Replace Comic-Gen with Chinese→Malay Video Dubber (CPU Basic)

Dockerfile

@@ -1,9 +1,10 @@
 # HF-recommended base for Gradio Spaces
 FROM python:3.10-slim
 
-# Install system dependencies
+# Install system dependencies (ffmpeg for audio/video processing)
 RUN apt-get update && apt-get install -y --no-install-recommends \
     git \
+    ffmpeg \
     && rm -rf /var/lib/apt/lists/*
 
 # Create user with UID 1000 (HF requirement)
@@ -14,7 +15,7 @@ ENV HOME=/home/user \
 
 WORKDIR $HOME/app
 
-# Copy requirements and install (single pip install, pin Gradio 4.44.1)
+# Copy requirements and install
 COPY --chown=user requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt \
     "uvicorn>=0.14.0" \
@@ -27,4 +28,4 @@ ENV GRADIO_SERVER_NAME=0.0.0.0 \
     GRADIO_SERVER_PORT=7860 \
     PYTHONUNBUFFERED=1
 
-CMD ["python", "app.py"]
+CMD ["python", "app.py"]

README.md

@@ -1,70 +1,28 @@
 ---
-title: Comic Scene Generator
-emoji: 🎨
-colorFrom: purple
-colorTo: pink
-sdk: docker
-sdk_version: "5.0.0"
+title: Chinese To Malay Video Dubber CPU
+emoji: 🇨🇳🇲🇾
+colorFrom: blue
+colorTo: green
+sdk: gradio
 app_file: app.py
 pinned: false
-license: apache-2.0
-tags:
-  - stable-diffusion
-  - sdxl
-  - noobai
-  - manga
-  - anime
-  - cpu-basic
 ---
 
-# Comic Scene Generator
+# Chinese → Bahasa Melayu Video Dubber — CPU Basic
 
-最小可用版：输入一句中文场景描述 → 自动扩写为适合 NoobAI-XL 的英文提示词 → 生成 640×960 漫画/二次元场景图。
+Designed for Hugging Face Spaces CPU Basic (2 vCPU / 16 GB RAM).
 
-## 功能特性
+Pipeline:
 
-- **Prompt 自动增强**：中文场景自动翻译 + 质量标签 + 风格标签（masterpiece, best quality, dynamic composition 等）
-- **NoobAI-XL 1.1 SDNQ UINT4**：量化模型 ~2.56 GB，Diffusers 仓库 ~3.63 GB，适配 16GB RAM
-- **CPU Basic 优化**：
-  - `low_cpu_mem_usage=True` 全链路
-  - VAE slicing + tiling
-  - Attention slicing
-  - `enable_model_cpu_offload()` 逐层卸载
-  - 单张生成，关闭并行
-- **默认配置**：640×960、20 steps、Euler Ancestral、guidance 5.0
-- **输出**：PNG 格式
+Chinese MP4 → Whisper Small → Chinese → Malay translation → Meta MMS Malay TTS → FFmpeg → Malay MP4
 
-## 部署说明
+Models:
+- `openai/whisper-small`
+- `Helsinki-NLP/opus-mt-zh-ms`
+- `facebook/mms-tts-zlm`
 
-1. 在 Hugging Face 创建 **Gradio Space**（SDK: Gradio, Hardware: CPU Basic）
-2. 上传三个文件：`app.py`、`requirements.txt`、`README.md`
-3. 等待 Build 完成（首次下载模型约 3.6 GB）
-4. 打开网页，输入中文场景，点击 Generate
+This version does not clone the original speaker's voice. It is intentionally lightweight for CPU Basic.
 
-## 使用示例
+Recommended test: 1–3 minute videos. Longer videos can take a long time on CPU.
 
-| 输入 | 风格 | 分辨率 | Steps |
-|------|------|--------|-------|
-| 一只愤怒的霸王龙站在暴雨中的史前森林里仰头怒吼，闪电照亮天空，低角度镜头，强烈的漫画构图，动态线条，丰富背景细节 | Manga / Anime | 640×960 | 20 |
-| 赛博朋克城市夜晚，霓虹灯招牌倒映在湿漉漉的街道上，孤独的少女撑着透明伞，电影感构图 | Manga / Anime | 640×960 | 20 |
-| 古风仙侠，白衣剑仙御剑飞行穿越云海，日出金光洒在云层上，宏大壮观，工笔画风 | Manga / Anime | 640×960 | 25 |
-
-## 性能预期 (CPU Basic: 2 vCPU / 16GB RAM)
-
-| 分辨率 | Steps | 预估耗时 |
-|--------|-------|----------|
-| 640×960 | 20 | ~3-5 分钟 |
-| 640×960 | 30 | ~5-7 分钟 |
-| 512×768 | 20 | ~2-3 分钟 |
-
-> 首次冷启动需额外下载模型权重（~3.6 GB），之后缓存复用。
-
-## 模型来源
-
-- **NoobAI-XL/NoobAI-XL-1.1-SDNQ-UINT4** (Apache-2.0)
-- SDNQ 量化：INT4 权重 + FP16 激活，体积约 1/3 原版
-
-## 许可证
-
-Apache-2.0（遵循 NoobAI-XL 原模型许可）
-# Rebuild trigger Tue Sep  1 15:58:37 UTC 2026
+FFmpeg/ffprobe are required by the Space runtime.

app.py

@@ -1,221 +1,229 @@
 import os
-import torch
-import gradio as gr
-from diffusers import StableDiffusionXLPipeline, AutoencoderKL
-from diffusers.schedulers import EulerAncestralDiscreteScheduler
-from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
-from safetensors.torch import load_file
-import gc
-
-# ─── Config ──────────────────────────────────────────────────────────
-MODEL_ID = "Laxhar/noobai-XL-1.1"  # HF repo (public, diffusers-compatible)
-DEVICE = "cpu"
-DTYPE = torch.bfloat16  # CPU 不支持 fp16 运算，改用 bfloat16（同样 2 字节/参数，省内存且 CPU 原生支持）
-DEFAULT_W, DEFAULT_H = 640, 960
-DEFAULT_STEPS = 20
-DEFAULT_GUIDANCE = 5.0
-
-# ─── Prompt enhancement (Chinese → English + quality tags) ───────────
-QUALITY_TAGS = (
-    "masterpiece, best quality, ultra high res, "
-    "dynamic composition, dramatic lighting, "
-    "detailed background, manga style, anime style, "
-    "sharp focus, vivid colors, professional illustration"
-)
-
-NEGATIVE_PROMPT = (
-    "low quality, worst quality, bad anatomy, bad composition, "
-    "blurry, pixelated, watermark, text, signature, logo, "
-    "ugly, deformed, noise, artifacts, lowres, error, "
-    "missing fingers, extra digits, fewer digits, cropped, "
-    "jpeg artifacts, out of frame, mutation, mutated, "
-    "extra limbs, malformed limbs, long neck, disfigured"
-)
-
-STYLE_PROMPTS = {
-    "Manga / Anime": "manga style, anime style, screentone, ink lines, dynamic panel composition",
-    "Western Comics": "western comic style, bold outlines, halftone dots, vibrant colors, dynamic action",
-    "Manhwa": "manhwa style, webtoon style, clean lines, soft shading, vertical composition",
-    "Sketch / Lineart": "sketch style, lineart, black and white, pencil drawing, detailed strokes",
-}
-
-def enhance_prompt(user_prompt: str, style: str) -> str:
-    """中文场景 → 英文提示词 + 质量标签 + 风格标签"""
-    # 简单的中文关键词映射（可扩展为更完整的翻译/扩写）
-    cn_en_map = {
-        "恐龙": "dinosaur",
-        "霸王龙": "T-Rex",
-        "暴雨": "heavy rain",
-        "史前森林": "prehistoric forest",
-        "仰头": "looking up",
-        "怒吼": "roaring",
-        "闪电": "lightning",
-        "照亮": "illuminating",
-        "天空": "sky",
-        "低角度": "low angle",
-        "镜头": "camera angle",
-        "强烈": "intense",
-        "漫画": "manga",
-        "构图": "composition",
-        "动态": "dynamic",
-        "线条": "lines",
-        "丰富": "rich",
-        "背景": "background",
-        "细节": "details",
-        "愤怒": "angry",
-        "站在": "standing in",
-        "里": "in",
-        "的": "",
-        "一只": "a",
-        "在": "in",
-        "中": "in",
-        "和": "and",
-        "与": "and",
-    }
-    # 粗略翻译：把常见中文词替换成英文
-    translated = user_prompt
-    for cn, en in cn_en_map.items():
-        translated = translated.replace(cn, en)
-    # 如果全是中文，做简单拼接；如果已经有英文，保留
-    if any('\u4e00' <= c <= '\u9fff' for c in translated):
-        # 还含有中文，简单处理
-        prompt_en = translated
-    else:
-        prompt_en = translated
-    style_tag = STYLE_PROMPTS.get(style, STYLE_PROMPTS["Manga / Anime"])
-    return f"{prompt_en}, {style_tag}, {QUALITY_TAGS}"
-
-# ─── Load pipeline (lazy, cached) ────────────────────────────────────
-_pipe = None
-
-def get_pipeline():
-    global _pipe
-    if _pipe is not None:
-        return _pipe
-
-    print("Loading NoobAI-XL SDNQ pipeline on CPU (bf16)...")
-
-    # 1) VAE (单独加载，启用 slicing)
-    vae = AutoencoderKL.from_pretrained(
-        MODEL_ID,
-        subfolder="vae",
-        torch_dtype=DTYPE,
-        low_cpu_mem_usage=True,
-    )
-    vae.enable_slicing()
-    vae.enable_tiling()
-
-    # 2) Text encoders & tokenizers (SDXL 双编码器)
-    tokenizer_1 = CLIPTokenizer.from_pretrained(MODEL_ID, subfolder="tokenizer")
-    tokenizer_2 = CLIPTokenizer.from_pretrained(MODEL_ID, subfolder="tokenizer_2")
-    text_encoder_1 = CLIPTextModel.from_pretrained(
-        MODEL_ID, subfolder="text_encoder", torch_dtype=DTYPE, low_cpu_mem_usage=True
-    )
-    text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(
-        MODEL_ID, subfolder="text_encoder_2", torch_dtype=DTYPE, low_cpu_mem_usage=True
-    )
-
-    # 3) Pipeline
-    pipe = StableDiffusionXLPipeline.from_pretrained(
-        MODEL_ID,
-        vae=vae,
-        text_encoder=text_encoder_1,
-        text_encoder_2=text_encoder_2,
-        tokenizer=tokenizer_1,
-        tokenizer_2=tokenizer_2,
-        scheduler=EulerAncestralDiscreteScheduler.from_pretrained(MODEL_ID, subfolder="scheduler"),
-        torch_dtype=DTYPE,
-        low_cpu_mem_usage=True,
-    )
-
-    # CPU 优化
-    # 注意：enable_model_cpu_offload() 依赖 CUDA（会把模块按需搬到 GPU），
-    # 纯 CPU 环境不能用，否则会报 "Found no NVIDIA driver"
-    pipe.enable_vae_slicing()
-    pipe.enable_vae_tiling()
-    pipe.enable_attention_slicing("auto")     # 注意力切片
-    # 关闭 xformers / flash attention（CPU 不支持）
-    try:
-        pipe.disable_xformers_memory_efficient_attention()
-    except Exception:
-        pass
+import re
+import subprocess
+import tempfile
+from pathlib import Path
 
-    pipe.to(DEVICE)
-    _pipe = pipe
-    print("Pipeline ready.")
-    return pipe
-
-# ─── Generation ──────────────────────────────────────────────────────
-def generate(prompt, style, width, height, steps, guidance, progress=gr.Progress()):
-    if not prompt.strip():
-        return None, "请输入场景描述"
+import gradio as gr
+import torch
+import torchaudio
+from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+from transformers import VitsModel
 
-    progress(0.1, desc="增强提示词...")
-    full_prompt = enhance_prompt(prompt, style)
+ASR_MODEL = os.getenv("ASR_MODEL", "openai/whisper-small")
+TRANSLATION_MODEL = os.getenv("TRANSLATION_MODEL", "Helsinki-NLP/opus-mt-zh-ms")
+TTS_MODEL = os.getenv("TTS_MODEL", "facebook/mms-tts-zlm")
 
-    progress(0.2, desc="加载模型...")
-    pipe = get_pipeline()
+torch.set_num_threads(2)
+DEVICE = "cpu"
 
-    progress(0.4, desc="生成中（CPU 慢，请耐心等待）...")
+_asr_processor = _asr_model = None
+_tr_tokenizer = _tr_model = None
+_tts_tokenizer = _tts_model = None
+
+def run(cmd):
+    p = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if p.returncode:
+        raise RuntimeError(p.stderr[-5000:])
+    return p.stdout
+
+def video_duration(path):
+    return float(run(["ffprobe","-v","error","-show_entries","format=duration",
+                      "-of","default=noprint_wrappers=1:nokey=1",str(path)]).strip())
+
+def extract_audio(video, wav):
+    run(["ffmpeg","-y","-i",str(video),"-vn","-ac","1","-ar","16000",str(wav)])
+
+def load_asr():
+    global _asr_processor, _asr_model
+    if _asr_model is None:
+        _asr_processor = AutoProcessor.from_pretrained(ASR_MODEL)
+        _asr_model = AutoModelForSpeechSeq2Seq.from_pretrained(
+            ASR_MODEL, low_cpu_mem_usage=True
+        ).to(DEVICE).eval()
+    return _asr_processor, _asr_model
+
+def transcribe(wav):
+    processor, model = load_asr()
+    audio, sr = torchaudio.load(str(wav))
+    audio = audio.mean(0)
+    if sr != 16000:
+        audio = torchaudio.functional.resample(audio, sr, 16000)
+
+    total = audio.numel() / 16000
+    pieces = []
+    chunk = 30
+
+    for start in range(0, int(total + 0.999), chunk):
+        x = audio[start*16000:min(int(total*16000),(start+chunk)*16000)].numpy()
+        if len(x) < 400:
+            continue
+        inputs = processor(x, sampling_rate=16000, return_tensors="pt")
+        with torch.inference_mode():
+            ids = model.generate(
+                inputs.input_features,
+                language="zh",
+                task="transcribe",
+                max_new_tokens=256
+            )
+        text = processor.batch_decode(ids, skip_special_tokens=True)[0].strip()
+        if text:
+            pieces.append(text)
+    return " ".join(pieces)
+
+def load_translation():
+    global _tr_tokenizer, _tr_model
+    if _tr_model is None:
+        _tr_tokenizer = AutoTokenizer.from_pretrained(TRANSLATION_MODEL)
+        _tr_model = AutoModelForSeq2SeqLM.from_pretrained(
+            TRANSLATION_MODEL, low_cpu_mem_usage=True
+        ).to(DEVICE).eval()
+    return _tr_tokenizer, _tr_model
+
+def translate(text):
+    tokenizer, model = load_translation()
+    parts = [x.strip() for x in re.split(r'(?<=[。！？!?])\s*', text) if x.strip()]
+    if not parts:
+        parts = [text]
+
+    result = []
+    for part in parts:
+        inputs = tokenizer(part, return_tensors="pt", truncation=True, max_length=256)
+        with torch.inference_mode():
+            ids = model.generate(**inputs, max_new_tokens=256, num_beams=4,
+                                 renormalize_logits=True)
+        result.append(tokenizer.decode(ids[0], skip_special_tokens=True).strip())
+    return " ".join(result)
+
+def load_tts():
+    global _tts_tokenizer, _tts_model
+    if _tts_model is None:
+        _tts_tokenizer = AutoTokenizer.from_pretrained(TTS_MODEL)
+        _tts_model = VitsModel.from_pretrained(
+            TTS_MODEL, low_cpu_mem_usage=True
+        ).to(DEVICE).eval()
+    return _tts_tokenizer, _tts_model
+
+def tts(text, wav):
+    tokenizer, model = load_tts()
+    inputs = tokenizer(text, return_tensors="pt")
     with torch.inference_mode():
-        image = pipe(
-            prompt=full_prompt,
-            negative_prompt=NEGATIVE_PROMPT,
-            width=width,
-            height=height,
-            num_inference_steps=steps,
-            guidance_scale=guidance,
-            num_images_per_prompt=1,
-        ).images[0]
-
-    progress(0.9, desc="保存 PNG...")
-    # 保存到临时文件供 Gradio 返回
-    out_path = "/tmp/comic_scene.png"
-    image.save(out_path, "PNG")
-    progress(1.0, desc="完成")
-    gc.collect()
-    torch.cuda.empty_cache() if torch.cuda.is_available() else None
-    return out_path, f"✅ 生成完成\nPrompt: {full_prompt[:200]}..."
-
-# ─── Gradio UI ───────────────────────────────────────────────────────
-css = """
-#title { text-align: center; font-size: 1.5rem; font-weight: 700; margin-bottom: 0.5rem; }
-#subtitle { text-align: center; color: #666; margin-bottom: 1rem; }
-.generate-btn { background: linear-gradient(90deg, #667eea 0%, #764ba2 100%) !important; }
-"""
-
-with gr.Blocks(css=css, title="Comic Scene Generator") as demo:
-    gr.HTML('<div id="title">🎨 Comic Scene Generator</div>')
-    gr.HTML('<div id="subtitle">输入一句中文场景 → 自动扩写 → NoobAI-XL 生成漫画场景图</div>')
-
+        audio = model(**inputs).waveform.cpu()
+    torchaudio.save(str(wav), audio, model.config.sampling_rate)
+
+def fit_audio(src, dst, target):
+    target = max(float(target), 0.1)
+    ratio = video_duration(src) / target
+    filters = []
+    while ratio > 2:
+        filters.append("atempo=2.0")
+        ratio /= 2
+    while ratio < 0.5:
+        filters.append("atempo=0.5")
+        ratio /= 0.5
+    filters.append(f"atempo={ratio:.6f}")
+    run(["ffmpeg","-y","-i",str(src),"-filter:a",",".join(filters),
+         "-t",f"{target:.3f}","-ar","24000","-ac","1",str(dst)])
+
+def make_dub(ms, total, work):
+    parts = [x.strip() for x in re.split(r'(?<=[。！？!?])\s*|(?<=[，,])\s+', ms) if x.strip()]
+    if not parts:
+        parts = [ms]
+    weights = [max(1,len(re.sub(r'\s+','',x))) for x in parts]
+    total_w = sum(weights)
+
+    silent = work/"silent.wav"
+    run(["ffmpeg","-y","-f","lavfi","-i","anullsrc=r=24000:cl=mono",
+         "-t",f"{total:.3f}","-ar","24000","-ac","1",str(silent)])
+
+    inputs = ["-i",str(silent)]
+    labels = ["[0:a]"]
+    start = 0.0
+
+    for i, part in enumerate(parts, 1):
+        slot = total * weights[i-1] / total_w
+        raw, fitted = work/f"tts_{i}.wav", work/f"fit_{i}.wav"
+        tts(part, raw)
+        fit_audio(raw, fitted, slot)
+        inputs += ["-i",str(fitted)]
+        delay = int(start*1000)
+        labels.append(f"[{i}:a]adelay={delay}|{delay}[d{i}]")
+        start += slot
+
+    mix = "[0:a]" + "".join(f"[d{i}]" for i in range(1,len(parts)+1))
+    filt = ";".join(labels) + f";{mix}amix=inputs={len(parts)+1}:duration=first[a]"
+    out = work/"dub.wav"
+    run(["ffmpeg","-y",*inputs,"-filter_complex",filt,"-map","[a]",
+         "-ar","24000","-ac","1",str(out)])
+    return out
+
+def process(video, keep_original, original_volume, progress=gr.Progress()):
+    if not video:
+        raise gr.Error("请上传中文视频。")
+    work = Path(tempfile.mkdtemp(prefix="zhms_cpu_"))
+    try:
+        src = Path(video)
+        audio = work/"source.wav"
+        progress(0.05, desc="提取声音…")
+        extract_audio(src, audio)
+        total = video_duration(src)
+
+        progress(0.15, desc="中文语音识别（CPU）…")
+        zh = transcribe(audio)
+        if not zh:
+            raise RuntimeError("没有识别到中文语音。")
+
+        progress(0.42, desc="翻译成 Bahasa Melayu…")
+        ms = translate(zh)
+
+        progress(0.58, desc="生成马来文配音…")
+        dub = make_dub(ms, total, work)
+
+        out = work/"malay_dub.mp4"
+        progress(0.9, desc="合成 MP4…")
+        if keep_original:
+            filt = f"[0:a]volume={float(original_volume):.3f}[o];[o][1:a]amix=inputs=2:duration=first[a]"
+            run(["ffmpeg","-y","-i",str(src),"-i",str(dub),"-filter_complex",filt,
+                 "-map","0:v:0","-map","[a]","-c:v","copy","-c:a","aac","-b:a","160k",
+                 "-shortest",str(out)])
+        else:
+            run(["ffmpeg","-y","-i",str(src),"-i",str(dub),
+                 "-map","0:v:0","-map","1:a:0","-c:v","copy","-c:a","aac",
+                 "-b:a","160k","-shortest",str(out)])
+
+        txt = work/"translation.txt"
+        txt.write_text("=== 中文识别 ===\n"+zh+"\n\n=== Bahasa Melayu ===\n"+ms,encoding="utf-8")
+        progress(1, desc="完成")
+        return str(out), str(txt), zh, ms
+    except Exception as e:
+        raise gr.Error(str(e))
+
+with gr.Blocks(title="中文 → Bahasa Melayu（CPU Basic）") as demo:
+    gr.Markdown("""# 🇨🇳 → 🇲🇾 中文视频 → Bahasa Melayu
+**Hugging Face CPU Basic 专用版本**
+
+上传中文视频 → 中文语音识别 → 马来文翻译 → 马来文 TTS → 输出 MP4。
+
+- 不需要 GPU
+- 不需要 API Key
+- 不使用声音克隆
+- 使用专门的 Chinese → Malay 翻译模型
+- 使用 Meta MMS Malay TTS
+
+⚠️ CPU Basic 只有 2 vCPU / 16 GB RAM，建议先测试 1–3 分钟视频。
+""")
+    video = gr.Video(label="上传中文视频", sources=["upload"], type="filepath")
     with gr.Row():
-        with gr.Column(scale=1):
-            prompt = gr.Textbox(
-                label="Scene",
-                placeholder="例：一只愤怒的霸王龙站在暴雨中的史前森林里仰头怒吼，闪电照亮天空，低角度镜头，强烈的漫画构图，动态线条，丰富背景细节",
-                lines=4,
-            )
-            style = gr.Dropdown(
-                label="Style",
-                choices=list(STYLE_PROMPTS.keys()),
-                value="Manga / Anime",
-            )
-            with gr.Row():
-                width = gr.Slider(label="Width", minimum=320, maximum=1024, value=DEFAULT_W, step=64)
-                height = gr.Slider(label="Height", minimum=320, maximum=1536, value=DEFAULT_H, step=64)
-            steps = gr.Slider(label="Steps", minimum=10, maximum=50, value=DEFAULT_STEPS, step=1)
-            guidance = gr.Slider(label="Guidance Scale", minimum=1.0, maximum=15.0, value=DEFAULT_GUIDANCE, step=0.5)
-            btn = gr.Button("Generate", elem_classes="generate-btn", variant="primary")
-
-        with gr.Column(scale=1):
-            output_img = gr.Image(label="生成的漫画场景", type="filepath", interactive=False)
-            status = gr.Textbox(label="Status", interactive=False)
-
-    btn.click(
-        generate,
-        inputs=[prompt, style, width, height, steps, guidance],
-        outputs=[output_img, status],
-    )
+        keep = gr.Checkbox(label="保留原声音作为背景", value=False)
+        volume = gr.Slider(0,0.3,value=0.06,step=0.01,label="原声音量")
+    button = gr.Button("🚀 开始翻译成马来文", variant="primary")
+    output = gr.Video(label="🇲🇾 马来文视频")
+    text_file = gr.File(label="识别 + 翻译文本")
+    with gr.Row():
+        zh_box = gr.Textbox(label="中文识别结果", lines=8)
+        ms_box = gr.Textbox(label="Bahasa Melayu", lines=8)
+    button.click(process,[video,keep,volume],[output,text_file,zh_box,ms_box])
 
 if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)
+    demo.queue(max_size=2).launch()

requirements.txt

@@ -1,14 +1,7 @@
-# Core ML dependencies
-torch==2.4.0
-torchvision==0.19.0
-torchaudio==2.4.0
-
-diffusers==0.31.0
-transformers==4.44.0
-accelerate==0.33.0
-safetensors==0.4.4
-pillow==10.4.0
-
-# Gradio (use latest 5.x with latest gradio_client)
-gradio>=5.0.0
-gradio_client>=1.0.0
+gradio>=5,<7
+torch>=2.2
+torchaudio>=2.2
+transformers>=4.48,<5
+sentencepiece>=0.2
+safetensors>=0.4
+numpy<2.3