app.py

import os
import subprocess
import sys

# ZeroGPU: torch.compile / dynamo unsupported — disable before any torch import.
os.environ["TORCH_COMPILE_DISABLE"] = "1"
os.environ["TORCHDYNAMO_DISABLE"] = "1"

# (removed runtime xformers install -> would pull torch 2.8 and break the AOTI .pt2; SDPA used)

# --- clone + install the NATIVE LTX-2 codebase at the pinned commit the working ZeroGPU spaces use ---
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"
if not os.path.exists(LTX_REPO_DIR):
    subprocess.run(["git", "clone", LTX_REPO_URL, LTX_REPO_DIR], check=True)
    subprocess.run(["git", "-C", LTX_REPO_DIR, "checkout", LTX_COMMIT], check=True)
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

import numpy as np
import imageio.v3 as iio
from PIL import Image, ImageOps

import torch
torch._dynamo.config.suppress_errors = True
torch._dynamo.config.disable = True

import spaces
import gradio as gr
from huggingface_hub import hf_hub_download, snapshot_download

# Import LTX modules in the proven order — importing ltx_core.quantization/loader FIRST hits a
# circular import (fp8_cast <-> loader.fuse_loras). Importing the model modules first forces the
# correct init order (mirrors the working reference Space).
from ltx_core.model.video_vae import TilingConfig, get_video_chunks_number, decode_video as _vae_decode_video  # noqa: F401
from ltx_core.model.upsampler import upsample_video as _upsample_video  # noqa: F401
from ltx_core.model.audio_vae import encode_audio as _vae_encode_audio  # noqa: F401
from ltx_core.quantization import QuantizationPolicy
from ltx_core.loader import LoraPathStrengthAndSDOps, LTXV_LORA_COMFY_RENAMING_MAP
from ltx_pipelines.ic_lora import ICLoraPipeline
from ltx_pipelines.utils.media_io import encode_video

# --- ZeroGPU loader patch -------------------------------------------------------------
# The native loader opens safetensors directly on the CUDA device
# (safe_open(path, device="cuda")), doing the host->device copy in safetensors' own C++
# (cudaMemcpy) — bypassing torch.Tensor.to, the call ZeroGPU patches to virtualise + pack
# weights at module scope. Result: "No CUDA GPUs are available" at startup, nothing packs.
# Patch it to open on CPU then move via torch.Tensor.to (ZeroGPU-virtualisable).
import safetensors as _safetensors
import ltx_core.loader.sft_loader as _sft
from ltx_core.loader.primitives import StateDict as _StateDict

def _zerogpu_safe_load(self, path, sd_ops, device=None):
    device = device or torch.device("cpu")
    sd, size, dtype = {}, 0, set()
    model_paths = path if isinstance(path, list) else [path]
    for shard_path in model_paths:
        with _safetensors.safe_open(shard_path, framework="pt", device="cpu") as f:
            for name in f.keys():
                expected = name if sd_ops is None else sd_ops.apply_to_key(name)
                if expected is None:
                    continue
                value = f.get_tensor(name).to(device=device)  # torch path -> ZeroGPU-virtualised
                kvs = ((expected, value),)
                if sd_ops is not None:
                    kvs = sd_ops.apply_to_key_value(expected, value)
                for k, v in kvs:
                    size += v.nbytes
                    dtype.add(v.dtype)
                    sd[k] = v
    return _StateDict(sd=sd, device=device, size=size, dtype=dtype)

_sft.SafetensorsStateDictLoader.load = _zerogpu_safe_load
print("[PATCH] safetensors loader -> CPU-open + torch.to (ZeroGPU-virtualisable)")
# --------------------------------------------------------------------------------------

# --- attention backend patch (FA3 crashes on Blackwell ZeroGPU; use xformers/SDPA) ---
import torch.nn.functional as F
from ltx_core.model.transformer import attention as _attn_mod

def _sdpa_as_mea(query, key, value, attn_bias=None, scale=None, **kwargs):
    q, k, v = query.transpose(1, 2), key.transpose(1, 2), value.transpose(1, 2)
    return F.scaled_dot_product_attention(q, k, v, scale=scale).transpose(1, 2)

# IMPORTANT (ZeroGPU): never query CUDA at module scope. SDPA works on every GPU (incl.
# Blackwell ZeroGPU, where FA3 crashes), so patch it unconditionally.
_attn_mod.memory_efficient_attention = _sdpa_as_mea
print("[ATTN] SDPA (patched at module scope, no CUDA query)")

logging.getLogger().setLevel(logging.INFO)

# =========================== PER-LORA CONFIG (colorize) ===========================
TITLE = "LTX-2.3 Add Water (native LTX-2)"
LORA_REPO = "Lightricks/LTX-2.3-22b-IC-LoRA-Water-Simulation"
LORA_FILE = "ltx-2.3-22b-ic-lora-water-simulation-0.9.safetensors"
LORA_SCALE = 1.0
SKIP_STAGE_2 = True
GRAYSCALE_REF = False
RES_PRESETS = {"960×544 (fast)": (960, 544), "1216×704 (recommended)": (1216, 704)}
DEFAULT_PRESET = "1216×704 (recommended)"
FRAME_CHOICES = [49, 73, 97, 121]
DEFAULT_FRAMES = 73

def build_prompt(p):
    return (
        "Reference shows the dry scene. Edited shows the same scene with realistic, naturally-moving water added. "
        f"ADD WATER {p.strip()}. "
        "Subject identity, framing and motion are identical to the reference; only water is added."
    )

EXAMPLES = [
    ["examples/landscape_dry.mp4",
     "a wide river flooding across the valley with glassy rippling reflections and drifting foam, mist rising; flowing water and a distant waterfall",
     "1216×704 (recommended)", 73, 42, False],
    ["examples/man_dancing_dry.mp4",
     "a clear shallow stream rushing and braiding around their legs with white foam and splashes; rushing water and rhythmic splashes",
     "1216×704 (recommended)", 73, 42, False],
]
# =================================================================================

FPS = 24.0
MAX_SEED = np.iinfo(np.int32).max
HF_TOKEN = os.environ.get("HF_TOKEN")
LTX_MODEL_REPO = "Lightricks/LTX-2.3"
GEMMA_REPO = "google/gemma-3-12b-it-qat-q4_0-unquantized"


def _src_fps(path, default=FPS):
    try:
        return float(iio.immeta(path, plugin="pyav").get("fps", default)) or default
    except Exception:
        return default


def _prep_reference(path, width, height, num_frames):
    """Resample to 24fps, aspect-fit/crop to WxH, NF frames; (optionally grayscale); write temp mp4."""
    vid = iio.imread(path, plugin="pyav")
    src_fps = _src_fps(path)
    n = len(vid)
    out = []
    for i in range(num_frames):
        idx = min(int(round(i / FPS * src_fps)), n - 1)
        im = Image.fromarray(vid[idx]).convert("RGB")
        im = ImageOps.fit(im, (width, height), Image.LANCZOS)
        if GRAYSCALE_REF:
            im = im.convert("L").convert("RGB")
        out.append(np.array(im))
    tmp = tempfile.mktemp(suffix=".mp4")
    iio.imwrite(tmp, np.stack(out), fps=FPS, plugin="pyav", codec="libx264")
    return tmp


def _pick_resolution(path, preset):
    w, h = RES_PRESETS[preset]
    try:
        f0 = iio.imread(path, plugin="pyav", index=0)
        if f0.shape[0] > f0.shape[1]:  # portrait
            w, h = h, w
    except Exception:
        pass
    return w, h


# --- Load native pipeline + IC-LoRA once at module scope (ZeroGPU packs weights here) ---
print("Downloading checkpoints…")
checkpoint_path = hf_hub_download(LTX_MODEL_REPO, "ltx-2.3-22b-distilled-1.1.safetensors", token=HF_TOKEN)
spatial_upsampler_path = hf_hub_download(LTX_MODEL_REPO, "ltx-2.3-spatial-upscaler-x2-1.1.safetensors", token=HF_TOKEN)
gemma_root = snapshot_download(GEMMA_REPO, token=HF_TOKEN)
lora_path = hf_hub_download(LORA_REPO, LORA_FILE, token=HF_TOKEN)

print("Building ICLoraPipeline…")
pipeline = ICLoraPipeline(
    distilled_checkpoint_path=checkpoint_path,
    spatial_upsampler_path=spatial_upsampler_path,
    gemma_root=gemma_root,
    loras=[LoraPathStrengthAndSDOps(lora_path, LORA_SCALE, LTXV_LORA_COMFY_RENAMING_MAP)],
    # bf16 (NOT fp8): the IC-LoRA is fused into the transformer at MODULE SCOPE (the GPU
    # worker can't re-open the checkpoint file). fp8_cast()'s fusion runs a custom CUDA kernel
    # that can't be ZeroGPU-virtualised; the bf16 fuse rule is pure torch -> virtualisable.
    quantization=None,
)


def _preload_pin(ledger, tag):
    if ledger is None:
        return
    for name in ["transformer", "video_encoder", "video_decoder", "audio_encoder",
                 "audio_decoder", "vocoder", "spatial_upsampler", "text_encoder",
                 "gemma_embeddings_processor"]:
        fn = getattr(ledger, name, None)
        if callable(fn):
            try:
                obj = fn()
                setattr(ledger, name, (lambda o=obj: o))
                print(f"[preload {tag}] {name} ✓")
            except Exception as e:
                print(f"[preload {tag}] {name} skipped: {e}")

# Preload stage 1 always; preload stage 2 only when two-stage is used (skip_stage_2=False).
# Eagerly pinning both ledgers materializes TWO ~46GB transformers — too big for the ZeroGPU pack.
_preload_pin(getattr(pipeline, "stage_1_model_ledger", None), "stage1")
if not SKIP_STAGE_2:
    _preload_pin(getattr(pipeline, "stage_2_model_ledger", None), "stage2")
print("Pipeline ready.")

# ============================ AOTI (native bf16 transformer graph) ============================
AOTI_REPO = os.environ.get("AOTI_REPO", "linoyts/LTX-2.3-Native-Transformer-GroupA-sm120-cu130-r20")
import types as _types
from dataclasses import replace as _dc_replace
from ltx_core.model.transformer.transformer_args import TransformerArgs as _TA
_TA_FIELDS = list(_TA.__dataclass_fields__.keys())
def _flatten_ta(ta):
    out = []
    for f in _TA_FIELDS:
        v = getattr(ta, f)
        if torch.is_tensor(v):
            out.append(v)
        elif isinstance(v, tuple) and len(v) > 0 and all(torch.is_tensor(x) for x in v):
            out.extend(v)
    return out
def _install_aoti():
    velocity = pipeline.stage_1_model_ledger.transformer().velocity_model
    spaces.aoti_load(module=velocity, repo_id=AOTI_REPO)
    def _proc(self, video, audio, perturbations):
        for blk in self.transformer_blocks:
            o = blk(*(_flatten_ta(video) + _flatten_ta(audio)))
            video = _dc_replace(video, x=o[0]); audio = _dc_replace(audio, x=o[1])
        return video, audio
    velocity._process_transformer_blocks = _types.MethodType(_proc, velocity)
    print(f"[AOTI] loaded {AOTI_REPO} + patched block loop", flush=True)
print(f"[AOTI] base torch={torch.__version__} cuda={torch.version.cuda}", flush=True)
try:
    _install_aoti(); print("[AOTI] OK", flush=True)
except Exception as _e:
    import traceback; traceback.print_exc(); print(f"[AOTI] FAILED ({_e!r}) -> EAGER", flush=True)
# ==============================================================================================


def _duration(*args, **kwargs):
    nf = next((a for a in args if isinstance(a, int) and a in FRAME_CHOICES), DEFAULT_FRAMES)
    return int(60 + nf * 1.2)


@spaces.GPU(duration=_duration)
@torch.inference_mode()
def add_water(video, prompt, preset, num_frames, seed, randomize, progress=gr.Progress(track_tqdm=True)):
    if video is None:
        raise gr.Error("Please upload a video.")
    if not prompt.strip():
        raise gr.Error("Describe the result (e.g. 'a brown rabbit on grey rocks, soft birdsong').")
    seed = random.randint(0, MAX_SEED) if randomize else int(seed)
    num_frames = int(num_frames)
    width, height = _pick_resolution(video, preset)
    ref_path = _prep_reference(video, width, height, num_frames)
    tiling = TilingConfig.default()
    # skip_stage_2 outputs at half the passed dims -> pass 2x so output matches the preset.
    gen_w, gen_h = (width * 2, height * 2) if SKIP_STAGE_2 else (width, height)
    video_out, audio_out = pipeline(
        prompt=build_prompt(prompt),
        seed=seed, height=gen_h, width=gen_w,
        num_frames=num_frames, frame_rate=FPS,
        images=[], video_conditioning=[(ref_path, 1.0)],
        skip_stage_2=SKIP_STAGE_2, tiling_config=tiling,
    )
    out_path = tempfile.mktemp(suffix=".mp4")
    encode_video(video=video_out, fps=FPS, audio=audio_out, output_path=out_path,
                 video_chunks_number=get_video_chunks_number(num_frames, tiling))
    return out_path, seed


# --- UI config (match the public Space exactly) ---
RES_PRESETS = {"960×544 (fast)": (960, 544), "1216×704 (recommended)": (1216, 704),
               "1536×864 (high)": (1536, 864), "1920×1088 (native)": (1920, 1088)}
FRAME_CHOICES = [49, 73, 97, 121]
LORA_SCALE = 1.2  # native fixed scale (matches public default)


with gr.Blocks(title="LTX-2.3 Water Simulation") as demo:
    gr.Markdown(
        "# 🌊 LTX-2.3 Water Simulation\n"
        "Add believable, naturally-moving water to a dry clip — rivers, surf, rain, waterfalls, floods, "
        "splashes — that interacts with the moving scene, while maintaining subject and framing identity. "
        "Using [LTX 2.3 Distilled](https://huggingface.co/Lightricks/LTX-2.3) with the "
        "[Water Simulation IC-LoRA](https://huggingface.co/Lightricks/LTX-2.3-22b-IC-LoRA-Water-Simulation)."
    )
    gr.Markdown("⚡ **Accelerated with [AOTI](https://huggingface.co/linoyts/LTX-2.3-Native-Transformer-GroupA-sm120-cu130-r20)** — precompiled transformer for faster inference.")
    with gr.Row():
        with gr.Column():
            video_in = gr.Video(label="Dry input video")
            prompt = gr.Textbox(
                label="Describe the water — type, motion, how it interacts, plus any sounds", lines=3,
                placeholder="a clear shallow stream braiding around their legs with white foam crests and glistening wet ground; rushing water, gentle splashing",
            )
            with gr.Accordion("Settings", open=False):
                preset = gr.Dropdown(list(RES_PRESETS), value="1216×704 (recommended)", label="Resolution")
                num_frames = gr.Dropdown(FRAME_CHOICES, value=73, label="Frames (24fps)")
                randomize = gr.Checkbox(True, label="Randomize seed")
                seed = gr.Slider(0, MAX_SEED, value=42, step=1, label="Seed")
            run = gr.Button("Add water", variant="primary")
        with gr.Column():
            video_out = gr.Video(label="Result with water")

    run.click(add_water, inputs=[video_in, prompt, preset, num_frames, seed, randomize],
              outputs=[video_out, seed])

    gr.Examples(
        examples=[
            ['examples/man_dancing_dry.mp4', 'a clear, shallow stream rushing and braiding around their legs — cold mountain water swirling with white foam crests and glassy ripples, the wet floor glistening and throwing back reflections, bright droplets kicked up with every step; lively rushing water and rhythmic splashes as they move', '1216×704 (recommended)', 73, 42, False],
            ['examples/landscape_dry.mp4', 'a wide river flooding across the valley floor — the water spreading in glassy sheets with rippling reflections of the sky and drifting ribbons of white foam, a soft mist rising off the surface in the cool light; steadily flowing water and the distant rush of a waterfall', '1216×704 (recommended)', 73, 42, False],
        ],
        inputs=[video_in, prompt, preset, num_frames, seed, randomize],
        outputs=[video_out, seed], fn=add_water, cache_examples=True, cache_mode="lazy",
    )

if __name__ == "__main__":
    demo.launch(show_error=True)