Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+examples/landscape_compressed.mp4 filter=lfs diff=lfs merge=lfs -text
+examples/people_compressed.mp4 filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,13 +1,20 @@
 ---
-title: Ltx 2.3 Decompress
-emoji: 🌖
+title: LTX-2.3 Video Decompress
+emoji: 🧼
 colorFrom: gray
-colorTo: pink
+colorTo: blue
 sdk: gradio
-sdk_version: 6.18.0
-python_version: '3.12'
+sdk_version: 6.13.0
+python_version: "3.12"
 app_file: app.py
 pinned: false
+hardware: zero-a10g
+short_description: Remove compression artifacts with an LTX-2.3 IC-LoRA
+models:
+  - diffusers/LTX-2.3-Distilled-Diffusers
+  - linoyts/LTX-2.3-loras
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# 🧼 LTX-2.3 Video Decompression
+Removes macroblocking, chroma bleed, ringing and banding from low-bitrate video via the decompression
+IC-LoRA on distilled LTX-2.3 (`LTX2InContextPipeline`, 8-step). Optional generated audio.

app.py

@@ -0,0 +1,178 @@
+import os
+
+os.environ.setdefault("TORCH_COMPILE_DISABLE", "1")
+os.environ.setdefault("TORCHDYNAMO_DISABLE", "1")
+
+import random
+import tempfile
+
+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
+
+# --- Config -----------------------------------------------------------------
+BASE_MODEL = "diffusers/LTX-2.3-Distilled-Diffusers"
+LORA_REPO = "linoyts/LTX-2.3-loras"
+LORA_FILE = "ltx-2.3-22b-ic-lora-decompression-0.9.safetensors"
+LORA_SCALE = 1.0
+FPS = 24
+NUM_STEPS = len(DISTILLED_SIGMA_VALUES)
+MAX_SEED = np.iinfo(np.int32).max
+HF_TOKEN = os.environ.get("HF_TOKEN")
+
+RES_PRESETS = {"Fast (768×448)": (768, 448), "Quality (960×544)": (960, 544)}
+FRAME_CHOICES = [49, 73, 97, 121]
+
+pipe = LTX2InContextPipeline.from_pretrained(BASE_MODEL, torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+pipe.vae.enable_tiling()
+_lora_path = hf_hub_download(LORA_REPO, LORA_FILE, token=HF_TOKEN)
+pipe.load_lora_weights(load_file(_lora_path), adapter_name="decompress")
+pipe.set_adapters("decompress", LORA_SCALE)
+
+
+def _src_fps(path, default=FPS):
+    try:
+        return float(iio.immeta(path, plugin="pyav").get("fps", default)) or default
+    except Exception:
+        return default
+
+
+def _load_frames(path, num_frames, width, height):
+    frames = load_video(path)
+    if not frames:
+        return []
+    fps = _src_fps(path)
+    out = []
+    for i in range(num_frames):
+        idx = min(int(round(i / FPS * fps)), len(frames) - 1)
+        out.append(ImageOps.fit(frames[idx].convert("RGB"), (width, height), Image.LANCZOS))
+    return out
+
+
+def _pick_resolution(first_frame, preset):
+    w, h = RES_PRESETS[preset]
+    if first_frame.height > first_frame.width:
+        w, h = h, w
+    return w, h
+
+
+def _build_prompt(scene, audio):
+    scene = scene.strip() or "the scene"
+    p = (
+        f"Reference shows {scene}, heavily compressed with visible macroblocking, chroma bleed, and ringing artifacts. "
+        f"Edited shows the same scene restored to high quality with sharp detail, clean edges, and no compression artifacts. "
+        f"ENHANCE QUALITY {scene}. "
+        f"Subject identity, framing, and background geometry are identical to the reference; "
+        f"only compression artifacts and image quality differ between reference and edited."
+    )
+    if audio.strip():
+        p += f" Audio: {audio.strip()}."
+    return p
+
+
+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)
+
+
+def _duration(*args, **kwargs):
+    preset = next((a for a in args if a in RES_PRESETS), "Fast")
+    num_frames = next((a for a in args if a in FRAME_CHOICES), 73)
+    per_frame = 1.6 if "Quality" in str(preset) else 1.0
+    return int(60 + int(num_frames) * per_frame)
+
+
+@spaces.GPU(duration=_duration)
+def decompress(video, scene, audio, preset, num_frames, seed, randomize,
+               progress=gr.Progress(track_tqdm=True)):
+    if video is None:
+        raise gr.Error("Please upload a compressed / low-bitrate video.")
+    if randomize:
+        seed = random.randint(0, MAX_SEED)
+    seed = 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 = _load_frames(video, num_frames, width, height)
+    prompt = _build_prompt(scene, audio)
+
+    def _cb(p, i, t, kw):
+        progress((i + 1) / NUM_STEPS, desc=f"Restoring — step {i + 1}/{NUM_STEPS}")
+        return {}
+
+    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, callback_on_step_end=_cb,
+    )
+    out_path = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False).name
+    _export(video_out[0], audio_out, out_path)
+    return out_path, seed
+
+
+with gr.Blocks(title="LTX-2.3 Decompress") as demo:
+    gr.Markdown(
+        "# 🧼 LTX-2.3 Video Decompression\n"
+        "Remove compression artifacts — macroblocking, chroma bleed, ringing, banding — from low-bitrate / "
+        "heavily-compressed video, restoring sharp detail while keeping subject, framing, and geometry intact. "
+        "Optionally describe the soundscape for generated audio. "
+        "IC-LoRA: [`linoyts/LTX-2.3-loras`](https://huggingface.co/linoyts/LTX-2.3-loras) · base: distilled LTX-2.3."
+    )
+    with gr.Row():
+        with gr.Column():
+            video_in = gr.Video(label="Compressed / low-bitrate video")
+            scene = gr.Textbox(label="Scene description (optional)", lines=2,
+                               placeholder="a busy city street with pedestrians and storefronts in daylight")
+            audio = gr.Textbox(label="Sound / audio (optional)", lines=1,
+                               placeholder="city ambience, footsteps, distant traffic")
+            with gr.Accordion("Settings", open=False):
+                preset = gr.Dropdown(list(RES_PRESETS), value="Fast (768×448)", 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("Restore", variant="primary")
+        with gr.Column():
+            video_out = gr.Video(label="Restored result")
+            used_seed = gr.Number(label="Seed used", interactive=False)
+
+    run.click(decompress, inputs=[video_in, scene, audio, preset, num_frames, seed, randomize],
+              outputs=[video_out, used_seed])
+
+    gr.Examples(
+        examples=[
+            ["examples/people_compressed.mp4",
+             "two people slow dancing together in a warm-lit room, smiling",
+             "soft music, gentle footsteps, a quiet room tone",
+             "Fast (768×448)", 73, 42, False],
+            ["examples/landscape_compressed.mp4",
+             "a misty green mountain landscape over calm still water",
+             "gentle wind over water, distant birdsong",
+             "Fast (768×448)", 73, 42, False],
+        ],
+        inputs=[video_in, scene, audio, preset, num_frames, seed, randomize],
+        outputs=[video_out, used_seed], fn=decompress, cache_examples=True, cache_mode="lazy",
+    )
+
+if __name__ == "__main__":
+    demo.launch(show_error=True)

examples/landscape_compressed.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:147eb9a424e835629eba177c99d210f36c7409699fa862b0cbbdfd6c96b2f074
+size 352741

examples/people_compressed.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97bf8f9791287330119c57ac820dff95782ebb6c5e5102b878779ff7ec824093
+size 387634

requirements.txt

@@ -0,0 +1,9 @@
+git+https://github.com/huggingface/diffusers
+transformers
+accelerate
+peft
+safetensors
+sentencepiece
+imageio
+imageio-ffmpeg
+av