video_vae_kind
if video_vae_kind == "audio":
errors.append("video_vae is an H3 audio VAE; connect the H3 video VAE")
if audio_vae is not None:
audio_vae_kind = classify_h3_vae(audio_vae)
facts["audio_vae_kind"] = audio_vae_kind
if audio_vae_kind == "video":
errors.append("audio_vae is an H3 video VAE; connect the H3 audio VAE")
facts["audio_vae_sample_rate"] = int(getattr(audio_vae, "audio_sample_rate", 32000))
if audio_mode != "native" and drive_audio is None:
errors.append(f"audio mode {audio_mode} requires drive_audio")
if drive_audio is not None:
try:
waveform, sample_rate = validate_audio(drive_audio, "drive_audio")
facts["drive_audio_sample_rate"] = sample_rate
facts["drive_audio_duration_seconds"] = waveform.shape[-1] / sample_rate
if not torch.isfinite(waveform).all():
errors.append("drive_audio contains NaN or infinite samples")
else:
peak = float(waveform.abs().max())
rms = float(torch.sqrt(torch.mean(waveform.float().square())))
facts["drive_audio_peak"] = peak
facts["drive_audio_rms"] = rms
if peak > 1.0:
warnings.append(f"drive_audio peak {peak:.3f} exceeds 0 dBFS and may clip")
if rms < 1e-4:
warnings.append("drive_audio is effectively silent")
except ValueError as error:
errors.append(str(error))
pictures = sorted_autogrow_values(ref_images)
videos = sorted_autogrow_values(ref_videos)
audios = sorted_autogrow_values(ref_audios)
facts.update({"reference_images": len(pictures), "reference_videos": len(videos), "reference_audios": len(audios)})
if len(pictures) > 9 or len(videos) > 3 or len(audios) > 3:
errors.append("reference counts exceed H3 limits (9 pictures / 3 videos / 3 audios)")
for index, frames in enumerate(videos, 1):
if not isinstance(frames, torch.Tensor) or frames.ndim != 4:
errors.append(f"ref_video_{index} is not an IMAGE frame batch")
elif not 48 <= frames.shape[0] <= 360:
warnings.append(f"ref_video_{index} has {frames.shape[0]} frames; official guidance is 48-360 at 24fps")
ready = not errors
report = json.dumps(
{"ready": ready, "errors": errors, "warnings": warnings, "facts": facts},
ensure_ascii=False,
indent=2,
)
return ready, len(warnings), report