Switch backend to native LTX-2 (ICLoraPipeline)

README.md

@@ -11,7 +11,7 @@ pinned: false
 hardware: zero-a10g
 short_description: Colorize B&W video with an LTX-2.3 IC-LoRA
 models:
-  - diffusers/LTX-2.3-Distilled-Diffusers
+  - Lightricks/LTX-2.3
   - Lightricks/LTX-2.3-22b-IC-LoRA-Colorization
 ---
 
@@ -21,6 +21,6 @@ Restores natural color to grayscale, monochrome, or desaturated video while keep
 subject identity, framing, and motion untouched — only color changes.
 
 Runs the **colorization IC-LoRA** from [`linoyts/LTX-2.3-loras`](https://huggingface.co/linoyts/LTX-2.3-loras)
-on the distilled [`diffusers/LTX-2.3-Distilled-Diffusers`](https://huggingface.co/diffusers/LTX-2.3-Distilled-Diffusers)
-checkpoint via `LTX2InContextPipeline`. Upload any clip — it's desaturated to a grayscale
+on the distilled [`Lightricks/LTX-2.3`](https://huggingface.co/Lightricks/LTX-2.3)
+checkpoint via the native LTX-2 pipeline. Upload any clip — it's desaturated to a grayscale
 reference and recolored from your prompt. 8-step distilled schedule, LoRA scale 1.0.

app.py

@@ -1,56 +1,147 @@
 import os
-
-# ZeroGPU: torch.compile / dynamo are unsupported — disable before torch import.
-os.environ.setdefault("TORCH_COMPILE_DISABLE", "1")
-os.environ.setdefault("TORCHDYNAMO_DISABLE", "1")
-
+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"
+
+# memory-efficient attention
+subprocess.run([sys.executable, "-m", "pip", "install", "xformers==0.0.32.post2", "--no-build-isolation"], check=False)
+
+# --- 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 threading
-import time
 
 import numpy as np
 import imageio.v3 as iio
-import spaces
-import torch
-import gradio as gr
 from PIL import Image, ImageOps
-from huggingface_hub import hf_hub_download
-from safetensors.torch import load_file
 
-from diffusers import LTX2InContextPipeline
-from diffusers.pipelines.ltx2.pipeline_ltx2_ic_lora import LTX2ReferenceCondition
-from diffusers.pipelines.ltx2.utils import DISTILLED_SIGMA_VALUES
-from diffusers.utils import load_video, encode_video
+import torch
+torch._dynamo.config.suppress_errors = True
+torch._dynamo.config.disable = True
 
-# --- Config -----------------------------------------------------------------
-BASE_MODEL = "diffusers/LTX-2.3-Distilled-Diffusers"
+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:
+#     safetensors.safe_open(path, framework="pt", device="cuda")
+# which performs the host->device copy inside safetensors' own C++ (cudaMemcpy), BYPASSING
+# torch.Tensor.to — the exact call ZeroGPU patches to virtualise + pack weights at module
+# scope. The result: "No CUDA GPUs are available" at startup and nothing gets packed.
+# Patch the loader to open on CPU and move via torch.Tensor.to (which ZeroGPU virtualises),
+# so the module-scope preload packs correctly — matching diffusers / the reference Space.
+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): do NOT query CUDA at module scope. torch.cuda.get_device_capability()
+# forces torch._C._cuda_init() in the GPU-less main process, which poisons ZeroGPU's CUDA
+# virtualization — the module-scope model preload then fails with "No CUDA GPUs are available"
+# and ZeroGPU can't pack the weights. SDPA works on every GPU (incl. Blackwell ZeroGPU, where
+# FA3 crashes), so patch it unconditionally without ever touching torch.cuda here.
+_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 Colorize (native LTX-2)"
 LORA_REPO = "Lightricks/LTX-2.3-22b-IC-LoRA-Colorization"
 LORA_FILE = "ltx-2.3-22b-ic-lora-colorization-0.9.safetensors"
 LORA_SCALE = 1.0
-FPS = 24
-NUM_STEPS = len(DISTILLED_SIGMA_VALUES)  # 8-step distilled schedule
-MAX_SEED = np.iinfo(np.int32).max
-HF_TOKEN = os.environ.get("HF_TOKEN")
-
+SKIP_STAGE_2 = True          # restoration LoRA: stage-1-only native hi-res (per card)
+GRAYSCALE_REF = True         # colorize conditions on a B&W reference
 RES_PRESETS = {"960×544 (recommended)": (960, 544), "768×448 (fast)": (768, 448)}
+DEFAULT_PRESET = "960×544 (recommended)"
 FRAME_CHOICES = [49, 73, 97, 121]
+DEFAULT_FRAMES = 121
 
-# --- Load pipeline once at module scope (ZeroGPU registers it) ---------------
-pipe = LTX2InContextPipeline.from_pretrained(BASE_MODEL, torch_dtype=torch.bfloat16)
-pipe.to("cuda")
-pipe.vae.enable_tiling()
+def build_prompt(p):
+    return (
+        "Reference shows the same scene in high-contrast monochrome with soft natural daylight. "
+        "Edited shows the same scene with natural colors restored. "
+        f"COLORIZE {p.strip()}. "
+        "Subject identity, framing, and background geometry are identical to the reference; "
+        "only color information differs between reference and edited."
+    )
 
-_lora_path = hf_hub_download(LORA_REPO, LORA_FILE, token=HF_TOKEN)
-pipe.load_lora_weights(load_file(_lora_path), adapter_name="colorize")
-pipe.fuse_lora(lora_scale=LORA_SCALE)
-pipe.unload_lora_weights()
-# AOTI: load precompiled transformer blocks at ROOT level (ZeroGPU loads on cuda at
-# module scope; do NOT lazy-load or move to cuda inside @spaces.GPU).
-spaces.aoti_load(module=pipe.transformer, repo_id="ltx-community/LTX-2.3-Transformer-GroupA-sm120-cu130-r9e")
+EXAMPLES = [
+    ["examples/rabbit_rocks_gray.mp4",
+     "a young brown cottontail rabbit with warm tan and grey-brown fur, a pale cream underside and soft pink inner ears, perched on weathered grey granite boulders flecked with green and ochre lichen in warm late-afternoon sun; gentle wind, distant birdsong and soft rustling grass",
+     "960×544 (recommended)", 121, 42, False],
+    ["examples/surfing_gray.mp4",
+     "a surfer in a black wetsuit riding a curling turquoise ocean wave, bright white foam spraying off the crest, deep blue sky and sunlit teal water; powerful ocean waves crashing, rushing water, wind and distant seagulls",
+     "960×544 (recommended)", 121, 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"
 
 
-# --- Helpers ----------------------------------------------------------------
 def _src_fps(path, default=FPS):
     try:
         return float(iio.immeta(path, plugin="pyav").get("fps", default)) or default
@@ -58,105 +149,126 @@ def _src_fps(path, default=FPS):
         return default
 
 
-def _load_frames(path, num_frames, width, height):
-    """Natural-speed (real-time at 24fps), aspect-preserving (center-crop) frames."""
-    frames = load_video(path)
-    if not frames:
-        return []
-    fps = _src_fps(path)
+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 = int(round(i / FPS * fps))
-        idx = min(idx, len(frames) - 1)
-        out.append(ImageOps.fit(frames[idx].convert("RGB"), (width, height), Image.LANCZOS))
-    return out
-
-
-def _to_grayscale(img):
-    return img.convert("L").convert("RGB")
-
-
-def _pick_resolution(first_frame, preset):
+        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]
-    if first_frame.height > first_frame.width:
-        w, h = h, w
+    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
 
 
-def _build_prompt(prompt):
-    # COLORIZE convention; the user's single elaborate prompt (colors + any sounds) goes after the keyword.
-    return (
-        "Reference shows the same scene in high-contrast monochrome with soft natural daylight. "
-        "Edited shows the same scene with natural colors restored. "
-        f"COLORIZE {prompt.strip()}. "
-        "Subject identity, framing, and background geometry are identical to the reference; "
-        "only color information differs between reference and edited."
-    )
-
-
-def _export(video_np, audio, path):
-    kw = {}
-    if audio is not None:
-        kw = dict(audio=audio[0].float().cpu(), audio_sample_rate=pipe.vocoder.config.output_sampling_rate)
-    encode_video(video_np, fps=FPS, output_path=path, **kw)
+# --- 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 must be 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 ("CUDA error: no CUDA-capable device"), but the bf16
+    # fuse rule is pure torch matmul/add → virtualisable + packable. ~53GB pack (fits H200).
+    quantization=None,
+)
+
+
+# All components (incl. the bf16-fused transformer) load + pin at MODULE SCOPE so ZeroGPU
+# packs them (~53GB) and transfers them into each GPU worker — the worker can't re-open the
+# checkpoint file, so nothing may be built there. The CPU-open loader patch above makes the
+# host->device moves virtualisable; bf16 keeps the LoRA fusion virtualisable too.
+def _preload_pin(ledger, tag):
+    if ledger is None:
+        return
+    for name in ["video_encoder", "video_decoder", "audio_encoder", "audio_decoder",
+                 "vocoder", "spatial_upsampler", "text_encoder", "gemma_embeddings_processor",
+                 "transformer"]:
+        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_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 (all components preloaded + pinned for ZeroGPU packing).")
 
 
 def _duration(*args, **kwargs):
-    preset = next((a for a in args if isinstance(a, str) and a in RES_PRESETS), None)
-    num_frames = next((a for a in args if isinstance(a, int) and a in FRAME_CHOICES), 121)
-    w, h = RES_PRESETS.get(preset, (960, 544))
-    per_frame = max(1.0, (w * h) / (768 * 448))
-    return int(25 + int(num_frames) * per_frame * 0.65)  # ~2.3x measured runtime
-
+    nf = next((a for a in args if isinstance(a, int) and a in FRAME_CHOICES), DEFAULT_FRAMES)
+    return int(60 + nf * 1.2)
 
-# --- Inference --------------------------------------------------------------
 
 @spaces.GPU(duration=_duration)
-def colorize(video, prompt, preset, num_frames, seed, randomize,
-             progress=gr.Progress(track_tqdm=True)):
+@torch.inference_mode()
+def colorize(video, prompt, preset, num_frames, seed, randomize, progress=gr.Progress(track_tqdm=True)):
     if video is None:
-        raise gr.Error("Please upload a video to colorize.")
+        raise gr.Error("Please upload a video.")
     if not prompt.strip():
-        raise gr.Error("Describe the colorized scene (e.g. 'a brown rabbit on grey granite rocks, soft birdsong').")
-
-    if randomize:
-        seed = random.randint(0, MAX_SEED)
-    seed = int(seed)
+        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)
-
-    probe = load_video(video)
-    if not probe:
-        raise gr.Error("Could not read any frames from that video.")
-    width, height = _pick_resolution(probe[0], preset)
-
-    ref = [_to_grayscale(f) for f in _load_frames(video, num_frames, width, height)]
-    prompt = _build_prompt(prompt)
-
-
-    video_out, audio_out = pipe(
-        prompt=prompt,
-        negative_prompt="",
-        reference_conditions=[LTX2ReferenceCondition(frames=ref, strength=1.0)],
-        reference_downscale_factor=1,
-        width=width, height=height, num_frames=num_frames, frame_rate=FPS,
-        num_inference_steps=NUM_STEPS, sigmas=DISTILLED_SIGMA_VALUES,
-        guidance_scale=1.0, stg_scale=0.0, audio_guidance_scale=1.0, audio_stg_scale=0.0,
-        generator=torch.Generator(device="cuda").manual_seed(seed),
-        output_type="np", return_dict=False,
+    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 (height//2, width//2) — pass 2× so the
+    # final video matches the chosen preset. (Two-stage demos pass the preset directly.)
+    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.NamedTemporaryFile(suffix=".mp4", delete=False).name
-    _export(video_out[0], audio_out, out_path)
+    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 ---------------------------------------------------------------------
+# --- UI config (match the public Space exactly) ---
+RES_PRESETS = {"960×544 (recommended)": (960, 544), "768×448 (fast)": (768, 448)}
+FRAME_CHOICES = [49, 73, 97, 121]
+
+
 with gr.Blocks(title="LTX-2.3 Colorize") as demo:
     gr.Markdown(
         "# 🎨 LTX-2.3 Video Colorization\n"
         "Restore natural color to black-and-white or desaturated footage while keeping subject, framing and "
-        "motion identity. Using [LTX 2.3 Distilled](https://huggingface.co/diffusers/LTX-2.3-Distilled-Diffusers) "
-        "with the [Colorization IC-LoRA](https://huggingface.co/Lightricks/LTX-2.3-22b-IC-LoRA-Colorization), via diffusers 🧨."
+        "motion identity. Using [LTX 2.3 Distilled](https://huggingface.co/Lightricks/LTX-2.3) "
+        "with the [Colorization IC-LoRA](https://huggingface.co/Lightricks/LTX-2.3-22b-IC-LoRA-Colorization)."
     )
     with gr.Row():
         with gr.Column():

requirements.txt

@@ -1,9 +1,11 @@
-git+https://github.com/huggingface/diffusers
-transformers
+transformers==4.57.6
 accelerate
-peft
-safetensors
-sentencepiece
-imageio
-imageio-ffmpeg
+torch==2.8.0
+torchaudio==2.8.0
+einops
+scipy
 av
+scikit-image>=0.25.2
+flashpack==0.1.2
+imageio[ffmpeg]
+pillow