258 lines
13 KiB
Python
258 lines
13 KiB
Python
from comfy_execution.graph_utils import GraphBuilder
|
|
from .file_utils import get_path_without_base, get_file_extension
|
|
from .model_metadata import pull_and_update_model_timestamp
|
|
from .lora_stack import norm_lora_stack
|
|
from .model_info_utils import as_list, normalize_model_info_list, unwrap_single_item
|
|
|
|
from .logger import get_logger
|
|
|
|
logger = get_logger('helpers.graph')
|
|
|
|
# Utility function
|
|
def flatten_model_info(model_info):
|
|
"""Flatten model_info to a list of dictionaries."""
|
|
return normalize_model_info_list(model_info)
|
|
|
|
# Add individual model loader nodes
|
|
def add_ckpt_node(graph: GraphBuilder, ckpt_info):
|
|
ckpt_node = None
|
|
if ckpt_info is not None:
|
|
ckpt_info = unwrap_single_item(ckpt_info)
|
|
|
|
ckpt_name = ckpt_info["path"]
|
|
ckpt_fixed_name = get_path_without_base("checkpoints", ckpt_name)
|
|
# No GGUF support for checkpoints yet that I see.
|
|
logger.debug(f"Added node CheckpointLoaderSimple with ckpt_name: {ckpt_fixed_name}")
|
|
ckpt_node = graph.node("CheckpointLoaderSimple",ckpt_name=ckpt_fixed_name)
|
|
else:
|
|
logger.debug("No checkpoint found in model info, skipping checkpoint loading.")
|
|
return ckpt_node
|
|
|
|
def add_unet_node(graph: GraphBuilder, unet_info):
|
|
unet_node = None
|
|
if unet_info is not None:
|
|
unet_info = unwrap_single_item(unet_info)
|
|
|
|
unet_name = unet_info["path"]
|
|
unet_weight_dtype = unet_info["weight_dtype"]
|
|
unet_fixed_name = get_path_without_base("diffusion_models", unet_name)
|
|
if get_file_extension(unet_fixed_name) == "gguf":
|
|
# If the UNET is a GGUF, we need to use the GGUFLoader node. It doesn't ask for weight_dtype.
|
|
try:
|
|
logger.debug(f"Added node UnetLoaderGGUF with unet_name: {unet_fixed_name}")
|
|
unet_node = graph.node("UnetLoaderGGUF", unet_name=unet_fixed_name)
|
|
except Exception:
|
|
logger.debug("Unable to load UNET as GGUF. Do you have ComfyUI-GGUF installed?")
|
|
raise ValueError("Unable to load UNET as GGUF. Do you have ComfyUI-GGUF installed?")
|
|
else:
|
|
logger.debug(f"Added node UNETLoader with unet_name: {unet_fixed_name} and weight_dtype: {unet_weight_dtype}")
|
|
unet_node = graph.node("UNETLoader", unet_name=unet_fixed_name, weight_dtype=unet_weight_dtype)
|
|
return unet_node
|
|
|
|
def add_clip_node(graph: GraphBuilder, clip_info):
|
|
clip_node = None
|
|
if clip_info is not None:
|
|
clip_info = unwrap_single_item(clip_info)
|
|
logger.debug(f"Adding CLIP node with info: {clip_info}")
|
|
# 1+2 are in nodes.py, 3 is in node_sd3.py, and 4 is in nodes_hidream.py for legacy reasons.
|
|
clip_path = [get_path_without_base("clip", path) for path in as_list(clip_info["path"])]
|
|
clip_num = len(clip_path)
|
|
clip_type = clip_info["clip_type"]
|
|
logger.debug(f"Clip Path: {clip_path} Clip Type: {clip_type}")
|
|
|
|
# If any of the clip paths are GGUF, we need to use the GGUF loader. It can also handle non-GGUF paths.
|
|
if any(get_file_extension(path) == "gguf" for path in clip_path):
|
|
logger.debug("GGUF Clip!")
|
|
try:
|
|
if clip_num == 1:
|
|
logger.debug(f"Added node CLIPLoaderGGUF with clip_name: {clip_path[0]}, type: {clip_type}")
|
|
clip_node = graph.node("CLIPLoaderGGUF", clip_name=clip_path[0], type=clip_type)
|
|
elif clip_num == 2:
|
|
logger.debug(f"Added node DualCLIPLoaderGGUF with clip_name1: {clip_path[0]}, clip_name2: {clip_path[1]}, type: {clip_type}")
|
|
clip_node = graph.node("DualCLIPLoaderGGUF", clip_name1=clip_path[0], clip_name2=clip_path[1], type=clip_type)
|
|
elif clip_num == 3:
|
|
logger.debug(f"Added node TripleCLIPLoaderGGUF with clip_name1: {clip_path[0]}, clip_name2: {clip_path[1]}, clip_name3: {clip_path[2]}")
|
|
clip_node = graph.node("TripleCLIPLoaderGGUF", clip_name1=clip_path[0], clip_name2=clip_path[1], clip_name3=clip_path[2])
|
|
else:
|
|
logger.debug(f"Added node QuadrupleCLIPLoaderGGUF with clip_name1: {clip_path[0]}, clip_name2: {clip_path[1]}, clip_name3: {clip_path[2]}, clip_name4: {clip_path[3]}")
|
|
clip_node = graph.node("QuadrupleCLIPLoaderGGUF", clip_name1=clip_path[0], clip_name2=clip_path[1], clip_name3=clip_path[2], clip_name4=clip_path[3])
|
|
except Exception:
|
|
logger.debug("Unable to load CLIP as GGUF. Do you have ComfyUI-GGUF installed?")
|
|
raise ValueError("Unable to load CLIP as GGUF. Do you have ComfyUI-GGUF installed?")
|
|
else:
|
|
if clip_num == 1:
|
|
logger.debug(f"Added node CLIPLoader with clip_name: {clip_path[0]}, type: {clip_type}")
|
|
clip_node = graph.node("CLIPLoader", clip_name=clip_path[0], type=clip_type)
|
|
elif clip_num == 2:
|
|
logger.debug(f"Added node DualCLIPLoader with clip_name1: {clip_path[0]}, clip_name2: {clip_path[1]}, type: {clip_type}")
|
|
clip_node = graph.node("DualCLIPLoader", clip_name1=clip_path[0], clip_name2=clip_path[1], type=clip_type)
|
|
elif clip_num == 3:
|
|
logger.debug(f"Added node TripleCLIPLoader with clip_name1: {clip_path[0]}, clip_name2: {clip_path[1]}, clip_name3: {clip_path[2]}")
|
|
clip_node = graph.node("TripleCLIPLoader", clip_name1=clip_path[0], clip_name2=clip_path[1], clip_name3=clip_path[2])
|
|
else:
|
|
logger.debug(f"Added node QuadrupleCLIPLoader with clip_name1: {clip_path[0]}, clip_name2: {clip_path[1]}, clip_name3: {clip_path[2]}, clip_name4: {clip_path[3]}")
|
|
clip_node = graph.node("QuadrupleCLIPLoader", clip_name1=clip_path[0], clip_name2=clip_path[1], clip_name3=clip_path[2], clip_name4=clip_path[3])
|
|
return clip_node
|
|
|
|
def add_vae_node(graph: GraphBuilder, vae_info):
|
|
vae_node = None
|
|
if vae_info is not None:
|
|
vae_info = unwrap_single_item(vae_info)
|
|
|
|
vae_name = vae_info["path"]
|
|
vae_fixed_name = get_path_without_base("vae", vae_name)
|
|
# There is no GGUF for VAE currently, so we can just use VAELoader.
|
|
logger.debug(f"Added node VAELoader with vae_name: {vae_fixed_name}")
|
|
vae_node = graph.node("VAELoader", vae_name=vae_fixed_name)
|
|
return vae_node
|
|
|
|
# Add model loader nodes for checkpoints, unet, clip, or vae.
|
|
def create_model_loader_nodes(graph: GraphBuilder, model_info):
|
|
ckpt_node = unet_node = clip_node = vae_node = None
|
|
unet_out = clip_out = vae_out = None
|
|
model_present = {"CKPT": None, "UNET": None, "CLIP": None, "VAE": None}
|
|
|
|
# model_info is either a list of dicts, or a list of tuples with dicts inside. Either way, we need to go over each dict, and check the "type" key.
|
|
model_info = flatten_model_info(model_info)
|
|
logger.debug(f"Flattened model_info: {model_info}")
|
|
|
|
for idx, item in enumerate(model_info):
|
|
logger.debug(f"Processing model_info item {idx}: {item}")
|
|
model_present[item["type"]] = item
|
|
|
|
if model_present["CKPT"] is not None:
|
|
ckpt_node = add_ckpt_node(graph, model_present["CKPT"])
|
|
if ckpt_node is None:
|
|
raise ValueError("Checkpoint info is missing or invalid.")
|
|
else:
|
|
pull_and_update_model_timestamp(model_present["CKPT"]["path"], model_type="ckpt")
|
|
unet_out = ckpt_node.out(0)
|
|
clip_out = ckpt_node.out(1)
|
|
vae_out = ckpt_node.out(2)
|
|
|
|
# If we have a UNET, load it with the UNETLoader node.
|
|
if model_present["UNET"] is not None:
|
|
unet_node = add_unet_node(graph, model_present["UNET"])
|
|
|
|
if unet_node is None:
|
|
raise ValueError("UNET info is missing or invalid.")
|
|
else:
|
|
pull_and_update_model_timestamp(model_present["UNET"]["path"], model_type="unet")
|
|
unet_out = unet_node.out(0)
|
|
|
|
# If we have a CLIP, load it with the appropriate CLIPLoader node.
|
|
if model_present["CLIP"] is not None:
|
|
clip_node = add_clip_node(graph, model_present["CLIP"])
|
|
if clip_node is None:
|
|
raise ValueError("CLIP info is missing or invalid.")
|
|
else:
|
|
pull_and_update_model_timestamp(model_present["CLIP"]["path"], model_type="clip")
|
|
clip_out = clip_node.out(0)
|
|
|
|
# If we have a VAE, load it with the VAELoader node.
|
|
if model_present["VAE"] is not None:
|
|
vae_node = add_vae_node(graph, model_present["VAE"])
|
|
if vae_node is None:
|
|
raise ValueError("VAE info is missing or invalid.")
|
|
else:
|
|
pull_and_update_model_timestamp(model_present["VAE"]["path"], model_type="vae")
|
|
vae_out = vae_node.out(0)
|
|
return unet_out, clip_out, vae_out
|
|
|
|
# Add lora nodes for a lora stack.
|
|
def create_lora_nodes(graph: GraphBuilder, unet_in, clip_in=None, lora_stack=None):
|
|
exit_unet = unet_in
|
|
exit_clip = clip_in
|
|
exit_node = None
|
|
|
|
lora_stack = norm_lora_stack(lora_stack)
|
|
if lora_stack is None:
|
|
logger.info("No loras in stack.")
|
|
return exit_node, exit_unet, exit_clip
|
|
|
|
if exit_clip is not None:
|
|
logger.info("Using CLIP with loras.")
|
|
for lora in lora_stack:
|
|
logger.info(f"Applying lora: {lora[0]}, unet: {lora[1]}, clip: {lora[2]}")
|
|
exit_node = graph.node("LoraLoader", model=exit_unet, clip=exit_clip, lora_name=lora[0], strength_model=lora[1], strength_clip=lora[2])
|
|
exit_unet = exit_node.out(0)
|
|
exit_clip = exit_node.out(1)
|
|
return exit_node, exit_unet, exit_clip
|
|
|
|
logger.info("Using Model Only loras.")
|
|
for lora in lora_stack:
|
|
logger.info(f"Applying lora: {lora[0]}, unet: {lora[1]}")
|
|
exit_node = graph.node("LoraLoaderModelOnly", model=exit_unet, lora_name=lora[0], strength_model=lora[1])
|
|
exit_unet = exit_node.out(0)
|
|
return exit_node, exit_unet, exit_clip
|
|
|
|
# Add an actual model shift node.
|
|
def create_shift_node(graph: GraphBuilder, unet_in, model_shifts):
|
|
exit_node = None
|
|
unet_out = unet_in
|
|
|
|
if model_shifts is not None and model_shifts["shift_type"] != "None":
|
|
if model_shifts["shift_type"] == "x1":
|
|
logger.info(f"Applying x1 shift - AuraFlow/Lumina2. Shift: {model_shifts['shift']}")
|
|
exit_node = graph.node("ModelSamplingAuraFlow", model=unet_out, shift=model_shifts["shift"])
|
|
unet_out = exit_node.out(0)
|
|
elif model_shifts["shift_type"] == "x1000":
|
|
logger.info(f"Applying x1000 shift - SD3. Shift: {model_shifts['shift']}")
|
|
exit_node = graph.node("ModelSamplingSD3", model=unet_out, shift=model_shifts["shift"])
|
|
unet_out = exit_node.out(0)
|
|
|
|
return exit_node, unet_out
|
|
|
|
# Add FreeU v2 node if specified.
|
|
def create_freeu_v2_node(graph: GraphBuilder, unet_in, model_shifts):
|
|
exit_node = None
|
|
unet_out = unet_in
|
|
|
|
if model_shifts is not None and model_shifts.get("freeu_v2", False) == True:
|
|
logger.info(f"FreeU v2 is enabled, applying to model. b1: {model_shifts['b1']}, b2: {model_shifts['b2']}, s1: {model_shifts['s1']}, s2: {model_shifts['s2']}")
|
|
exit_node = graph.node("FreeU_V2",
|
|
model=unet_out,
|
|
b1=model_shifts["b1"],
|
|
b2=model_shifts["b2"],
|
|
s1=model_shifts["s1"],
|
|
s2=model_shifts["s2"]
|
|
)
|
|
unet_out = exit_node.out(0)
|
|
|
|
return exit_node, unet_out
|
|
|
|
# Add model shift nodes, which I'm including FreeU v2 in.
|
|
def create_model_shift_nodes(graph, unet_in, model_shifts):
|
|
exit_node = None
|
|
unet_out = unet_in
|
|
if model_shifts is None:
|
|
logger.info("No model shifts to apply.")
|
|
return exit_node, unet_out
|
|
|
|
exit_node, unet_out = create_shift_node(graph, unet_out, model_shifts)
|
|
exit_node, unet_out = create_freeu_v2_node(graph, unet_out, model_shifts)
|
|
return exit_node, unet_out
|
|
|
|
# Add lora nodes as well as model shift nodes.
|
|
def create_lora_shift_nodes(graph: GraphBuilder, unet_in, clip_in=None, lora_stack=None, model_shifts=None):
|
|
exit_node, exit_unet, exit_clip = create_lora_nodes(graph, unet_in, clip_in, lora_stack)
|
|
exit_node, exit_unet = create_model_shift_nodes(graph, exit_unet, model_shifts)
|
|
return exit_node, exit_unet, exit_clip
|
|
|
|
# Add nodes to add to a lora stack.
|
|
def add_lora_stack_node(graph: GraphBuilder, args, idx, lora_stack = None):
|
|
lora_enabled = args[f"enabled_{idx}"]
|
|
lora_name = args[f"lora_{idx}_name"]
|
|
model_weight = args[f"model_{idx}_weight"]
|
|
# If clip weight is not provided, it defaults to 1.0
|
|
clip_weight = args.get(f"clip_{idx}_weight", 1.0)
|
|
logger.info(f"Creating Lora stack node: {lora_name}, enabled: {lora_enabled}, model_weight: {model_weight}, clip_weight: {clip_weight}")
|
|
|
|
return graph.node("Sage_LoraStack",
|
|
enabled = lora_enabled,
|
|
lora_name = lora_name,
|
|
model_weight = model_weight,
|
|
clip_weight = clip_weight,
|
|
lora_stack = lora_stack
|
|
)
|