webp #4
@@ -1,12 +1,11 @@
|
||||
import os
|
||||
import torch
|
||||
import numpy as np
|
||||
from PIL import Image, ExifTags
|
||||
from PIL import Image
|
||||
from PIL.PngImagePlugin import PngInfo
|
||||
import concurrent.futures
|
||||
import re
|
||||
import time
|
||||
import json
|
||||
|
||||
class FastAbsoluteSaver:
|
||||
@classmethod
|
||||
@@ -20,11 +19,14 @@ class FastAbsoluteSaver:
|
||||
# --- FORMAT SWITCH ---
|
||||
"save_format": (["png", "webp"], ),
|
||||
|
||||
# --- PERFORMANCE ---
|
||||
"max_threads": ("INT", {"default": 0, "min": 0, "max": 128, "step": 1, "label": "Max Threads (0=Auto)"}),
|
||||
|
||||
# --- COMMON OPTIONS ---
|
||||
"filename_with_score": ("BOOLEAN", {"default": False, "label": "Append Score to Filename"}),
|
||||
"metadata_key": ("STRING", {"default": "sharpness_score"}),
|
||||
|
||||
# --- WEBP SPECIFIC (-z 6 -q 100 -lossless) ---
|
||||
# --- WEBP SPECIFIC ---
|
||||
"webp_lossless": ("BOOLEAN", {"default": True, "label": "WebP Lossless"}),
|
||||
"webp_quality": ("INT", {"default": 100, "min": 0, "max": 100, "step": 1, "label": "WebP Quality (-q)"}),
|
||||
"webp_method": ("INT", {"default": 4, "min": 0, "max": 6, "step": 1, "label": "WebP Compression (-z)"}),
|
||||
@@ -60,56 +62,18 @@ class FastAbsoluteSaver:
|
||||
|
||||
return frames[:batch_size], scores[:batch_size]
|
||||
|
||||
def get_webp_exif(self, key, value):
|
||||
"""
|
||||
Creates a basic Exif header to store the score in UserComment.
|
||||
ComfyUI standard metadata handling for WebP is complex,
|
||||
so we use a simple JSON dump inside the UserComment tag (ID 0x9286).
|
||||
"""
|
||||
# Create a basic exif dict
|
||||
exif_data = {
|
||||
0x9286: f"{key}: {value}".encode("utf-8") # UserComment
|
||||
}
|
||||
|
||||
# Convert to bytes manually to avoid requiring 'piexif' library
|
||||
# This is a minimal TIFF header structure for Exif.
|
||||
# If this is too hacky, we can just skip metadata for WebP,
|
||||
# but this usually works for basic viewers.
|
||||
|
||||
# ACTUALLY: Pillow's image.save(exif=...) expects raw bytes.
|
||||
# Generating raw Exif bytes from scratch is error-prone.
|
||||
# Simpler Strategy: We will create a fresh Image and modify its info.
|
||||
|
||||
# Since generating raw Exif without a library is risky,
|
||||
# we will skip internal metadata for WebP in this "No-Dependency" version
|
||||
# and rely on the filename.
|
||||
# *However*, if you strictly need it, we return None here and rely on filename
|
||||
# unless you have 'piexif' installed.
|
||||
return None
|
||||
|
||||
def save_single_image(self, tensor_img, full_path, score, key_name, fmt, lossless, quality, method):
|
||||
try:
|
||||
array = 255. * tensor_img.cpu().numpy()
|
||||
img = Image.fromarray(np.clip(array, 0, 255).astype(np.uint8))
|
||||
|
||||
if fmt == "png":
|
||||
# PNG METADATA (Robust)
|
||||
metadata = PngInfo()
|
||||
metadata.add_text(key_name, str(score))
|
||||
metadata.add_text("software", "ComfyUI_Parallel_Node")
|
||||
# PNG uses compress_level (0-9). Level 1 is fastest.
|
||||
img.save(full_path, format="PNG", pnginfo=metadata, compress_level=1)
|
||||
|
||||
elif fmt == "webp":
|
||||
# WEBP SAVING
|
||||
# Pillow options map directly to cwebp parameters:
|
||||
# method=6 -> -z 6 (Slowest, best compression)
|
||||
# quality=100 -> -q 100
|
||||
# lossless=True -> -lossless
|
||||
|
||||
# Note: WebP metadata in Pillow is finicky.
|
||||
# We save purely visual data here.
|
||||
# The score is in the filename (if option selected).
|
||||
img.save(full_path, format="WEBP",
|
||||
lossless=lossless,
|
||||
quality=quality,
|
||||
@@ -120,8 +84,8 @@ class FastAbsoluteSaver:
|
||||
print(f"xx- Error saving {full_path}: {e}")
|
||||
return False
|
||||
|
||||
def save_images_fast(self, images, output_path, filename_prefix, save_format, filename_with_score, metadata_key,
|
||||
webp_lossless, webp_quality, webp_method, scores_info=None):
|
||||
def save_images_fast(self, images, output_path, filename_prefix, save_format, max_threads,
|
||||
filename_with_score, metadata_key, webp_lossless, webp_quality, webp_method, scores_info=None):
|
||||
|
||||
output_path = output_path.strip('"')
|
||||
if not os.path.exists(output_path):
|
||||
@@ -130,12 +94,20 @@ class FastAbsoluteSaver:
|
||||
except OSError:
|
||||
raise ValueError(f"Could not create directory: {output_path}")
|
||||
|
||||
# --- AUTO-SCALING LOGIC ---
|
||||
if max_threads == 0:
|
||||
# os.cpu_count() returns None on some rare systems, so we default to 4 just in case
|
||||
cpu_cores = os.cpu_count() or 4
|
||||
# For WebP (CPU intensive), stick to core count.
|
||||
# For PNG (Disk intensive), we could technically go higher, but core count is safe.
|
||||
max_threads = cpu_cores
|
||||
|
||||
print(f"xx- FastSaver: Using {max_threads} Threads for saving.")
|
||||
|
||||
batch_size = len(images)
|
||||
frame_indices, scores_list = self.parse_info(scores_info, batch_size)
|
||||
|
||||
print(f"xx- FastSaver: Saving {batch_size} images ({save_format}) to {output_path}...")
|
||||
|
||||
with concurrent.futures.ThreadPoolExecutor(max_workers=16) as executor:
|
||||
with concurrent.futures.ThreadPoolExecutor(max_workers=max_threads) as executor:
|
||||
futures = []
|
||||
|
||||
for i, img_tensor in enumerate(images):
|
||||
@@ -151,22 +123,20 @@ class FastAbsoluteSaver:
|
||||
if real_frame_num == 0 and scores_info is None:
|
||||
base_name = f"{filename_prefix}_{int(time.time())}_{i:03d}"
|
||||
|
||||
# Append correct extension
|
||||
ext = ".webp" if save_format == "webp" else ".png"
|
||||
fname = f"{base_name}{ext}"
|
||||
full_path = os.path.join(output_path, fname)
|
||||
|
||||
# Submit
|
||||
futures.append(executor.submit(
|
||||
self.save_single_image,
|
||||
img_tensor,
|
||||
full_path,
|
||||
current_score,
|
||||
metadata_key,
|
||||
save_format, # fmt
|
||||
webp_lossless, # lossless
|
||||
webp_quality, # quality
|
||||
webp_method # method
|
||||
save_format,
|
||||
webp_lossless,
|
||||
webp_quality,
|
||||
webp_method
|
||||
))
|
||||
|
||||
concurrent.futures.wait(futures)
|
||||
|
||||
Reference in New Issue
Block a user