Sammir Alejandro Femat RoblesandClaude Opus 4.8
1dfe723e1a
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
102 lines
4.9 KiB
Python
102 lines
4.9 KiB
Python
# Copyright (c) 2026 — ComfyUI-BerniniR. Apache-2.0.
|
||
"""Constantes canónicas de Bernini-R (tomadas de bernini/cli.py y config.json).
|
||
|
||
Fuente de verdad: github.com/bytedance/Bernini (las DEFAULTS del CLI ganan
|
||
sobre las defaults internas del pipeline)."""
|
||
|
||
# --- Defaults del modelo (config.json del repo Bernini-R) ---------------------
|
||
SWITCH_DIT_BOUNDARY = 0.875 # frontera high/low-noise (× num_train_timesteps)
|
||
MAX_SEQUENCE_LENGTH = 512 # tokens de texto (UMT5), zero-padded a 512
|
||
SHIFT = 3.0 # flow shift EFECTIVO con UniPC (NO 5.0; ver nota)
|
||
USE_UNIPC = True
|
||
USE_SRC_ID_ROTARY_EMB = True
|
||
NUM_TRAIN_TIMESTEPS = 1000
|
||
|
||
# --- Defaults de generación (CLI; canónicos de cara al usuario) ---------------
|
||
DEFAULTS = dict(
|
||
num_frames=81,
|
||
num_inference_steps=40,
|
||
max_image_size=848,
|
||
height=480,
|
||
width=848,
|
||
omega_V=1.25,
|
||
omega_I=4.5,
|
||
omega_TI=4.0,
|
||
omega_scale=0.8, # escala los 3 omegas UNA vez, al cambiar a low-noise
|
||
flow_shift=5.0, # SOLO se aplica en el scheduler FlowMatch (no-UniPC). Ver nota.
|
||
eta=0.5, # peso de la componente paralela en APG
|
||
norm_threshold=(50.0, 50.0, 50.0),
|
||
momentum=0.0,
|
||
seed=42,
|
||
fps=16,
|
||
)
|
||
|
||
# NOTA IMPORTANTE sobre el shift:
|
||
# Con use_unipc=True (default del modelo), UniPCMultistepScheduler fija su
|
||
# flow_shift = config.shift = 3.0 en construcción, y el flow_shift=5.0 del CLI
|
||
# queda MUERTO. Reproducimos esto: el shift efectivo por defecto es 3.0.
|
||
|
||
# --- Negative prompt estándar de Wan2.2 (cli.DEFAULT_NEG_PROMPT) --------------
|
||
DEFAULT_NEG_PROMPT = (
|
||
"色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，"
|
||
"最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，"
|
||
"画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，"
|
||
"杂乱的背景，三条腿，背景人很多，倒着走"
|
||
)
|
||
|
||
# --- System prompts por tarea (prompt_enhancer.SYSTEM_PROMPTS) -----------------
|
||
# OJO: en Bernini-R estos strings se CONCATENAN literalmente (sin espacio)
|
||
# delante del prompt positivo antes de codificar con UMT5; NO van en el negativo.
|
||
SYSTEM_PROMPTS = {
|
||
"default": "You are a helpful assistant.",
|
||
"t2i": "You are a helpful assistant specialized in text-to-image generation.",
|
||
"t2v": "You are a helpful assistant specialized in text-to-video generation.",
|
||
"i2i": "You are a helpful assistant specialized in image editing.",
|
||
"r2i": "You are a helpful assistant specialized in subject-to-image generation.",
|
||
"i2v": "You are a helpful assistant specialized in image-to-video generation.",
|
||
"v2v": "You are a helpful assistant specialized in video editing.",
|
||
"r2v": "You are a helpful assistant specialized in subject-to-video generation.",
|
||
"vi2v": "You are a helpful assistant specialized in video editing on content propagation.",
|
||
"rv2v": "You are a helpful assistant specialized in video editing with reference.",
|
||
"ads2v": "You are a helpful assistant specialized in ads insertion.",
|
||
"vrc2v": "You are a helpful assistant for editing. You may need to adjust the subject's action or position.",
|
||
"mv2v": ("You are a helpful assistant for editing. You might need to adjust the video's "
|
||
"style, lighting, colors, textures, and the subject's pose or action."),
|
||
}
|
||
|
||
|
||
def get_system_prompt_for_task(task_type: str) -> str:
|
||
return SYSTEM_PROMPTS.get(task_type, SYSTEM_PROMPTS["default"])
|
||
|
||
|
||
# --- Tipos de tarea y modos de guía -------------------------------------------
|
||
TASK_TYPES = ["t2i", "i2i", "t2v", "v2v", "mv2v", "rv2v", "r2v"]
|
||
GUIDANCE_MODES = ["rv2v", "v2v", "v2v_chain", "t2v", "r2v_apg", "v2v_apg", "t2v_apg"]
|
||
|
||
# guidance_mode recomendado por task_type (cuando el usuario no lo fija a mano).
|
||
TASK_TO_GUIDANCE = {
|
||
"t2i": "t2v", # t2i = t2v con num_frames=1
|
||
"t2v": "t2v",
|
||
"i2i": "v2v", # imagen única -> stream VI, CFG sobre texto
|
||
"v2v": "v2v",
|
||
"mv2v": "v2v_apg", # cambia movimiento del sujeto -> APG ayuda
|
||
"rv2v": "rv2v", # video + referencia(s) -> CFG encadenado de 4 forwards
|
||
"r2v": "r2v_apg", # solo referencias -> APG encadenado en x-space
|
||
}
|
||
|
||
# Para cada guidance_mode: qué "combos" de stream y texto se usan en cada forward,
|
||
# y cómo se combinan. Los combos posibles son:
|
||
# none = solo target ; V = (1er video)+target ; I = imágenes+target ;
|
||
# VI = videos+imágenes+target
|
||
# y el texto es "uncond" (neg) o "cond" (pos). El plan exacto vive en sampler.py;
|
||
# esta tabla solo declara cuántos forwards y si usa APG (x-space).
|
||
GUIDANCE_PLAN = {
|
||
"rv2v": dict(n_fwd=4, apg=False),
|
||
"v2v": dict(n_fwd=2, apg=False),
|
||
"v2v_chain": dict(n_fwd=3, apg=False),
|
||
"t2v": dict(n_fwd=2, apg=False),
|
||
"r2v_apg": dict(n_fwd=3, apg=True),
|
||
"v2v_apg": dict(n_fwd=2, apg=True),
|
||
"t2v_apg": dict(n_fwd=2, apg=True),
|
||
}
|