File size: 33,914 Bytes
f8c73f9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 | """Native PySide Oasis Action Player, backed by the connected Oasis project."""
from __future__ import annotations
import importlib.util
import io
import json
import math
import sys
import threading
import time
import traceback
from pathlib import Path
from PIL import Image, ImageDraw, ImageOps
from PySide6.QtCore import QEvent, QThread, Qt, QTimer, Signal
from PySide6.QtGui import QPixmap
from PySide6.QtWidgets import (
QCheckBox, QComboBox, QFileDialog, QFormLayout, QFrame, QHBoxLayout,
QLabel, QLineEdit, QMessageBox, QPushButton, QScrollArea, QSlider,
QSpinBox, QVBoxLayout, QWidget,
)
from adam.oasis_player import capture_path, discover_models, frame_size, model_info, random_roblox_dataset_frame
def _external_player_module(root: Path):
script = root / "roblox_action_flow_app.py"
if not script.is_file():
raise RuntimeError("Oasis player code was not found. Connect Oasis-Game-Trainer in Settings.")
# The connected project owns oasis_input.py, temporal_oasis.py and its VAE.
# Its directory must be importable before executing roblox_action_flow_app.py.
if str(root) not in sys.path:
sys.path.insert(0, str(root))
spec = importlib.util.spec_from_file_location("adam_oasis_external_player", script)
if spec is None or spec.loader is None:
raise RuntimeError("ADAM could not load the Oasis player helpers.")
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
return module
class OasisPlaybackWorker(QThread):
status = Signal(str)
frames_ready = Signal(object, float, str, str)
comparison_ready = Signal(bytes, str)
failed = Signal(str)
def __init__(self, adam_root: Path, oasis_root: Path, model: Path, image_bytes: bytes, settings: dict, state: dict, lock: threading.Lock, input_timeline=None):
super().__init__()
# The player code and trained models intentionally stay in Oasis, while
# captures belong to ADAM. These paths must never be interchangeable.
self.adam_root, self.oasis_root = adam_root, oasis_root
self.model, self.image_bytes = model, image_bytes
self.settings, self.state, self.lock = settings, state, lock
self.input_timeline = input_timeline
self._running = True
self._compare = threading.Event()
def stop(self):
self._running = False
def request_comparison(self):
self._compare.set()
@staticmethod
def _png(image: Image.Image) -> bytes:
data = io.BytesIO()
image.save(data, format="PNG")
return data.getvalue()
def _snapshot(self):
with self.lock:
settings = dict(self.settings)
state = dict(self.state)
self.state["mouse_dx"] = 0.0
self.state["mouse_dy"] = 0.0
self.state["zoom"] = 0.0
return settings, state
def _action(self, external, info: dict, state: dict):
binary = external.BINARY_ACTION_NAMES
enabled = set(info.get("enabled_action_names") or info.get("action_names") or external.ACTION_NAMES)
require_drag = bool(self.settings.get("require_drag", True))
camera_allowed = not require_drag or bool(state.get("right_mouse", False))
mouse_dx = max(-1.0, min(1.0, float(state.get("mouse_dx", 0.0)) / external.PLAYER_MOUSE_SCALE)) if camera_allowed else 0.0
mouse_dy = max(-1.0, min(1.0, float(state.get("mouse_dy", 0.0)) / external.PLAYER_MOUSE_SCALE)) if camera_allowed else 0.0
zoom = max(-1.0, min(1.0, float(state.get("zoom", 0.0)) / external.PLAYER_ZOOM_SCALE))
held = set(state.get("keys", set()))
values = [1.0 if ({"jump": "space"}.get(name, name) in held) else 0.0 for name in binary]
values += [mouse_dx, mouse_dy, zoom]
values = [value if name in enabled else 0.0 for name, value in zip(external.ACTION_NAMES, values)]
moving = any(value > .5 for value in values[:len(binary)]) or any(abs(value) > .02 for value in values[-3:])
labels = [external.ACTION_DISPLAY_NAMES.get(name, name.upper()) for name, value in zip(binary, values) if value > .5]
return values, moving, labels
def _comparison(self, external, source, model, previous, info, torch, dtype, device, generator):
actions = [("Idle", [0.0] * external.ACTION_DIM)]
enabled = set(info.get("enabled_action_names") or info.get("action_names") or external.ACTION_NAMES)
for index, name in enumerate(external.BINARY_ACTION_NAMES):
if name in enabled:
values = [0.0] * external.ACTION_DIM
values[index] = 1.0
actions.append((external.ACTION_DISPLAY_NAMES.get(name, name.upper()), values))
actions = actions[:9]
width, height = frame_size(info)
panel = Image.new("RGB", (width * 3, (height + 24) * math.ceil(len(actions) / 3)), (9, 19, 29))
draw = ImageDraw.Draw(panel)
base_noise = torch.randn(previous.shape, generator=generator, device=device, dtype=dtype)
for index, (label, values) in enumerate(actions):
action = torch.tensor([values], device=device, dtype=dtype)
generated = external.sample_guided_action_frame(
model, previous, action, int(self.settings["steps"]), device, dtype, generator,
method=self.settings["method"], initial_noise=base_noise.clone(),
action_input_scale=float(info.get("action_input_scale", 1.0)), guidance=float(self.settings["guidance"]),
)
image = source.tensor_to_pil(generated[0].clamp(-1, 1))
x, y = (index % 3) * width, (index // 3) * (height + 24)
panel.paste(image, (x, y + 24)); draw.text((x + 5, y + 5), label, fill=(237, 245, 251))
path = capture_path(self.adam_root, "oasis_action_compare")
panel.save(path)
self.comparison_ready.emit(self._png(panel), str(path))
def _run_temporal(self, external, info: dict, current: Image.Image, torch) -> None:
"""Run the external temporal format without applying legacy guidance/noise UI tricks."""
if self.input_timeline is None:
raise RuntimeError("Temporal input timing was not initialized.")
if info.get("action_names") != external.ACTION_NAMES:
raise RuntimeError("This temporal Oasis model uses a control schema ADAM does not recognize.")
from oasis_input import interval_vector
pixel_engine = info.get("model_type") == "action_conditioned_temporal_pixel_flow"
if pixel_engine:
from temporal_pixel_oasis import TemporalPixelPlayback as TemporalPlayback, load_model
else:
from temporal_oasis import TemporalPlayback, load_model
from vae_cpu_lite import load_vae
import numpy as np
required_source = info.get("mouse_input_source")
if required_source == "windows_raw_input" and not bool(self.settings.get("raw_mouse_ready")):
raise RuntimeError("This model was trained with raw mouse counts. Raw mouse input is unavailable in ADAM, so cursor pixels will not be substituted.")
device, dtype = torch.device("cuda" if torch.cuda.is_available() else "cpu"), torch.float32
self.status.emit("Loading temporal Oasis model…")
model = load_model(self.model, device, dtype)
vae = None if pixel_engine else load_vae(self.model, device, dtype)
pixels = torch.from_numpy(np.array(current.convert("RGB"), copy=True)).permute(2, 0, 1).float() / 127.5 - 1
generator = torch.Generator(device=device).manual_seed(int(self.settings["seed"]))
world = (TemporalPlayback(model, info, pixels[None], device, dtype, generator) if pixel_engine
else TemporalPlayback(model, vae, info, pixels[None], device, dtype, generator))
dt = float(info["step_seconds"])
self.status.emit(f"Temporal preview ready: {info['context_frames']} frames of history, {dt:.4f}s per world step.")
world_step = 0
while self._running:
deadline = time.perf_counter() + dt
while self._running and time.perf_counter() < deadline:
time.sleep(min(0.01, max(0.0, deadline - time.perf_counter())))
if not self._running:
break
settings, _state = self._snapshot()
interval = self.input_timeline.snapshot(time.perf_counter())
if not interval["valid"]:
raise RuntimeError("Temporal input timing overflowed. Stop and restart the player.")
values = interval_vector(interval, external.BINARY_ACTION_NAMES, info["mouse_scale"], info["scroll_scale"])
has_input = any(value > 1e-6 for value in values[:len(external.BINARY_ACTION_NAMES)]) or any(
abs(value) > 1e-6 for value in values[len(external.BINARY_ACTION_NAMES):]
)
if not settings.get("temporal_advance_idle", True) and not has_input:
self.status.emit("Temporal preview paused for no input. Move the mouse or press a control to advance.")
continue
snapshot = world.snapshot()
rgb = world.step(values, int(settings["steps"]), str(settings["method"]))
if not self._running:
world.restore(snapshot)
break
array = ((rgb[0].detach().float().cpu().clamp(-1, 1) + 1) * 127.5).round().byte().permute(1, 2, 0).numpy()
current = Image.fromarray(array)
world_step += 1
text = f"Temporal input | mouse {values[-3]:+.2f}, {values[-2]:+.2f} | world time {world_step * dt:.2f}s"
self.frames_ready.emit([self._png(current)], 1 / dt, text, "Learned history; raw generated frames")
def run(self):
try:
import torch
external = _external_player_module(self.oasis_root)
info = model_info(self.model)
width, height = frame_size(info)
current = ImageOps.fit(Image.open(io.BytesIO(self.image_bytes)).convert("RGB"), (width, height), method=Image.Resampling.LANCZOS)
if info.get("model_type") in {"action_conditioned_temporal_latent_flow", "action_conditioned_temporal_pixel_flow"}:
import torch
self._run_temporal(external, info, current, torch)
return
source, _ = external.load_source_module()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.float16 if device.type == "cuda" else torch.float32
self.status.emit("Loading Oasis action model…")
model = external.load_action_unet(self.model, device, dtype)
generator = torch.Generator(device=device).manual_seed(int(self.settings["seed"]))
noise_state = None
self.status.emit(f"Running {info.get('name', self.model.name)} on {device.type.upper()}.")
while self._running:
started = time.perf_counter()
settings, state = self._snapshot()
previous_frame = current.copy()
previous = source.pil_to_normalized_tensor(current, device, dtype)
values, moving, labels = self._action(external, info, state)
if not moving and not settings["animate_idle"]:
result = current.copy()
else:
refresh = float(settings["motion_noise"]) * (float(settings["active_noise"]) if moving else float(settings["idle_noise"]))
if noise_state is None or tuple(noise_state.shape) != tuple(previous.shape):
noise_state = torch.randn(previous.shape, generator=generator, device=device, dtype=dtype)
elif refresh:
fresh = torch.randn(previous.shape, generator=generator, device=device, dtype=dtype)
noise_state = noise_state * math.sqrt(max(0.0, 1 - refresh * refresh)) + fresh * refresh
action = torch.tensor([external.subdivide_action_values(values, int(settings["subdivision"]))], device=device, dtype=dtype)
generated = external.sample_guided_action_frame(model, previous, action, int(settings["steps"]), device, dtype, generator, method=settings["method"], initial_noise=noise_state.detach().clone(), action_input_scale=float(info.get("action_input_scale", 1.0)), guidance=float(settings["guidance"]))
result = source.tensor_to_pil(generated[0])
if settings["stability"]:
try: result = external.motion_aware_stabilize(current, result, float(settings["stability"]))
except Exception: pass
current = result
if self._compare.is_set():
self._compare.clear(); self._comparison(external, source, model, previous, info, torch, dtype, device, generator)
fps = max(.5, float(settings["fps"])); period = 1 / fps
try:
display = external.interpolate_motion_frames(previous_frame, current, max(1, int(settings["display_multiplier"]) if moving else 1))
except Exception:
display = [current]
text = "Guided input: " + (" + ".join(labels) or "IDLE") + f" | yaw {values[-3]:+.2f} pitch {values[-2]:+.2f} zoom {values[-1]:+.2f}"
self.frames_ready.emit([self._png(frame) for frame in display], fps, text, f"Motion-aware stability {float(settings['stability']) * 100:.0f}%")
remaining = period - (time.perf_counter() - started)
if remaining > 0: time.sleep(remaining)
except Exception:
self.failed.emit(traceback.format_exc())
class OasisPlayerPage(QWidget):
"""The playable Oasis experience; training deliberately remains in Training Studio."""
def __init__(self, root, assets, config):
super().__init__()
self.root, self.assets, self.config = Path(root), assets, config
self.models: dict[str, Path] = {}; self.frame: bytes | None = None; self.worker = None
self.state = {"keys": set(), "mouse_dx": 0.0, "mouse_dy": 0.0, "zoom": 0.0, "right_mouse": False}
self.lock = threading.Lock(); self.sequence = 0
self.temporal_timeline = None; self.raw_mouse_input = None; self.temporal_mouse_source = None
self._restoring = True
self._build(); self.refresh(); self._restore()
self._restoring = False
self._model_changed()
def _build(self):
layout = QVBoxLayout(self); layout.setContentsMargins(22, 18, 22, 18)
title = QLabel("Oasis Player"); title.setObjectName("PageTitle"); layout.addWidget(title)
sub = QLabel("Explore a trained action-world model from a starting screenshot. Training stays in Training Studio."); sub.setProperty("muted", True); layout.addWidget(sub)
body = QHBoxLayout(); layout.addLayout(body, 1)
controls = QFrame(); controls.setProperty("card", True); controls.setMaximumWidth(350)
cl = QVBoxLayout(controls); cl.setContentsMargins(15, 15, 15, 15)
label = QLabel("PLAYER CONTROLS"); label.setObjectName("CardTitle"); cl.addWidget(label)
self.model = QComboBox(); self.model.currentIndexChanged.connect(self._model_changed); cl.addWidget(self.model)
refresh = QPushButton("Refresh models"); refresh.clicked.connect(self.refresh); cl.addWidget(refresh)
self.image = QLineEdit(); self.image.setPlaceholderText("Choose a starting game screenshot"); cl.addWidget(self.image)
browse = QPushButton("Choose starting image"); browse.clicked.connect(self._browse_image); cl.addWidget(browse)
self.random_dataset = QCheckBox("Use a random Roblox Dataset frame when starting"); self.random_dataset.setChecked(True); cl.addWidget(self.random_dataset)
self.random_dataset.toggled.connect(lambda _checked: self._save())
find_frame = QPushButton("Find New Frame"); find_frame.clicked.connect(self._find_new_frame); cl.addWidget(find_frame)
form = QFormLayout(); self.steps = self._combo(["1", "2", "4", "6", "8", "10", "12"], "4"); self.method = self._combo(["Euler", "Heun"], "Euler"); self.guidance = self._combo(["1.0", "1.5", "2.0", "2.5", "3.0", "3.5"], "2.0"); self.fps = self._combo(["2", "4", "6", "8", "12", "24"], "12"); self.subdivision = self._combo(["1", "2", "3", "4"], "3"); self.multiplier = self._combo(["1", "2", "3", "4"], "1")
for name, field in (("Flow steps", self.steps), ("Denoise method", self.method), ("Action guidance", self.guidance), ("Gameplay FPS", self.fps), ("Motion subdivision", self.subdivision), ("Display interpolation", self.multiplier)): form.addRow(name, field)
cl.addLayout(form)
self.stability = QSlider(Qt.Horizontal); self.stability.setRange(0, 100); self.stability.setValue(25); cl.addWidget(QLabel("Motion-aware stability")); cl.addWidget(self.stability)
self.idle = QCheckBox("Keep no-input scenes animated"); self.idle.setChecked(True); self.drag = QCheckBox("Camera requires right-mouse drag"); self.drag.setChecked(True); cl.addWidget(self.idle); cl.addWidget(self.drag)
self.temporal_idle = QCheckBox("Temporal: advance during no-input"); self.temporal_idle.setChecked(True); cl.addWidget(self.temporal_idle)
self.temporal_hint = QLabel("When enabled, idle time tests coasting, falling, and other learned dynamics. When disabled, the preview waits for an input."); self.temporal_hint.setProperty("muted", True); self.temporal_hint.setWordWrap(True); cl.addWidget(self.temporal_hint)
self.start = QPushButton("Start Action Player"); self.start.setProperty("primary", True); self.start.clicked.connect(self._toggle); cl.addWidget(self.start)
reset = QPushButton("Reset frame"); reset.clicked.connect(self._reset_frame); cl.addWidget(reset)
capture = QPushButton("Capture frame"); capture.clicked.connect(self._capture); compare = QPushButton("Compare actions (same noise)"); compare.clicked.connect(self._compare); cl.addWidget(capture); cl.addWidget(compare); cl.addStretch()
body.addWidget(controls)
display = QFrame(); display.setProperty("card", True); dl = QVBoxLayout(display); dl.setContentsMargins(12, 12, 12, 12)
top = QHBoxLayout(); self.status = QLabel("Choose an Oasis model and starting image."); self.status.setProperty("muted", True); self.metrics = QLabel("AI FPS —"); self.metrics.setProperty("accent", True); top.addWidget(self.status); top.addStretch(); top.addWidget(self.metrics); dl.addLayout(top)
self.preview = QLabel("Select a model and screenshot to begin."); self.preview.setAlignment(Qt.AlignCenter); self.preview.setMinimumSize(620, 420); self.preview.setStyleSheet("background: #03080d; border: 1px solid #183246; border-radius: 8px;"); self.preview.setFocusPolicy(Qt.StrongFocus); self.preview.installEventFilter(self); dl.addWidget(self.preview, 1)
self.input = QLabel("Guided input: IDLE"); self.input.setProperty("accent", True); dl.addWidget(self.input); body.addWidget(display, 1)
for control in (self.steps, self.method, self.guidance, self.fps, self.subdivision, self.multiplier):
control.currentTextChanged.connect(lambda _value: self._save())
self.stability.valueChanged.connect(lambda _value: self._save())
self.idle.toggled.connect(lambda _checked: self._save())
self.temporal_idle.toggled.connect(lambda _checked: self._save())
self.drag.toggled.connect(lambda _checked: self._save())
@staticmethod
def _combo(items, current):
box = QComboBox(); box.addItems(items); box.setCurrentText(current); return box
def refresh(self):
oasis = Path(str(self.config.get("tool_folders", {}).get("oasis_trainer", "")))
current = self.model.currentData(); self.models = dict(discover_models(oasis, self.assets.assets)) if oasis.is_dir() else {}
self.model.blockSignals(True); self.model.clear()
for name, path in self.models.items(): self.model.addItem(name, str(path))
if current:
index = self.model.findData(current)
if index >= 0: self.model.setCurrentIndex(index)
self.model.blockSignals(False); self._model_changed()
def _restore(self):
saved = self.config.get("oasis_player_settings", {})
if isinstance(saved, dict):
self.image.setText(str(saved.get("starting_image", "")))
self.random_dataset.setChecked(bool(saved.get("random_roblox_dataset", True)))
for key, control in (
("steps", self.steps), ("method", self.method), ("guidance", self.guidance),
("fps", self.fps), ("subdivision", self.subdivision), ("display_multiplier", self.multiplier),
):
value = str(saved.get(key, ""))
if control.findText(value) >= 0:
control.setCurrentText(value)
try:
self.stability.setValue(max(0, min(100, int(saved.get("stability", 25)))))
except (TypeError, ValueError):
pass
self.idle.setChecked(bool(saved.get("animate_idle", True)))
self.temporal_idle.setChecked(bool(saved.get("temporal_advance_idle", True)))
self.drag.setChecked(bool(saved.get("require_drag", True)))
index = self.model.findData(str(saved.get("model", "")))
if index >= 0: self.model.setCurrentIndex(index)
def _save(self):
if getattr(self, "_restoring", False):
return
self.config.update({"oasis_player_settings": {
"model": self.model.currentData() or "",
"starting_image": self.image.text(),
"random_roblox_dataset": self.random_dataset.isChecked(),
"steps": self.steps.currentText(),
"method": self.method.currentText(),
"guidance": self.guidance.currentText(),
"fps": self.fps.currentText(),
"subdivision": self.subdivision.currentText(),
"display_multiplier": self.multiplier.currentText(),
"stability": self.stability.value(),
"animate_idle": self.idle.isChecked(),
"temporal_advance_idle": self.temporal_idle.isChecked(),
"require_drag": self.drag.isChecked(),
}})
def _model_changed(self):
path = self.model.currentData()
temporal = False
if path:
try:
temporal = model_info(path).get("model_type") in {"action_conditioned_temporal_latent_flow", "action_conditioned_temporal_pixel_flow"}
except ValueError:
pass
self.status.setText(f"Ready: {self.model.currentText()} · choose a matching starting screenshot.")
self.temporal_idle.setVisible(temporal); self.temporal_hint.setVisible(temporal)
self._save()
def _browse_image(self):
path, _ = QFileDialog.getOpenFileName(self, "Choose starting screenshot", "", "Images (*.png *.jpg *.jpeg *.webp *.bmp)")
if path: self.image.setText(path); self._reset_frame()
def _reset_frame(self):
try:
path = Path(self.image.text())
info = model_info(self.model.currentData()); width, height = frame_size(info)
with Image.open(path) as image:
image = ImageOps.fit(image.convert("RGB"), (width, height), method=Image.Resampling.LANCZOS)
output = io.BytesIO(); image.save(output, format="PNG"); self.frame = output.getvalue()
self._show(self.frame); self.status.setText("Frame reset. Press Start Action Player."); self._save()
except Exception as exc: QMessageBox.warning(self, "Starting image needed", str(exc))
def _settings(self):
return {"seed": int(time.time() * 1000) % 1_000_000_000, "steps": int(self.steps.currentText()), "method": self.method.currentText(), "guidance": float(self.guidance.currentText()), "fps": float(self.fps.currentText()), "subdivision": int(self.subdivision.currentText()), "display_multiplier": int(self.multiplier.currentText()), "stability": self.stability.value() / 100, "animate_idle": self.idle.isChecked(), "temporal_advance_idle": self.temporal_idle.isChecked(), "require_drag": self.drag.isChecked(), "motion_noise": .06, "active_noise": 1.0, "idle_noise": .18, "raw_mouse_ready": bool(self.raw_mouse_input and self.raw_mouse_input.available)}
def _oasis_root(self) -> Path:
raw = self.config.get("tool_folders", {}).get("oasis_trainer", "")
folder = Path(str(raw)).expanduser()
if not (folder / "roblox_action_flow_app.py").is_file():
raise ValueError("Oasis Game Trainer is not configured with roblox_action_flow_app.py. Check Settings → Tool Folders.")
return folder.resolve()
def _find_new_frame(self) -> bool:
"""Find and immediately prepare a new nested Roblox recording frame."""
try:
selected = random_roblox_dataset_frame(self._oasis_root())
self.image.setText(str(selected))
self._reset_frame()
self.status.setText(f"Random Roblox Dataset frame ready: {selected.name}")
return self.frame is not None
except (OSError, ValueError) as exc:
QMessageBox.warning(self, "Roblox Dataset frame unavailable", str(exc))
return False
def _toggle(self):
if self.worker and self.worker.isRunning(): self._stop(); return
if self.random_dataset.isChecked() and not self._find_new_frame():
return
if not self.model.currentData() or not self.frame: QMessageBox.warning(self, "Player needs setup", "Choose a valid Oasis model and reset a starting image first."); return
try:
oasis_root = self._oasis_root()
except ValueError as exc:
QMessageBox.warning(self, "Oasis connection needed", str(exc))
return
try:
info = model_info(self.model.currentData())
self._prepare_temporal_inputs(oasis_root, info)
except (ImportError, OSError, RuntimeError, ValueError) as exc:
QMessageBox.warning(self, "Temporal Oasis setup", str(exc)); return
self.worker = OasisPlaybackWorker(self.root, oasis_root, Path(self.model.currentData()), self.frame, self._settings(), self.state, self.lock, self.temporal_timeline)
self.worker.status.connect(self.status.setText); self.worker.frames_ready.connect(self._frames); self.worker.comparison_ready.connect(self._comparison); self.worker.failed.connect(self._failed); self.worker.finished.connect(lambda: self.start.setText("Start Action Player")); self.worker.start(); self.start.setText("Stop Action Player"); self.preview.setFocus()
def _stop(self):
if self.worker: self.worker.stop(); self.status.setText("Action player paused.")
self._close_temporal_inputs()
self.start.setText("Start Action Player")
def shutdown(self):
"""Stop inference before Qt tears down the page and its CUDA worker."""
if self.worker and self.worker.isRunning():
self.worker.stop()
self.worker.wait(3000)
self._close_temporal_inputs()
def _frames(self, frames, fps, text, mode):
self.sequence += 1; token = self.sequence; self.input.setText(text); self.metrics.setText(f"AI playback FPS: {fps:.1f} | {mode}")
for index, frame in enumerate(frames): QTimer.singleShot(round(index * 1000 / max(1, fps * len(frames))), lambda data=frame, current=token: self._show(data) if current == self.sequence else None)
def _show(self, data):
self.frame = data; pixmap = QPixmap(); pixmap.loadFromData(data, "PNG"); self.preview.setPixmap(pixmap.scaled(self.preview.size(), Qt.KeepAspectRatio, Qt.SmoothTransformation))
def _capture(self):
if not self.frame: return
path = capture_path(self.root)
path.write_bytes(self.frame); self.status.setText(f"Captured {path.name}")
def _compare(self):
try:
if model_info(self.model.currentData()).get("model_type") in {"action_conditioned_temporal_latent_flow", "action_conditioned_temporal_pixel_flow"}:
QMessageBox.information(self, "Temporal model", "Temporal models are checked through their history and generated-rollout metrics. The single-image comparison does not apply.")
return
except ValueError:
pass
if not self.worker or not self.worker.isRunning(): QMessageBox.information(self, "Start player first", "Start the player to load the model, then compare actions."); return
self.status.setText("Generating same-noise action comparison…"); self.worker.request_comparison()
def _comparison(self, data, path):
self._show(data); self.status.setText(f"Saved same-noise action comparison: {Path(path).name}")
def _failed(self, error):
self._close_temporal_inputs(); self.start.setText("Start Action Player"); self.status.setText("Action player failed."); QMessageBox.critical(self, "Oasis player error", error[-2500:])
def _prepare_temporal_inputs(self, oasis_root: Path, info: dict) -> None:
self._close_temporal_inputs()
if info.get("model_type") not in {"action_conditioned_temporal_latent_flow", "action_conditioned_temporal_pixel_flow"}:
return
_external_player_module(oasis_root)
from oasis_input import InputTimeline
from oasis_raw_mouse import RawMouseInput
source = str(info.get("mouse_input_source", ""))
if source not in {"windows_raw_input", "absolute_cursor_fallback"}:
raise ValueError("This temporal model has no recognized mouse input source.")
with self.lock:
buttons = {("jump" if key == "space" else key): 1 for key in self.state["keys"]}
buttons["mouse_right"] = int(self.state["right_mouse"])
self.temporal_timeline = InputTimeline(); self.temporal_timeline.reset(time.perf_counter(), buttons)
self.temporal_mouse_source = source
if source == "windows_raw_input":
self.raw_mouse_input = RawMouseInput(self._raw_mouse)
if not self.raw_mouse_input.available:
error = self.raw_mouse_input.error or "Raw mouse input could not start."
self._close_temporal_inputs()
raise RuntimeError(error)
def _close_temporal_inputs(self) -> None:
if self.raw_mouse_input is not None:
self.raw_mouse_input.close()
self.raw_mouse_input = None; self.temporal_timeline = None; self.temporal_mouse_source = None
def _raw_mouse(self, dx, dy, source_time=None) -> None:
timeline = self.temporal_timeline
if timeline is not None and self.temporal_mouse_source == "windows_raw_input":
timeline.push("mouse", dx=float(dx), dy=float(dy), source_time=source_time)
def eventFilter(self, watched, event):
if watched is self.preview:
with self.lock:
if event.type() == QEvent.KeyPress:
key = {Qt.Key_W: "w", Qt.Key_A: "a", Qt.Key_S: "s", Qt.Key_D: "d", Qt.Key_Space: "space", Qt.Key_Up: "arrow_up", Qt.Key_Left: "arrow_left", Qt.Key_Down: "arrow_down", Qt.Key_Right: "arrow_right"}.get(event.key())
if key:
self.state["keys"].add(key)
if self.temporal_timeline is not None: self.temporal_timeline.push("button", name="jump" if key == "space" else key, value=1)
return True
if event.type() == QEvent.KeyRelease:
key = {Qt.Key_W: "w", Qt.Key_A: "a", Qt.Key_S: "s", Qt.Key_D: "d", Qt.Key_Space: "space", Qt.Key_Up: "arrow_up", Qt.Key_Left: "arrow_left", Qt.Key_Down: "arrow_down", Qt.Key_Right: "arrow_right"}.get(event.key())
if key:
self.state["keys"].discard(key)
if self.temporal_timeline is not None: self.temporal_timeline.push("button", name="jump" if key == "space" else key, value=0)
return True
if event.type() == QEvent.MouseButtonPress and event.button() == Qt.RightButton:
self.state["right_mouse"] = True; self.state["position"] = event.position()
if self.temporal_timeline is not None: self.temporal_timeline.push("button", name="mouse_right", value=1)
return True
if event.type() == QEvent.MouseButtonRelease and event.button() == Qt.RightButton:
self.state["right_mouse"] = False
if self.temporal_timeline is not None: self.temporal_timeline.push("button", name="mouse_right", value=0)
return True
if event.type() == QEvent.MouseMove and self.state.get("right_mouse"):
old = self.state.get("position", event.position()); dx, dy = event.position().x() - old.x(), event.position().y() - old.y()
self.state["mouse_dx"] += dx; self.state["mouse_dy"] += dy; self.state["position"] = event.position()
if self.temporal_timeline is not None and self.temporal_mouse_source == "absolute_cursor_fallback": self.temporal_timeline.push("mouse", dx=dx, dy=dy)
return True
if event.type() == QEvent.Wheel:
scroll = event.angleDelta().y() / 120; self.state["zoom"] += scroll
if self.temporal_timeline is not None: self.temporal_timeline.push("scroll", value=scroll)
return True
return super().eventFilter(watched, event)
def closeEvent(self, event):
self.shutdown(); super().closeEvent(event)
|