Meisoftcoltd
29915698e5
This resolves aggressive caching bugs in newer ComfyUI versions.
Replaced float("NaN") with time.time() in image.py and loop.py
to ensure stateful memory/trigger nodes consistently fire.
101 lines
3.5 KiB
Python
101 lines
3.5 KiB
Python
import os
|
|
import time
|
|
import random
|
|
import numpy as np
|
|
import torch
|
|
from PIL import Image
|
|
import folder_paths
|
|
from . import register_node
|
|
|
|
# Variable global para almacenar el último fotograma en la memoria del sistema (RAM)
|
|
global_session_image = None
|
|
|
|
def tensor_to_temp_image(tensor_image, prefix="session_img"):
|
|
"""Convierte un tensor de ComfyUI [1, H, W, C] a PNG temporal para la UI."""
|
|
temp_dir = folder_paths.get_temp_directory()
|
|
filename = f"{prefix}_{random.randint(10000, 99999)}.png"
|
|
filepath = os.path.join(temp_dir, filename)
|
|
|
|
# Extraemos el primer (y único) frame del tensor y lo convertimos
|
|
img_array = 255. * tensor_image[0].cpu().numpy()
|
|
img = Image.fromarray(np.clip(img_array, 0, 255).astype(np.uint8))
|
|
img.save(filepath)
|
|
|
|
return {"filename": filename, "subfolder": "", "type": "temp"}
|
|
|
|
|
|
@register_node
|
|
class SessionImageReceiver:
|
|
"""Proporciona la imagen inicial o la última generada, leyendo de la memoria RAM."""
|
|
@classmethod
|
|
def INPUT_TYPES(s):
|
|
return {
|
|
"required": {
|
|
"initial_image": ("IMAGE",),
|
|
"current_loop_index": ("INT", {"default": 0, "min": 0, "max": 10000, "step": 1}),
|
|
},
|
|
}
|
|
|
|
RETURN_TYPES = ("IMAGE",)
|
|
RETURN_NAMES = ("current_image",)
|
|
OUTPUT_NODE = True
|
|
FUNCTION = "get_image"
|
|
CATEGORY = "🔁 Sequential Batcher/Image"
|
|
|
|
@classmethod
|
|
def IS_CHANGED(cls, **kwargs):
|
|
return time.time() # Garantiza 100% una ruptura de caché
|
|
|
|
def get_image(self, initial_image, current_loop_index):
|
|
global global_session_image
|
|
|
|
# Extracción segura obligatoria por si hay listas
|
|
loop_idx = current_loop_index[0] if isinstance(current_loop_index, list) else current_loop_index
|
|
is_first_cycle = (loop_idx == 0)
|
|
|
|
if is_first_cycle or global_session_image is None:
|
|
global_session_image = initial_image.clone().cpu() # Guardamos en CPU por seguridad
|
|
print(f"[Sequential Batcher] Receiver: Ciclo {loop_idx}. Iniciando sesión con la imagen original.")
|
|
selected = initial_image
|
|
else:
|
|
print(f"[Sequential Batcher] Receiver: Ciclo {loop_idx}. Usando el fotograma rescatado de la RAM.")
|
|
# Movemos de vuelta a la GPU (o dejamos que ComfyUI lo asigne)
|
|
selected = global_session_image
|
|
|
|
ui_image = tensor_to_temp_image(selected, "receiver")
|
|
return {"ui": {"images": [ui_image]}, "result": (selected, )}
|
|
|
|
|
|
@register_node
|
|
class SessionImageSender:
|
|
"""Extrae, guarda en memoria del sistema (CPU) y muestra el último fotograma."""
|
|
@classmethod
|
|
def INPUT_TYPES(s):
|
|
return {
|
|
"required": {
|
|
"generated_images": ("IMAGE",),
|
|
},
|
|
}
|
|
|
|
RETURN_TYPES = ()
|
|
RETURN_NAMES = ()
|
|
OUTPUT_NODE = True
|
|
FUNCTION = "set_image"
|
|
CATEGORY = "🔁 Sequential Batcher/Image"
|
|
|
|
@classmethod
|
|
def IS_CHANGED(cls, **kwargs):
|
|
return time.time() # Garantiza 100% una ruptura de caché
|
|
|
|
def set_image(self, generated_images):
|
|
global global_session_image
|
|
|
|
# CRÍTICO: .cpu() asegura que ComfyUI no destruya el tensor al vaciar la VRAM de la gráfica
|
|
last_frame = generated_images[-1:].clone().cpu()
|
|
global_session_image = last_frame
|
|
|
|
print(f"[Sequential Batcher] Sender: Último fotograma capturado y asegurado en la RAM del sistema.")
|
|
|
|
ui_image = tensor_to_temp_image(last_frame, "sender")
|
|
return {"ui": {"images": [ui_image]}}
|