EasyCascadeLoader¶
Documentation¶
- Class name:
easy cascadeLoader - Category:
EasyUse/Loaders - Output node:
False
The easy cascadeLoader node is designed to facilitate the loading and management of cascading models, specifically focusing on integrating multiple model stages for enhanced processing capabilities. It abstracts the complexity of handling various model stages, making it easier to implement sophisticated model cascading strategies within the ComfyUI framework.
Input types¶
Required¶
stage_c- Specifies the combined list of filenames from the 'unet' and 'cascade' model directories. This input is crucial for determining the sequence and composition of models to be loaded for cascading operations, directly influencing the loader's behavior and the resulting model integration.
- Comfy dtype:
COMBO[STRING] - Python dtype:
List[str]
stage_b- Represents the filenames for the second stage of the cascading models, playing a critical role in the sequential processing and integration of model stages.
- Comfy dtype:
COMBO[STRING] - Python dtype:
List[str]
stage_a- Denotes the filenames for the initial stage of the cascading models, essential for setting up the foundational processing layer in the cascading strategy.
- Comfy dtype:
COMBO[STRING] - Python dtype:
List[str]
clip_name- Specifies the filenames for the CLIP models to be used in conjunction with the cascading models, enhancing the model's understanding and processing capabilities.
- Comfy dtype:
COMBO[STRING] - Python dtype:
List[str]
lora_name- Specifies the LoRA model names available for enhancing the model's capabilities, particularly in terms of adaptability and fine-tuning.
- Comfy dtype:
COMBO[STRING] - Python dtype:
List[str]
lora_model_strength- Determines the strength of the LoRA model adjustments, influencing the model's performance and adaptability.
- Comfy dtype:
FLOAT - Python dtype:
float
lora_clip_strength- Sets the strength of the CLIP adjustments when LoRA models are used, affecting the overall model performance.
- Comfy dtype:
FLOAT - Python dtype:
float
resolution- Defines the resolution for the model's output, impacting the quality and detail of the generated images.
- Comfy dtype:
COMBO[STRING] - Python dtype:
str
empty_latent_width- Specifies the width of the empty latent space, crucial for initializing the model's processing capabilities.
- Comfy dtype:
INT - Python dtype:
int
empty_latent_height- Determines the height of the empty latent space, essential for the model's initial setup and processing.
- Comfy dtype:
INT - Python dtype:
int
compression- Sets the level of compression for the model's output, balancing between quality and efficiency.
- Comfy dtype:
INT - Python dtype:
int
positive- Captures positive prompts or keywords to guide the model's generation process towards desired outcomes.
- Comfy dtype:
STRING - Python dtype:
str
negative- Includes negative prompts or keywords to steer the model away from undesired elements in the generation process.
- Comfy dtype:
STRING - Python dtype:
str
batch_size- Specifies the number of instances to be processed in a single batch, affecting the model's efficiency and throughput.
- Comfy dtype:
INT - Python dtype:
int
Optional¶
optional_lora_stack- Allows for the optional inclusion of a stack of LoRA models, offering enhanced customization and fine-tuning capabilities.
- Comfy dtype:
LORA_STACK - Python dtype:
List[str]
Output types¶
pipe- Comfy dtype:
PIPE_LINE - The pipeline object that orchestrates the flow and integration of various model stages.
- Python dtype:
Pipeline
- Comfy dtype:
model_c- Comfy dtype:
MODEL - The model object for the final stage in the cascading process, ready for further operations or analysis.
- Python dtype:
Model
- Comfy dtype:
latent_c- Comfy dtype:
LATENT - The latent representation generated by the final stage model, capturing the processed information.
- Python dtype:
LatentRepresentation
- Comfy dtype:
vae- Comfy dtype:
VAE - The VAE model used in the cascading process, contributing to the generation or processing of data.
- Python dtype:
VAEModel
- Comfy dtype:
Usage tips¶
- Infra type:
CPU - Common nodes: unknown
Source code¶
class cascadeLoader:
def __init__(self):
pass
@classmethod
def INPUT_TYPES(s):
return {"required": {
"stage_c": (folder_paths.get_filename_list("unet") + folder_paths.get_filename_list("checkpoints"),),
"stage_b": (folder_paths.get_filename_list("unet") + folder_paths.get_filename_list("checkpoints"),),
"stage_a": (["Baked VAE"]+folder_paths.get_filename_list("vae"),),
"clip_name": (["None"] + folder_paths.get_filename_list("clip"),),
"lora_name": (["None"] + folder_paths.get_filename_list("loras"),),
"lora_model_strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}),
"lora_clip_strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}),
"resolution": (resolution_strings, {"default": "1024 x 1024"}),
"empty_latent_width": ("INT", {"default": 1024, "min": 16, "max": MAX_RESOLUTION, "step": 8}),
"empty_latent_height": ("INT", {"default": 1024, "min": 16, "max": MAX_RESOLUTION, "step": 8}),
"compression": ("INT", {"default": 42, "min": 32, "max": 64, "step": 1}),
"positive": ("STRING", {"default":"", "placeholder": "Positive", "multiline": True}),
"negative": ("STRING", {"default":"", "placeholder": "Negative", "multiline": True}),
"batch_size": ("INT", {"default": 1, "min": 1, "max": 64}),
},
"optional": {"optional_lora_stack": ("LORA_STACK",), },
"hidden": {"prompt": "PROMPT", "my_unique_id": "UNIQUE_ID"}
}
RETURN_TYPES = ("PIPE_LINE", "MODEL", "LATENT", "VAE")
RETURN_NAMES = ("pipe", "model_c", "latent_c", "vae")
FUNCTION = "adv_pipeloader"
CATEGORY = "EasyUse/Loaders"
def is_ckpt(self, name):
is_ckpt = False
path = folder_paths.get_full_path("checkpoints", name)
if path is not None:
is_ckpt = True
return is_ckpt
def adv_pipeloader(self, stage_c, stage_b, stage_a, clip_name, lora_name, lora_model_strength, lora_clip_strength,
resolution, empty_latent_width, empty_latent_height, compression,
positive, negative, batch_size, optional_lora_stack=None,prompt=None,
my_unique_id=None):
vae: VAE | None = None
model_c: ModelPatcher | None = None
model_b: ModelPatcher | None = None
clip: CLIP | None = None
can_load_lora = True
pipe_lora_stack = []
# Clean models from loaded_objects
easyCache.update_loaded_objects(prompt)
# Create Empty Latent
samples = sampler.emptyLatent(resolution, empty_latent_width, empty_latent_height, batch_size, compression)
if self.is_ckpt(stage_c):
model_c, clip, vae_c, clip_vision = easyCache.load_checkpoint(stage_c)
else:
model_c = easyCache.load_unet(stage_c)
vae_c = None
if self.is_ckpt(stage_b):
model_b, clip, vae_b, clip_vision = easyCache.load_checkpoint(stage_b)
else:
model_b = easyCache.load_unet(stage_b)
vae_b = None
if optional_lora_stack is not None and can_load_lora:
for lora in optional_lora_stack:
lora = {"lora_name": lora[0], "model": model_c, "clip": clip, "model_strength": lora[1], "clip_strength": lora[2]}
model_c, clip = easyCache.load_lora(lora)
lora['model'] = model_c
lora['clip'] = clip
pipe_lora_stack.append(lora)
if lora_name != "None" and can_load_lora:
lora = {"lora_name": lora_name, "model": model_c, "clip": clip, "model_strength": lora_model_strength,
"clip_strength": lora_clip_strength}
model_c, clip = easyCache.load_lora(lora)
pipe_lora_stack.append(lora)
model = (model_c, model_b)
# Load clip
if clip_name != 'None':
clip = easyCache.load_clip(clip_name, "stable_cascade")
# Load vae
if stage_a not in ["Baked VAE", "Baked-VAE"]:
vae_b = easyCache.load_vae(stage_a)
vae = (vae_c, vae_b)
# 判断是否连接 styles selector
is_positive_linked_styles_selector = is_linked_styles_selector(prompt, my_unique_id, 'positive')
is_negative_linked_styles_selector = is_linked_styles_selector(prompt, my_unique_id, 'negative')
log_node_warn("正在处理提示词...")
positive_seed = find_wildcards_seed(my_unique_id, positive, prompt)
# Translate cn to en
if has_chinese(positive):
positive = zh_to_en([positive])[0]
model_c, clip, positive, positive_decode, show_positive_prompt, pipe_lora_stack = process_with_loras(positive,
model_c, clip,
"positive",
positive_seed,
can_load_lora,
pipe_lora_stack,
easyCache)
positive_wildcard_prompt = positive_decode if show_positive_prompt or is_positive_linked_styles_selector else ""
negative_seed = find_wildcards_seed(my_unique_id, negative, prompt)
# Translate cn to en
if has_chinese(negative):
negative = zh_to_en([negative])[0]
model_c, clip, negative, negative_decode, show_negative_prompt, pipe_lora_stack = process_with_loras(negative,
model_c, clip,
"negative",
negative_seed,
can_load_lora,
pipe_lora_stack,
easyCache)
negative_wildcard_prompt = negative_decode if show_negative_prompt or is_negative_linked_styles_selector else ""
tokens = clip.tokenize(positive)
cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True)
positive_embeddings_final = [[cond, {"pooled_output": pooled}]]
tokens = clip.tokenize(negative)
cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True)
negative_embeddings_final = [[cond, {"pooled_output": pooled}]]
image = easySampler.pil2tensor(Image.new('RGB', (1, 1), (0, 0, 0)))
log_node_warn("处理结束...")
pipe = {
"model": model,
"positive": positive_embeddings_final,
"negative": negative_embeddings_final,
"vae": vae,
"clip": clip,
"samples": samples,
"images": image,
"seed": 0,
"loader_settings": {
"vae_name": stage_a,
"lora_name": lora_name,
"lora_model_strength": lora_model_strength,
"lora_clip_strength": lora_clip_strength,
"lora_stack": pipe_lora_stack,
"positive": positive,
"positive_token_normalization": 'none',
"positive_weight_interpretation": 'comfy',
"negative": negative,
"negative_token_normalization": 'none',
"negative_weight_interpretation": 'comfy',
"resolution": resolution,
"empty_latent_width": empty_latent_width,
"empty_latent_height": empty_latent_height,
"batch_size": batch_size,
"compression": compression
}
}
return {"ui": {"positive": positive_wildcard_prompt, "negative": negative_wildcard_prompt},
"result": (pipe, model_c, model_b, vae)}