Create app.py

app.py

@@ -0,0 +1,1013 @@
+import os
+import subprocess
+import sys
+
+# Disable torch.compile / dynamo before any torch import
+os.environ["TORCH_COMPILE_DISABLE"] = "1"
+os.environ["TORCHDYNAMO_DISABLE"] = "1"
+
+# Clone LTX-2 repo and install packages
+LTX_REPO_URL = "https://github.com/Lightricks/LTX-2.git"
+LTX_REPO_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "LTX-2")
+
+LTX_COMMIT = "ae855f8538843825f9015a419cf4ba5edaf5eec2"  # known working commit with decode_video
+
+if not os.path.exists(LTX_REPO_DIR):
+    print(f"Cloning {LTX_REPO_URL}...")
+    subprocess.run(["git", "clone", LTX_REPO_URL, LTX_REPO_DIR], check=True)
+    subprocess.run(["git", "checkout", LTX_COMMIT], cwd=LTX_REPO_DIR, check=True)
+
+print("Installing ltx-core and ltx-pipelines from cloned repo...")
+subprocess.run(
+    [sys.executable, "-m", "pip", "install", "--force-reinstall", "--no-deps", "-e",
+     os.path.join(LTX_REPO_DIR, "packages", "ltx-core"),
+     "-e", os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines")],
+    check=True,
+)
+
+sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines", "src"))
+sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-core", "src"))
+
+import logging
+import random
+import tempfile
+from pathlib import Path
+import gc
+import hashlib
+
+import torch
+torch._dynamo.config.suppress_errors = True
+torch._dynamo.config.disable = True
+
+import spaces
+import gradio as gr
+import numpy as np
+from huggingface_hub import hf_hub_download, snapshot_download
+from safetensors.torch import load_file, save_file
+from safetensors import safe_open
+import json
+import requests
+
+from ltx_core.components.diffusion_steps import EulerDiffusionStep
+from ltx_core.components.noisers import GaussianNoiser
+from ltx_core.components.protocols import DiffusionStepProtocol
+from ltx_core.model.audio_vae import decode_audio as vae_decode_audio
+from ltx_core.model.audio_vae import encode_audio as vae_encode_audio
+from ltx_core.model.upsampler import upsample_video
+from ltx_core.model.video_vae import TilingConfig, get_video_chunks_number, decode_video as vae_decode_video
+from ltx_core.quantization import QuantizationPolicy
+from ltx_core.types import Audio, LatentState, AudioLatentShape, VideoPixelShape
+from ltx_pipelines.distilled import DistilledPipeline
+from ltx_pipelines.utils import ModelLedger, euler_denoising_loop
+from ltx_pipelines.utils.args import ImageConditioningInput
+from ltx_pipelines.utils.constants import DISTILLED_SIGMA_VALUES, STAGE_2_DISTILLED_SIGMA_VALUES
+from ltx_pipelines.utils.helpers import (
+    cleanup_memory,
+    combined_image_conditionings,
+    denoise_video_only,
+    denoise_audio_video,
+    get_device,
+    encode_prompts,
+    simple_denoising_func,
+)
+from ltx_pipelines.utils.media_io import decode_audio_from_file, encode_video
+from ltx_core.loader.primitives import (
+    LoraPathStrengthAndSDOps,
+    LoraStateDictWithStrength,
+    StateDict,
+)
+from ltx_core.loader.sd_ops import LTXV_LORA_COMFY_RENAMING_MAP
+from ltx_core.loader.fuse_loras import apply_loras
+from ltx_core.loader.sft_loader import SafetensorsStateDictLoader
+
+from ltx_pipelines.utils.types import PipelineComponents
+
+logging.getLogger().setLevel(logging.INFO)
+
+MAX_SEED = np.iinfo(np.int32).max
+DEFAULT_PROMPT = (
+    "An astronaut hatches from a fragile egg on the surface of the Moon, "
+    "the shell cracking and peeling apart in gentle low-gravity motion. "
+    "Fine lunar dust lifts and drifts outward with each movement, floating "
+    "in slow arcs before settling back onto the ground."
+)
+DEFAULT_FRAME_RATE = 24.0
+
+# Resolution presets: (width, height)
+RESOLUTIONS = {
+    "high": {"16:9": (1536, 1024), "9:16": (1024, 1536), "1:1": (1024, 1024)},
+    "low": {"16:9": (768, 512), "9:16": (512, 768), "1:1": (768, 768)},
+}
+
+
+class LTX23DistilledA2VPipeline:
+    """DistilledPipeline with optional audio conditioning."""
+
+    def __init__(
+        self,
+        distilled_checkpoint_path: str,
+        gemma_root: str,
+        spatial_upsampler_path: str,
+        loras: tuple,
+        quantization: QuantizationPolicy | None = None,
+    ):
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self.dtype = torch.bfloat16
+
+        self.model_ledger = ModelLedger(
+            dtype=self.dtype,
+            device=self.device,
+            checkpoint_path=distilled_checkpoint_path,
+            spatial_upsampler_path=spatial_upsampler_path,
+            gemma_root_path=gemma_root,
+            loras=loras,
+            quantization=quantization,
+        )
+
+        self.pipeline_components = PipelineComponents(
+            dtype=self.dtype,
+            device=self.device,
+        )
+    
+    def __call__(
+        self,
+        prompt: str,
+        seed: int,
+        height: int,
+        width: int,
+        num_frames: int,
+        frame_rate: float,
+        images: list[ImageConditioningInput],
+        tiling_config: TilingConfig | None = None,
+        enhance_prompt: bool = False,
+    ):
+
+        generator = torch.Generator(device=self.device).manual_seed(seed)
+        noiser = GaussianNoiser(generator=generator)
+        stepper = EulerDiffusionStep()
+        dtype = torch.bfloat16
+
+        (ctx_p,) = encode_prompts(
+            [prompt],
+            self.model_ledger,
+            enhance_first_prompt=enhance_prompt,
+            enhance_prompt_image=images[0][0] if len(images) > 0 else None,
+        )
+        video_context, audio_context = ctx_p.video_encoding, ctx_p.audio_encoding
+
+        # Stage 1: Initial low resolution video generation.
+        video_encoder = self.model_ledger.video_encoder()
+        transformer = self.model_ledger.transformer()
+        stage_1_sigmas = torch.Tensor(DISTILLED_SIGMA_VALUES).to(self.device)
+
+        def denoising_loop(
+            sigmas: torch.Tensor, video_state: LatentState, audio_state: LatentState, stepper: DiffusionStepProtocol
+        ) -> tuple[LatentState, LatentState]:
+            return euler_denoising_loop(
+                sigmas=sigmas,
+                video_state=video_state,
+                audio_state=audio_state,
+                stepper=stepper,
+                denoise_fn=simple_denoising_func(
+                    video_context=video_context,
+                    audio_context=audio_context,
+                    transformer=transformer,  # noqa: F821
+                ),
+            )
+
+        stage_1_output_shape = VideoPixelShape(
+            batch=1,
+            frames=num_frames,
+            width=width,
+            height=height,
+            fps=frame_rate,
+        )
+        stage_1_conditionings = combined_image_conditionings(
+            images=images,
+            height=stage_1_output_shape.height,
+            width=stage_1_output_shape.width,
+            video_encoder=video_encoder,
+            dtype=dtype,
+            device=self.device,
+        )
+
+        video_state, audio_state = denoise_audio_video(
+            output_shape=stage_1_output_shape,
+            conditionings=stage_1_conditionings,
+            noiser=noiser,
+            sigmas=stage_1_sigmas,
+            stepper=stepper,
+            denoising_loop_fn=denoising_loop,
+            components=self.pipeline_components,
+            dtype=dtype,
+            device=self.device,
+        )
+
+        torch.cuda.synchronize()
+        del transformer
+        del video_encoder
+        cleanup_memory()
+
+        decoded_video = vae_decode_video(
+            video_state.latent, self.model_ledger.video_decoder(), tiling_config, generator
+        )
+        decoded_audio = vae_decode_audio(
+            audio_state.latent, self.model_ledger.audio_decoder(), self.model_ledger.vocoder()
+        )
+        return decoded_video, decoded_audio
+
+
+# Model repos
+LTX_MODEL_REPO = "Lightricks/LTX-2.3"
+GEMMA_REPO ="Lightricks/gemma-3-12b-it-qat-q4_0-unquantized"
+
+# Download model checkpoints
+print("=" * 80)
+print("Downloading LTX-2.3 distilled model + Gemma...")
+print("=" * 80)
+
+# LoRA cache directory and currently-applied key
+LORA_CACHE_DIR = Path("lora_cache")
+LORA_CACHE_DIR.mkdir(exist_ok=True)
+current_lora_key: str | None = None
+
+PENDING_LORA_KEY: str | None = None
+PENDING_LORA_STATE: dict[str, torch.Tensor] | None = None
+PENDING_LORA_STATUS: str = "No LoRA state prepared yet."
+
+# New: cached base transformer SD and in-memory LoRA SDs
+_base_transformer_sd: StateDict | None = None
+_lora_state_dicts: dict[str, StateDict] = {}
+_lora_loader = SafetensorsStateDictLoader()
+
+weights_dir = Path("weights")
+weights_dir.mkdir(exist_ok=True)
+checkpoint_path = hf_hub_download(
+    repo_id="TenStrip/LTX2.3-10Eros",
+    filename="10Eros_v1.2_bf16.safetensors",
+    local_dir=str(weights_dir),
+    local_dir_use_symlinks=False,
+)
+spatial_upsampler_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-spatial-upscaler-x2-1.1.safetensors")
+gemma_root = snapshot_download(repo_id=GEMMA_REPO)
+
+
+# ---- Insert block (LoRA downloads) between lines 268 and 269 ----
+# LoRA repo + download the requested LoRA adapters
+LORA_REPO = "dagloop5/LoRA"
+
+print("=" * 80)
+print("Downloading LoRA adapters from dagloop5/LoRA...")
+print("=" * 80)
+singularity_lora_path = hf_hub_download(repo_id="TenStrip/LTX2.3_JoyAI_Lora_Extracted", filename="JoyAI-Echo_r256.safetensors")
+teneros_lora_path = hf_hub_download(repo_id="TenStrip/LTX2.3_Distilled_Lora_1.1_Experiments", filename="ltx-2.3-22b-distilled-lora-1.1_fro90_ceil52_condsafe.safetensors")
+pose_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2_3_NSFW_furry_concat_v2.safetensors")
+general_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2.3_reasoning_I2V_V3.safetensors")
+motion_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Sulphur_LTX 2.3_better _NSFW_motion.safetensors")
+dreamlay_lora_path = hf_hub_download(repo_id="lynaNSFW/DR34ML4Y_AIO_NSFW_LTX23", filename="DR34ML4Y_LTXXX_V2.safetensors") # m15510n4ry, bl0wj0b, d0ubl3_bj, d0gg1e, c0wg1rl
+mself_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Furry Hyper Masturbation - LTX-2 I2V v1.safetensors") # Hyperfap
+dramatic_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX-2.3 - Orgasm.safetensors") # buddr
+fluid_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Cr3ampi3_animation_sulphur-2_i2v_v1.0.safetensors") # cr3ampi3 animation
+liquid_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="liquid_wet_dr1pp_ltx2_v1.0_scaled.safetensors") # wet dr1pp
+demopose_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="clapping-cheeks-audio-v001-alpha.safetensors")
+voice_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="hentai_voice_ltx23_v2.comfy.safetensors")
+realism_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="FurryenhancerLTX2.3V4.094fused.safetensors")
+transition_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX-2_takerpov_lora_v1.2.safetensors") # takerpov1, taker pov
+physics_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2.3_Physics_V2_000002000.safetensors")
+reasoning_lora_path = hf_hub_download(repo_id="LiconStudio/Ltx2.3-VBVR-lora-I2V", filename="Ltx2.3-Licon-VBVR-I2V-390K-R32.safetensors")
+twostep_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2.3_Multi_step_video_reasoning_V0.1.safetensors")
+mcfurry_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX-2_3_mc_Furry_realistic_lora_v1.safetensors")
+dm_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Doggy_mission_sulphur-2_v0.5.safetensors")
+praxis_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Penile_Praxis_V4.safetensors")
+threed_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="ltx2-3d-animations-12500-steps-k3nk.safetensors")
+concept_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="ltx23_nsfw_helper_multi_concept_lora_v2.safetensors")
+bulge_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="stomach_bulge_10eros_sulphur_v1.safetensors")
+
+print(f"Singularity LoRA: {singularity_lora_path}")
+print(f"10Eros LoRA: {teneros_lora_path}")
+print(f"Pose LoRA: {pose_lora_path}")
+print(f"General LoRA: {general_lora_path}")
+print(f"Motion LoRA: {motion_lora_path}")
+print(f"Dreamlay LoRA: {dreamlay_lora_path}")
+print(f"Mself LoRA: {mself_lora_path}")
+print(f"Dramatic LoRA: {dramatic_lora_path}")
+print(f"Fluid LoRA: {fluid_lora_path}")
+print(f"Liquid LoRA: {liquid_lora_path}")
+print(f"Demopose LoRA: {demopose_lora_path}")
+print(f"Voice LoRA: {voice_lora_path}")
+print(f"Realism LoRA: {realism_lora_path}")
+print(f"Transition LoRA: {transition_lora_path}")
+print(f"Physics LoRA: {physics_lora_path}")
+print(f"Reasoning LoRA: {reasoning_lora_path}")
+print(f"Twostep LoRA: {twostep_lora_path}")
+print(f"Mcfurry LoRA: {mcfurry_lora_path}")
+print(f"DM LoRA: {dm_lora_path}")
+print(f"Praxis LoRA: {praxis_lora_path}")
+print(f"ThreeD LoRA: {threed_lora_path}")
+print(f"Concept LoRA: {concept_lora_path}")
+print(f"Bulge LoRA: {bulge_lora_path}")
+
+print(f"Checkpoint: {checkpoint_path}")
+print(f"Spatial upsampler: {spatial_upsampler_path}")
+print(f"[Gemma] Root ready: {gemma_root}")
+
+# Initialize pipeline WITH text encoder and optional audio support
+# ---- Replace block (pipeline init) lines 275-281 ----
+pipeline = LTX23DistilledA2VPipeline(
+    distilled_checkpoint_path=checkpoint_path,
+    spatial_upsampler_path=spatial_upsampler_path,
+    gemma_root=gemma_root,
+    loras=[],
+    quantization=None,
+)
+# ----------------------------------------------------------------
+
+def _make_lora_key(singularity_strength: float, teneros_strength: float, pose_strength: float, general_strength: float, motion_strength: float, dreamlay_strength: float, mself_strength: float, dramatic_strength: float, fluid_strength: float, liquid_strength: float, demopose_strength: float, voice_strength: float, realism_strength: float, transition_strength: float, physics_strength: float, reasoning_strength: float, twostep_strength: float, mcfurry_strength: float, dm_strength: float, praxis_strength: float, threed_strength: float, concept_strength: float, bulge_strength: float) -> tuple[str, str]:
+    rx = round(float(singularity_strength), 2)
+    ra = round(float(teneros_strength), 2)
+    rp = round(float(pose_strength), 2)
+    rg = round(float(general_strength), 2)
+    rm = round(float(motion_strength), 2)
+    rd = round(float(dreamlay_strength), 2)
+    rs = round(float(mself_strength), 2)
+    rr = round(float(dramatic_strength), 2)
+    rf = round(float(fluid_strength), 2)
+    rl = round(float(liquid_strength), 2)
+    ro = round(float(demopose_strength), 2)
+    rv = round(float(voice_strength), 2)
+    re = round(float(realism_strength), 2)
+    rt = round(float(transition_strength), 2)
+    ry = round(float(physics_strength), 2)
+    ri = round(float(reasoning_strength), 2)
+    rw = round(float(twostep_strength), 2)
+    mc = round(float(mcfurry_strength), 2)
+    dm = round(float(dm_strength), 2)
+    pr = round(float(praxis_strength), 2)
+    td = round(float(threed_strength), 2)
+    co = round(float(concept_strength), 2)
+    bu = round(float(bulge_strength), 2)
+    key_str = f"{singularity_lora_path}:{rx}|{teneros_lora_path}:{ra}|{pose_lora_path}:{rp}|{general_lora_path}:{rg}|{motion_lora_path}:{rm}|{dreamlay_lora_path}:{rd}|{mself_lora_path}:{rs}|{dramatic_lora_path}:{rr}|{fluid_lora_path}:{rf}|{liquid_lora_path}:{rl}|{demopose_lora_path}:{ro}|{voice_lora_path}:{rv}|{realism_lora_path}:{re}|{transition_lora_path}:{rt}|{physics_lora_path}:{ry}|{reasoning_lora_path}:{ri}|{twostep_lora_path}:{rw}|{mcfurry_lora_path}:{mc}|{dm_lora_path}:{dm}|{praxis_lora_path}:{pr}|{threed_lora_path}:{td}|{concept_lora_path}:{co}|{bulge_lora_path}:{bu}"
+    key = hashlib.sha256(key_str.encode("utf-8")).hexdigest()
+    return key, key_str
+
+
+def prepare_lora_cache(
+    singularity_strength: float,
+    teneros_strength: float,
+    pose_strength: float,
+    general_strength: float,
+    motion_strength: float,
+    dreamlay_strength: float,
+    mself_strength: float,
+    dramatic_strength: float,
+    fluid_strength: float,
+    liquid_strength: float,
+    demopose_strength: float,
+    voice_strength: float,
+    realism_strength: float,
+    transition_strength: float,
+    physics_strength: float,
+    reasoning_strength: float,
+    twostep_strength: float,
+    mcfurry_strength: float,
+    dm_strength: float,
+    praxis_strength: float,
+    threed_strength: float,
+    concept_strength: float,
+    bulge_strength: float,
+    progress=gr.Progress(track_tqdm=True),
+):
+    """
+    Fast CPU step:
+    - checks on-disk cache
+    - if missing, fuses the selected LoRA strengths against the pre-cached base SD
+    - saves the fused SD so the next apply step is just load_state_dict
+    """
+    global PENDING_LORA_KEY, PENDING_LORA_STATE, PENDING_LORA_STATUS
+
+    key, _ = _make_lora_key(
+        singularity_strength, teneros_strength, pose_strength, general_strength,
+        motion_strength, dreamlay_strength, mself_strength, dramatic_strength,
+        fluid_strength, liquid_strength, demopose_strength, voice_strength,
+        realism_strength, transition_strength, physics_strength, reasoning_strength,
+        twostep_strength, mcfurry_strength, dm_strength, praxis_strength,
+        threed_strength, concept_strength, bulge_strength,
+    )
+    cache_path = LORA_CACHE_DIR / f"{key}.safetensors"
+
+    progress(0.05, desc="Preparing LoRA state")
+
+    if cache_path.exists():
+        try:
+            progress(0.20, desc="Loading cached fused state")
+            state = load_file(str(cache_path))
+            PENDING_LORA_KEY = key
+            PENDING_LORA_STATE = state
+            PENDING_LORA_STATUS = f"Loaded cached LoRA state: {cache_path.name}"
+            return PENDING_LORA_STATUS
+        except Exception as e:
+            print(f"[LoRA] Cache load failed: {type(e).__name__}: {e}")
+
+    strengths = {
+        "singularity": round(float(singularity_strength), 2),
+        "teneros":     round(float(teneros_strength), 2),
+        "pose":        round(float(pose_strength), 2),
+        "general":     round(float(general_strength), 2),
+        "motion":      round(float(motion_strength), 2),
+        "dreamlay":    round(float(dreamlay_strength), 2),
+        "mself":       round(float(mself_strength), 2),
+        "dramatic":    round(float(dramatic_strength), 2),
+        "fluid":       round(float(fluid_strength), 2),
+        "liquid":      round(float(liquid_strength), 2),
+        "demopose":    round(float(demopose_strength), 2),
+        "voice":       round(float(voice_strength), 2),
+        "realism":     round(float(realism_strength), 2),
+        "transition":  round(float(transition_strength), 2),
+        "physics":     round(float(physics_strength), 2),
+        "reasoning":   round(float(reasoning_strength), 2),
+        "twostep":     round(float(twostep_strength), 2),
+        "mcfurry":     round(float(mcfurry_strength), 2),
+        "dm":          round(float(dm_strength), 2),
+        "praxis":      round(float(praxis_strength), 2),
+        "threed":      round(float(threed_strength), 2),
+        "concept":     round(float(concept_strength), 2),
+        "bulge":       round(float(bulge_strength), 2),
+    }
+
+    if not any(v != 0.0 for v in strengths.values()):
+        PENDING_LORA_KEY = None
+        PENDING_LORA_STATE = None
+        PENDING_LORA_STATUS = "No non-zero LoRA strengths selected; nothing to prepare."
+        return PENDING_LORA_STATUS
+
+    try:
+        progress(0.25, desc="Fusing LoRAs into base state dict")
+        fused_state = _fuse_lora_state_dict(strengths)
+
+        progress(0.70, desc="Saving fused state to cache")
+        save_file(fused_state, str(cache_path))
+
+        PENDING_LORA_KEY = key
+        PENDING_LORA_STATE = fused_state
+        PENDING_LORA_STATUS = f"Built and cached LoRA state: {cache_path.name}"
+        return PENDING_LORA_STATUS
+
+    except Exception as e:
+        import traceback
+        print(f"[LoRA] Prepare failed: {type(e).__name__}: {e}")
+        print(traceback.format_exc())
+        PENDING_LORA_KEY = None
+        PENDING_LORA_STATE = None
+        PENDING_LORA_STATUS = f"LoRA prepare failed: {type(e).__name__}: {e}"
+        return PENDING_LORA_STATUS
+
+    finally:
+        gc.collect()
+
+
+def apply_prepared_lora_state_to_pipeline():
+    """
+    Fast step: copy the already prepared CPU state into the live transformer.
+    This is the only part that should remain near generation time.
+    """
+    global current_lora_key, PENDING_LORA_KEY, PENDING_LORA_STATE
+
+    if PENDING_LORA_STATE is None or PENDING_LORA_KEY is None:
+        print("[LoRA] No prepared LoRA state available; skipping.")
+        return False
+
+    if current_lora_key == PENDING_LORA_KEY:
+        print("[LoRA] Prepared LoRA state already active; skipping.")
+        return True
+
+    existing_transformer = _transformer
+    with torch.no_grad():
+        missing, unexpected = existing_transformer.load_state_dict(PENDING_LORA_STATE, strict=False)
+        if missing or unexpected:
+            print(f"[LoRA] load_state_dict mismatch: missing={len(missing)}, unexpected={len(unexpected)}")
+
+    current_lora_key = PENDING_LORA_KEY
+    print("[LoRA] Prepared LoRA state applied to the pipeline.")
+    return True
+
+def _capture_base_transformer_state() -> None:
+    """
+    After _transformer has been preloaded with no LoRAs, snapshot its complete
+    state_dict to CPU. This is the reference "W_base" for all future LoRA fusing.
+    """
+    global _base_transformer_sd
+    assert _transformer is not None, "transformer must be preloaded first"
+
+    sd_dict = {
+        k: v.detach().to(device=torch.device("cpu"), dtype=torch.bfloat16).contiguous()
+        for k, v in _transformer.state_dict().items()
+    }
+    _base_transformer_sd = StateDict(
+        sd=sd_dict,
+        device=torch.device("cpu"),
+        size=sum(t.nbytes for t in sd_dict.values()),
+        dtype={torch.bfloat16},
+    )
+    print(f"[LoRA] Base transformer SD captured: {_base_transformer_sd.size / 1024**3:.2f} GB on CPU")
+
+
+def _preload_lora_state_dicts() -> None:
+    """
+    Load every downloaded LoRA file into system RAM once.
+    The files are renamed with the LTXV Comfy map so they match base-model keys.
+    """
+    print("=" * 80)
+    print("Pre-loading LoRA state dicts into CPU memory...")
+    print("=" * 80)
+
+    entries = [
+        ("singularity", singularity_lora_path),
+        ("teneros",     teneros_lora_path),
+        ("pose",        pose_lora_path),
+        ("general",     general_lora_path),
+        ("motion",      motion_lora_path),
+        ("dreamlay",    dreamlay_lora_path),
+        ("mself",       mself_lora_path),
+        ("dramatic",    dramatic_lora_path),
+        ("fluid",       fluid_lora_path),
+        ("liquid",      liquid_lora_path),
+        ("demopose",    demopose_lora_path),
+        ("voice",       voice_lora_path),
+        ("realism",     realism_lora_path),
+        ("transition",  transition_lora_path),
+        ("physics",     physics_lora_path),
+        ("reasoning",   reasoning_lora_path),
+        ("twostep",     twostep_lora_path),
+        ("mcfurry",     mcfurry_lora_path),
+        ("dm",          dm_lora_path),
+        ("praxis",      praxis_lora_path),
+        ("threed",      threed_lora_path),
+        ("concept",     concept_lora_path),
+        ("bulge",       bulge_lora_path),
+    ]
+    for label, path in entries:
+        _lora_state_dicts[label] = _lora_loader.load(
+            [path],
+            sd_ops=LTXV_LORA_COMFY_RENAMING_MAP,
+            device=torch.device("cpu"),
+        )
+        print(f"[LoRA] Pre-loaded {label}: {path}")
+
+
+def _fuse_lora_state_dict(strengths: dict[str, float]) -> dict[str, torch.Tensor]:
+    """
+    Compute W_eff = W_base + sum_i(alpha_i * B_i @ A_i) for all active LoRAs.
+    This calls the same primitive SingleGPUModelBuilder uses internally,
+    but without rebuilding the meta-model or reloading the base checkpoint.
+    """
+    assert _base_transformer_sd is not None, "Base transformer state dict not ready"
+
+    active = []
+    for label, strength in strengths.items():
+        if float(strength) == 0.0:
+            continue
+        active.append(
+            LoraStateDictWithStrength(
+                sd=_lora_state_dicts[label],
+                strength=float(strength),
+            )
+        )
+
+    if not active:
+        return {
+            k: v.detach().clone().contiguous()
+            for k, v in _base_transformer_sd.sd.items()
+        }
+
+    fused = apply_loras(
+        model_sd=_base_transformer_sd,
+        lora_sd_and_strengths=active,
+        dtype=torch.bfloat16,
+        destination_sd=None,
+    )
+
+    return {
+        k: v.detach().to(device=torch.device("cpu")).contiguous()
+        for k, v in fused.sd.items()
+    }
+
+# ---- REPLACE PRELOAD BLOCK START ----
+# Preload all models for ZeroGPU tensor packing.
+print("Preloading all models (including Gemma and audio components)...")
+ledger = pipeline.model_ledger
+
+# Save the original factory methods so we can rebuild individual components later.
+# These are bound callables on ledger that will call the builder when invoked.
+_orig_transformer_factory = ledger.transformer
+_orig_video_encoder_factory = ledger.video_encoder
+_orig_video_decoder_factory = ledger.video_decoder
+_orig_audio_encoder_factory = ledger.audio_encoder
+_orig_audio_decoder_factory = ledger.audio_decoder
+_orig_vocoder_factory = ledger.vocoder
+_orig_spatial_upsampler_factory = ledger.spatial_upsampler
+_orig_text_encoder_factory = ledger.text_encoder
+_orig_gemma_embeddings_factory = ledger.gemma_embeddings_processor
+
+# Call the original factories once to create the cached instances we will serve by default.
+_transformer = _orig_transformer_factory()
+_video_encoder = _orig_video_encoder_factory()
+_video_decoder = _orig_video_decoder_factory()
+_audio_encoder = _orig_audio_encoder_factory()
+_audio_decoder = _orig_audio_decoder_factory()
+_vocoder = _orig_vocoder_factory()
+_spatial_upsampler = _orig_spatial_upsampler_factory()
+_text_encoder = _orig_text_encoder_factory()
+_embeddings_processor = _orig_gemma_embeddings_factory()
+
+# Replace ledger methods with lightweight lambdas that return the cached instances.
+# We keep the original factories above so we can call them later to rebuild components.
+ledger.transformer = lambda: _transformer
+ledger.video_encoder = lambda: _video_encoder
+ledger.video_decoder = lambda: _video_decoder
+ledger.audio_encoder = lambda: _audio_encoder
+ledger.audio_decoder = lambda: _audio_decoder
+ledger.vocoder = lambda: _vocoder
+ledger.spatial_upsampler = lambda: _spatial_upsampler
+ledger.text_encoder = lambda: _text_encoder
+ledger.gemma_embeddings_processor = lambda: _embeddings_processor
+
+print("All models preloaded (including Gemma text encoder and audio encoder)!")
+# ---- REPLACE PRELOAD BLOCK END ----
+
+_capture_base_transformer_state()
+_preload_lora_state_dicts()
+cleanup_memory()
+
+print("=" * 80)
+print("Pipeline ready!")
+print("=" * 80)
+
+
+def log_memory(tag: str):
+    if torch.cuda.is_available():
+        allocated = torch.cuda.memory_allocated() / 1024**3
+        peak = torch.cuda.max_memory_allocated() / 1024**3
+        free, total = torch.cuda.mem_get_info()
+        print(f"[VRAM {tag}] allocated={allocated:.2f}GB peak={peak:.2f}GB free={free / 1024**3:.2f}GB total={total / 1024**3:.2f}GB")
+
+
+def detect_aspect_ratio(image) -> str:
+    if image is None:
+        return "16:9"
+    if hasattr(image, "size"):
+        w, h = image.size
+    elif hasattr(image, "shape"):
+        h, w = image.shape[:2]
+    else:
+        return "16:9"
+    ratio = w / h
+    candidates = {"16:9": 16 / 9, "9:16": 9 / 16, "1:1": 1.0}
+    return min(candidates, key=lambda k: abs(ratio - candidates[k]))
+
+
+def on_image_upload(first_image, last_image, high_res):
+    ref_image = first_image if first_image is not None else last_image
+    aspect = detect_aspect_ratio(ref_image)
+    tier = "high" if high_res else "low"
+    w, h = RESOLUTIONS[tier][aspect]
+    return gr.update(value=w), gr.update(value=h)
+
+
+def on_highres_toggle(first_image, last_image, high_res):
+    ref_image = first_image if first_image is not None else last_image
+    aspect = detect_aspect_ratio(ref_image)
+    tier = "high" if high_res else "low"
+    w, h = RESOLUTIONS[tier][aspect]
+    return gr.update(value=w), gr.update(value=h)
+
+
+def get_gpu_duration(
+    first_image,
+    last_image,
+    prompt: str,
+    duration: float,
+    gpu_duration: float,
+    enhance_prompt: bool = True,
+    seed: int = 42,
+    randomize_seed: bool = True,
+    height: int = 1024,
+    width: int = 1536,
+    singularity_strength: float = 0.0,
+    teneros_strength: float = 0.0,
+    pose_strength: float = 0.0,
+    general_strength: float = 0.0,
+    motion_strength: float = 0.0,
+    dreamlay_strength: float = 0.0,
+    mself_strength: float = 0.0,
+    dramatic_strength: float = 0.0,
+    fluid_strength: float = 0.0,
+    liquid_strength: float = 0.0,
+    demopose_strength: float = 0.0,
+    voice_strength: float = 0.0,
+    realism_strength: float = 0.0,
+    transition_strength: float = 0.0,
+    physics_strength: float = 0.0,
+    reasoning_strength: float = 0.0,
+    twostep_strength: float = 0.0,
+    mcfurry_strength: float = 0.0,
+    dm_strength: float = 0.0,
+    praxis_strength: float = 0.0,
+    threed_strength: float = 0.0,
+    concept_strength: float = 0.0,
+    bulge_strength: float = 0.0,
+    progress=None,
+):
+    return int(gpu_duration)
+
+@spaces.GPU(duration=get_gpu_duration)
+@torch.inference_mode()
+def generate_video(
+    first_image,
+    last_image,
+    prompt: str,
+    duration: float,
+    gpu_duration: float,
+    enhance_prompt: bool = True,
+    seed: int = 42,
+    randomize_seed: bool = True,
+    height: int = 1024,
+    width: int = 1536,
+    singularity_strength: float = 0.0,
+    teneros_strength: float = 0.0,
+    pose_strength: float = 0.0,
+    general_strength: float = 0.0,
+    motion_strength: float = 0.0,
+    dreamlay_strength: float = 0.0,
+    mself_strength: float = 0.0,
+    dramatic_strength: float = 0.0,
+    fluid_strength: float = 0.0,
+    liquid_strength: float = 0.0,
+    demopose_strength: float = 0.0,
+    voice_strength: float = 0.0,
+    realism_strength: float = 0.0,
+    transition_strength: float = 0.0,
+    physics_strength: float = 0.0,
+    reasoning_strength: float = 0.0,
+    twostep_strength: float = 0.0,
+    mcfurry_strength: float = 0.0,
+    dm_strength: float = 0.0,
+    praxis_strength: float = 0.0,
+    threed_strength: float = 0.0,
+    concept_strength: float = 0.0,
+    bulge_strength: float = 0.0,
+    progress=gr.Progress(track_tqdm=True),
+):
+    try:
+        torch.cuda.reset_peak_memory_stats()
+        log_memory("start")
+
+        current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
+
+        frame_rate = DEFAULT_FRAME_RATE
+        num_frames = int(duration * frame_rate) + 1
+        num_frames = ((num_frames - 1 + 7) // 8) * 8 + 1
+
+        print(f"Generating: {height}x{width}, {num_frames} frames ({duration}s), seed={current_seed}")
+
+        images = []
+        output_dir = Path("outputs")
+        output_dir.mkdir(exist_ok=True)
+
+        if first_image is not None:
+            temp_first_path = output_dir / f"temp_first_{current_seed}.jpg"
+            if hasattr(first_image, "save"):
+                first_image.save(temp_first_path)
+            else:
+                temp_first_path = Path(first_image)
+            images.append(ImageConditioningInput(path=str(temp_first_path), frame_idx=0, strength=1.0))
+
+        if last_image is not None:
+            temp_last_path = output_dir / f"temp_last_{current_seed}.jpg"
+            if hasattr(last_image, "save"):
+                last_image.save(temp_last_path)
+            else:
+                temp_last_path = Path(last_image)
+            images.append(ImageConditioningInput(path=str(temp_last_path), frame_idx=num_frames - 1, strength=1.0))
+
+        tiling_config = TilingConfig.default()
+        video_chunks_number = get_video_chunks_number(num_frames, tiling_config)
+
+        log_memory("before pipeline call")
+
+        apply_prepared_lora_state_to_pipeline()
+
+        video, audio = pipeline(
+            prompt=prompt,
+            seed=current_seed,
+            height=int(height),
+            width=int(width),
+            num_frames=num_frames,
+            frame_rate=frame_rate,
+            images=images,
+            tiling_config=tiling_config,
+            enhance_prompt=enhance_prompt,
+        )
+
+        log_memory("after pipeline call")
+
+        output_path = tempfile.mktemp(suffix=".mp4")
+        encode_video(
+            video=video,
+            fps=frame_rate,
+            audio=audio,
+            output_path=output_path,
+            video_chunks_number=video_chunks_number,
+        )
+
+        log_memory("after encode_video")
+        return str(output_path), current_seed
+
+    except Exception as e:
+        import traceback
+        log_memory("on error")
+        print(f"Error: {str(e)}\n{traceback.format_exc()}")
+        return None, current_seed
+
+
+with gr.Blocks(title="LTX-2.3 Distilled") as demo:
+    gr.Markdown("# LTX-2.3 F2LF with Fast Audio-Video Generation with Frame Conditioning")
+    
+
+    with gr.Row():
+        with gr.Column():
+            with gr.Row():
+                first_image = gr.Image(label="First Frame (Optional)", type="pil")
+                last_image = gr.Image(label="Last Frame (Optional)", type="pil")
+            prompt = gr.Textbox(
+                label="Prompt",
+                info="for best results - make it as elaborate as possible",
+                value="Make this image come alive with cinematic motion, smooth animation",
+                lines=3,
+                placeholder="Describe the motion and animation you want...",
+            )
+            duration = gr.Slider(label="Duration (seconds)", minimum=1.0, maximum=30.0, value=10.0, step=0.1)
+                
+
+            generate_btn = gr.Button("Generate Video", variant="primary", size="lg")
+
+            with gr.Accordion("Advanced Settings", open=False):
+                seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, value=10, step=1)
+                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
+                with gr.Row():
+                    width = gr.Number(label="Width", value=1536, precision=0)
+                    height = gr.Number(label="Height", value=1024, precision=0)
+                with gr.Row():
+                    enhance_prompt = gr.Checkbox(label="Enhance Prompt", value=False)
+                    high_res = gr.Checkbox(label="High Resolution", value=True)
+                with gr.Column():
+                    gr.Markdown("### LoRA adapter strengths (set to 0 to disable; slow and WIP)")
+                    singularity_strength = gr.Slider(
+                        label="JoyAI Lora strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    teneros_strength = gr.Slider(
+                        label="10Eros Distilled Lora strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    pose_strength = gr.Slider(
+                        label="Anthro Enhancer strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    general_strength = gr.Slider(
+                        label="Reasoning Enhancer strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    motion_strength = gr.Slider(
+                        label="Anthro Posing Helper strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    dreamlay_strength = gr.Slider(
+                        label="Dreamlay strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    mself_strength = gr.Slider(
+                        label="Mself strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    dramatic_strength = gr.Slider(
+                        label="Dramatic strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    fluid_strength = gr.Slider(
+                        label="Fluid Helper strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    liquid_strength = gr.Slider(
+                        label="Liquid Helper strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    demopose_strength = gr.Slider(
+                        label="Audio Helper strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    voice_strength = gr.Slider(
+                        label="Voice Helper strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    realism_strength = gr.Slider(
+                        label="Anthro Realism strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    transition_strength = gr.Slider(
+                        label="POV strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    physics_strength = gr.Slider(
+                        label="Physics strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    reasoning_strength = gr.Slider(
+                        label="Official Reasoning strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    twostep_strength = gr.Slider(
+                        label="Two Step Reasoning strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    mcfurry_strength = gr.Slider(
+                        label="t2v anthro realism strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    dm_strength = gr.Slider(
+                        label="DM3D strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    praxis_strength = gr.Slider(
+                        label="Praxis strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    threed_strength = gr.Slider(
+                        label="3D animation strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    concept_strength = gr.Slider(
+                        label="Conceptual strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    bulge_strength = gr.Slider(
+                        label="Bulge strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                prepare_lora_btn = gr.Button("Prepare / Load LoRA Cache", variant="secondary")
+                lora_status = gr.Textbox(
+                    label="LoRA Cache Status",
+                    value="No LoRA state prepared yet.",
+                    interactive=False,
+                )
+
+        with gr.Column():
+            output_video = gr.Video(label="Generated Video", autoplay=False)
+            gpu_duration = gr.Slider(
+                label="ZeroGPU duration (seconds; 10 second Img2Vid with 1024x1024 and LoRAs = ~70)",
+                minimum=30.0,
+                maximum=240.0,
+                value=75.0,
+                step=1.0,
+            )
+
+    first_image.change(
+        fn=on_image_upload,
+        inputs=[first_image, last_image, high_res],
+        outputs=[width, height],
+    )
+
+    last_image.change(
+        fn=on_image_upload,
+        inputs=[first_image, last_image, high_res],
+        outputs=[width, height],
+    )
+
+    high_res.change(
+        fn=on_highres_toggle,
+        inputs=[first_image, last_image, high_res],
+        outputs=[width, height],
+    )
+
+    prepare_lora_btn.click(
+        fn=prepare_lora_cache,
+        inputs=[singularity_strength, teneros_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength],
+        outputs=[lora_status],
+    )
+    
+    generate_btn.click(
+        fn=generate_video,
+        inputs=[
+            first_image, last_image, prompt, duration, gpu_duration, enhance_prompt,
+            seed, randomize_seed, height, width,
+            singularity_strength, teneros_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength,
+        ],
+        outputs=[output_video, seed],
+    )
+
+
+css = """
+.fillable{max-width: 1200px !important}
+"""
+
+if __name__ == "__main__":
+    demo.launch(theme=gr.themes.Citrus(), css=css)