Budi Hartono
36d1667eb0
- EmptyLatentAspectPreset gains a "model" widget; web/aspect_ratio_filter.js filters the "preset" combo to that model client-side, parsed from each preset's label so presets.py stays the single source of truth. - Fix latent channel count: SD1.5/SDXL use 4-channel latents, but Flux/HiDream/Krea/Qwen-Image use 16-channel latents (same family as ComfyUI's EmptySD3LatentImage) — was previously hardcoded to 4 for all models, breaking/corrupting generations on the newer models. - Drop Ideogram and ERNIE from presets.py: both are hosted API models with no local LATENT/diffusion pipeline in ComfyUI, so listing them implied false compatibility.
116 lines
4.1 KiB
Python
116 lines
4.1 KiB
Python
import torch
|
|
from .presets import PRESETS
|
|
|
|
def _validate_dim(v: int):
|
|
if v <= 0 or v % 8 != 0:
|
|
raise ValueError("Dimension must be >0 and divisible by 8")
|
|
|
|
|
|
class EmptyLatentAspectPreset:
|
|
"""Creates a blank latent using one of the predefined presets."""
|
|
# Built once at import time and shared by all instances/calls.
|
|
PRESET_MAP = {
|
|
f"{w}x{h} - {lbl} - {model}": (w, h)
|
|
for model, lbl, w, h in PRESETS
|
|
}
|
|
# Unique models in first-appearance order; the "model" widget below is purely a
|
|
# client-side filter (see web/aspect_ratio_filter.js) for the "preset" dropdown,
|
|
# which already encodes the model in its label — parsed there, not duplicated.
|
|
MODELS = list(dict.fromkeys(model for model, _, _, _ in PRESETS))
|
|
|
|
# Latent channel count per model family. SD1.5/SDXL use the legacy 4-channel VAE;
|
|
# Flux/HiDream/Qwen-Image/Krea use a 16-channel VAE (same family as ComfyUI's own
|
|
# EmptySD3LatentImage) — feeding a 4-channel latent into those samplers fails or
|
|
# silently produces garbage.
|
|
LATENT_CHANNELS = {
|
|
"SD15": 4,
|
|
"SDXL": 4,
|
|
"HiDream": 16,
|
|
"Flux.1": 16,
|
|
"Flux.2": 16,
|
|
"Krea": 16,
|
|
"Qwen-Image": 16,
|
|
}
|
|
DEFAULT_LATENT_CHANNELS = 4
|
|
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"model": (cls.MODELS,),
|
|
"preset": (list(cls.PRESET_MAP.keys()),),
|
|
"batch_size": ("INT", {"default": 1, "min": 1})
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("LATENT", "INT", "INT")
|
|
RETURN_NAMES = ("LATENT", "width", "height")
|
|
FUNCTION = "generate"
|
|
CATEGORY = "latent" # moved into ComfyUI's built-in "latent" category
|
|
|
|
def generate(self, model: str, preset: str, batch_size: int):
|
|
if preset not in self.PRESET_MAP:
|
|
raise ValueError(f"Unknown preset: {preset}")
|
|
w, h = self.PRESET_MAP[preset]
|
|
|
|
_validate_dim(w)
|
|
_validate_dim(h)
|
|
|
|
channels = self.LATENT_CHANNELS.get(model, self.DEFAULT_LATENT_CHANNELS)
|
|
latent = torch.zeros([batch_size, channels, h // 8, w // 8], dtype=torch.float32)
|
|
return ({"samples": latent}, w, h)
|
|
|
|
|
|
class EmptyLatentAspectByAxis:
|
|
"""Creates a blank latent by fixing one axis and computing the other from an aspect ratio."""
|
|
ASPECT_CHOICES = [
|
|
("1:1 Square", (1, 1)),
|
|
("3:2 Landscape", (3, 2)),
|
|
("2:3 Portrait", (2, 3)),
|
|
("4:3 Landscape", (4, 3)),
|
|
("3:4 Portrait", (3, 4)),
|
|
("16:9 Landscape", (16, 9)),
|
|
("9:16 Portrait", (9, 16)),
|
|
("5:4 Landscape", (5, 4)),
|
|
("4:5 Portrait", (4, 5)),
|
|
("21:9 Widescreen",(21, 9)),
|
|
("9:21 Portrait", (9, 21)),
|
|
("7:5 Landscape", (7, 5)),
|
|
("5:7 Portrait", (5, 7)),
|
|
]
|
|
REFERENCE_CHOICES = ["Width", "Height"]
|
|
|
|
@classmethod
|
|
def INPUT_TYPES(cls):
|
|
return {
|
|
"required": {
|
|
"primary_dim": ("INT", {"default": 512, "min": 8}),
|
|
"reference": (cls.REFERENCE_CHOICES,),
|
|
"aspect_ratio": ([lbl for lbl,_ in cls.ASPECT_CHOICES],),
|
|
"batch_size": ("INT", {"default": 1, "min": 1})
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("LATENT", "INT", "INT")
|
|
RETURN_NAMES = ("LATENT", "width", "height")
|
|
FUNCTION = "generate"
|
|
CATEGORY = "latent" # now appears under the built-in latent category
|
|
|
|
RATIO_MAP = dict(ASPECT_CHOICES)
|
|
|
|
def generate(self, primary_dim: int, reference: str, aspect_ratio: str, batch_size: int):
|
|
if aspect_ratio not in self.RATIO_MAP:
|
|
raise ValueError(f"Unknown aspect ratio: {aspect_ratio}")
|
|
wr, hr = self.RATIO_MAP[aspect_ratio]
|
|
|
|
_validate_dim(primary_dim)
|
|
if reference == "Width":
|
|
w, h = primary_dim, round(primary_dim * hr / wr)
|
|
else:
|
|
h, w = primary_dim, round(primary_dim * wr / hr)
|
|
|
|
_validate_dim(w)
|
|
_validate_dim(h)
|
|
|
|
latent = torch.zeros([batch_size, 4, h // 8, w // 8], dtype=torch.float32)
|
|
return ({"samples": latent}, w, h) |