CLI mode, CFG slider 1.0-1.5, better neg prompt, lazy example, MCP, local model paths

.gitignore

@@ -1 +1,3 @@
 models/
+*.png
+test.*

app.py

@@ -1,42 +1,39 @@
-"""Z-Anime 6B Image Generation (CPU) via sd-cli binary"""
+"""Z-Anime 6B Image Generation (CPU/GPU) via sd-cli binary
 
-import os, time, subprocess, tempfile, threading, mmap
-from PIL import Image
-import gradio as gr
+CLI:  python app.py "prompt" --seed 42 --cfg 1.0
+GUI:  python app.py --gradio
+"""
+
+import os, sys, time, subprocess, tempfile, threading, argparse
 
 # ---------------------------------------------------------------------------
-# Model paths (downloaded at build time)
+# Paths — auto-detect local vs Docker
 # ---------------------------------------------------------------------------
-DIFFUSION = "/app/models/z-anime-4step-q5_0.gguf"
-LLM = "/app/models/qwen3_4b_iq4xs.gguf"
-VAE = "/app/models/ae.safetensors"
-
-# Warm up page cache — read all model files so --mmap loads from RAM
-print("[init] Preloading models into page cache...")
-t0 = time.time()
-for model_path in [DIFFUSION, LLM, VAE]:
-    if os.path.exists(model_path):
-        sz = os.path.getsize(model_path)
-        with open(model_path, "rb") as f:
-            mm = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
-            mm.read()
-            mm.close()
-        print(f"  {os.path.basename(model_path)}: {sz / 1e9:.2f} GB cached")
-print(f"[init] Page cache warm in {time.time() - t0:.1f}s")
+_LOCAL_MODELS = os.path.join(os.path.dirname(__file__), "models")
+_DOCKER_MODELS = "/app/models"
+MODELS_DIR = _LOCAL_MODELS if os.path.isdir(_LOCAL_MODELS) else _DOCKER_MODELS
+
+_LOCAL_SDCLI = os.path.join(os.path.dirname(__file__), "..", "sd-cpp-tools", "build", "bin", "Release", "sd-cli.exe")
+_DOCKER_SDCLI = "/app/sd-cli"
+SD_CLI = _LOCAL_SDCLI if os.path.isfile(_LOCAL_SDCLI) else _DOCKER_SDCLI
+
+DIFFUSION = os.path.join(MODELS_DIR, "z-anime-distill-4step-q5_0.gguf")
+LLM = os.path.join(MODELS_DIR, "qwen3_4b_iq4xs.gguf")
+VAE = os.path.join(MODELS_DIR, "ae.safetensors")
 
 RESOLUTIONS = ["512x512", "768x512", "512x768"]
 STEPS = 4
-CFG = 1.0
 TIMEOUT = 10800
 
 _active_proc = None
 _proc_lock = threading.Lock()
 
 # ---------------------------------------------------------------------------
-# Inference
+# Core generation (shared by CLI and GUI)
 # ---------------------------------------------------------------------------
 
-def generate(prompt, negative_prompt, resolution, seed):
+def generate_image(prompt, negative_prompt="", resolution="512x512",
+                   seed=-1, cfg=1.0, output_path=None):
     """Generate an anime image using Z-Anime 6B model.
 
     Args:
@@ -44,24 +41,29 @@ def generate(prompt, negative_prompt, resolution, seed):
         negative_prompt: Things to avoid in the generated image.
         resolution: Image resolution (512x512, 768x512, or 512x768).
         seed: Random seed (-1 for random).
+        cfg: CFG scale (1.0 recommended for distill, higher = slower).
+        output_path: Where to save the image (auto if None).
 
     Returns:
-        tuple: Generated image and status message.
+        tuple: (output_path, status_message)
     """
     global _active_proc
 
     if not prompt or not prompt.strip():
-        raise gr.Error("Please enter a prompt.")
+        raise ValueError("Please enter a prompt.")
 
     prompt = prompt.strip()[:500]
     w, h = (int(x) for x in resolution.split("x"))
-    seed = int(seed or -1) if seed is not None else -1
+    seed = int(seed or -1)
+    cfg = float(cfg)
 
-    with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as f:
+    if output_path is None:
+        f = tempfile.NamedTemporaryFile(suffix=".png", delete=False)
         output_path = f.name
+        f.close()
 
     cmd = [
-        "/app/sd-cli",
+        SD_CLI,
         "--diffusion-model", DIFFUSION,
         "--llm", LLM,
         "--vae", VAE,
@@ -70,7 +72,7 @@ def generate(prompt, negative_prompt, resolution, seed):
         "-W", str(w),
         "-H", str(h),
         "--steps", str(STEPS),
-        "--cfg-scale", str(CFG),
+        "--cfg-scale", str(cfg),
         "--sampling-method", "euler_a",
         "-o", output_path,
         "--diffusion-fa",
@@ -78,97 +80,163 @@ def generate(prompt, negative_prompt, resolution, seed):
         "--vae-tiling",
         "--vae-conv-direct",
         "--tensor-type-rules", "^vae=f32",
-        # "--offload-to-cpu",  # causes disk swapping with --mmap, slower
-        # "--mmap",            # use with preload only if --offload-to-cpu is off
         "-v",
     ]
     if seed >= 0:
         cmd += ["-s", str(seed)]
 
-    print(f"[gen] {w}x{h} steps={STEPS} seed={seed} prompt={prompt[:80]}")
+    print(f"[gen] {w}x{h} steps={STEPS} cfg={cfg} seed={seed} prompt={prompt[:80]}")
     t0 = time.time()
 
+    proc = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+    with _proc_lock:
+        _active_proc = proc
+
     try:
-        proc = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+        stdout, stderr = proc.communicate(timeout=TIMEOUT)
+    except subprocess.TimeoutExpired:
+        proc.kill()
+        proc.wait()
         with _proc_lock:
-            _active_proc = proc
+            _active_proc = None
+        raise RuntimeError(f"Generation timed out ({TIMEOUT // 60} min limit)")
 
-        try:
-            stdout, stderr = proc.communicate(timeout=TIMEOUT)
-        except subprocess.TimeoutExpired:
-            proc.kill()
-            proc.wait()
-            raise
+    elapsed = time.time() - t0
+    with _proc_lock:
+        _active_proc = None
 
-        elapsed = time.time() - t0
+    if proc.returncode != 0:
+        err = stderr.decode(errors="replace")[-500:] if stderr else "Unknown error"
+        if proc.returncode == -9:
+            raise RuntimeError("Out of memory (killed by OS). Try 512x512.")
+        raise RuntimeError(f"sd-cli failed (code {proc.returncode}): {err}")
 
-        with _proc_lock:
-            _active_proc = None
+    if not os.path.exists(output_path) or os.path.getsize(output_path) == 0:
+        raise RuntimeError("No output image generated")
 
-        if proc.returncode != 0:
-            err = stderr.decode(errors="replace")[-500:] if stderr else "Unknown error"
-            if proc.returncode == -9:
-                raise gr.Error("Out of memory (killed by OS). Try 512x512.")
-            raise gr.Error(f"sd-cli failed (code {proc.returncode}): {err}")
+    status = f"Generated in {elapsed:.1f}s ({w}x{h}, {STEPS} steps, cfg {cfg})"
+    print(f"[gen] {status}")
+    return output_path, status
 
-        if not os.path.exists(output_path) or os.path.getsize(output_path) == 0:
-            raise gr.Error("No output image generated")
 
-        img = Image.open(output_path)
-        status = f"Generated in {elapsed:.1f}s ({w}x{h}, {STEPS} steps)"
-        print(f"[gen] {status}")
-        return img, status
+# ---------------------------------------------------------------------------
+# CLI mode
+# ---------------------------------------------------------------------------
 
-    except subprocess.TimeoutExpired:
-        with _proc_lock:
-            _active_proc = None
-        raise gr.Error(f"Generation timed out ({TIMEOUT//60} min limit)")
-    except gr.Error:
-        with _proc_lock:
-            _active_proc = None
-        raise
+def cli_main():
+    parser = argparse.ArgumentParser(description="Z-Anime 6B Image Generation")
+    parser.add_argument("prompt", help="Text prompt for image generation")
+    parser.add_argument("-n", "--negative", default="lowres, bad anatomy, bad hands, text, error, worst quality, blurry",
+                        help="Negative prompt")
+    parser.add_argument("-r", "--resolution", default="512x512", choices=RESOLUTIONS)
+    parser.add_argument("-s", "--seed", type=int, default=-1, help="Random seed (-1=random)")
+    parser.add_argument("-c", "--cfg", type=float, default=1.0, help="CFG scale (1.0 recommended)")
+    parser.add_argument("-o", "--output", default=None, help="Output file path")
+    args = parser.parse_args()
+
+    if args.output is None:
+        args.output = f"z-anime_seed{args.seed}_cfg{args.cfg}.png"
+
+    try:
+        path, status = generate_image(
+            prompt=args.prompt,
+            negative_prompt=args.negative,
+            resolution=args.resolution,
+            seed=args.seed,
+            cfg=args.cfg,
+            output_path=args.output,
+        )
+        print(f"  Output: {path}")
     except Exception as e:
-        with _proc_lock:
-            _active_proc = None
-        raise gr.Error(f"Error: {e}")
+        print(f"ERROR: {e}", file=sys.stderr)
+        sys.exit(1)
+
 
 # ---------------------------------------------------------------------------
-# Gradio UI
+# Gradio GUI mode
 # ---------------------------------------------------------------------------
-with gr.Blocks(title="Z-Anime (CPU)") as demo:
-    gr.Markdown(
-        "**[Z-Anime 6B](https://huggingface.co/SeeSee21/Z-Anime)** S3-DiT Q5_0 GGUF "
-        "(distill 4-step) via [sd.cpp](https://github.com/leejet/stable-diffusion.cpp) | "
-        "~30 min at 512x512 on free CPU"
-    )
-    with gr.Row():
-        with gr.Column():
-            prompt_input = gr.Textbox(label="Prompt", lines=3,
-                placeholder="anime girl with silver hair, fantasy armor, dramatic lighting")
-            neg_input = gr.Textbox(label="Negative Prompt", lines=2,
-                value="lowres, bad anatomy, bad hands, text, error, worst quality, blurry")
-            with gr.Row():
-                res_input = gr.Dropdown(choices=RESOLUTIONS, value="512x512",
-                    label="Resolution")
-                seed_input = gr.Number(value=-1, label="Seed (-1=random)", precision=0)
-            gen_btn = gr.Button("Generate (4 steps, CFG 1)", variant="primary", size="lg")
-        with gr.Column():
-            output_img = gr.Image(type="pil", label="Output")
-            status_box = gr.Textbox(label="Status", interactive=False)
-
-    gen_btn.click(fn=generate,
-        inputs=[prompt_input, neg_input, res_input, seed_input],
-        outputs=[output_img, status_box],
-        concurrency_limit=1,
-        api_name="generate")
-
-    def _on_unload():
-        with _proc_lock:
-            proc = _active_proc
-        if proc and proc.poll() is None:
-            print("[cleanup] User disconnected, killing sd-cli process")
-            proc.kill()
 
-    demo.unload(_on_unload)
+def gradio_main():
+    import mmap
+    from PIL import Image
+    import gradio as gr
+
+    # Warm up page cache
+    print("[init] Preloading models into page cache...")
+    t0 = time.time()
+    for model_path in [DIFFUSION, LLM, VAE]:
+        if os.path.exists(model_path):
+            sz = os.path.getsize(model_path)
+            with open(model_path, "rb") as f:
+                mm = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
+                mm.read()
+                mm.close()
+            print(f"  {os.path.basename(model_path)}: {sz / 1e9:.2f} GB cached")
+    print(f"[init] Page cache warm in {time.time() - t0:.1f}s")
+
+    def gui_generate(prompt, negative_prompt, resolution, cfg, seed):
+        try:
+            path, status = generate_image(prompt, negative_prompt, resolution,
+                                          int(seed or -1), cfg=float(cfg or 1.0))
+            return Image.open(path), status
+        except Exception as e:
+            raise gr.Error(str(e))
+
+    with gr.Blocks(title="Z-Anime (CPU)") as demo:
+        gr.Markdown(
+            "**[Z-Anime 6B](https://huggingface.co/SeeSee21/Z-Anime)** S3-DiT Q5_0 GGUF "
+            "(distill 4-step) via [sd.cpp](https://github.com/leejet/stable-diffusion.cpp) | "
+            "~30 min at 512x512 on free CPU"
+        )
+        with gr.Row():
+            with gr.Column():
+                prompt_input = gr.Textbox(label="Prompt", lines=3,
+                    placeholder="anime girl with silver hair, fantasy armor, dramatic lighting, beautiful background art, professional anime illustration quality, cinematic composition, detailed shading, high quality anime art.")
+                neg_input = gr.Textbox(label="Negative Prompt", lines=2,
+                    value="worst quality, low quality, lowres, blurry, bad anatomy, deformed hands, extra fingers, fused fingers, missing fingers, bad proportions, wrong proportions, extra limbs, broken limbs, duplicate body parts, asymmetrical eyes, distorted face, warped features, poorly drawn face, mutated, extra eyes, cropped head, cut-off body, bad framing, jpeg artifacts, compression artifacts, watermark, logo, signature, text, error, noisy, oversmoothed, muddy colors, background clutter, censored, 3d, chibi, character doll, sepia, high contrast")
+                with gr.Row():
+                    res_input = gr.Dropdown(choices=RESOLUTIONS, value="512x512", label="Resolution")
+                    cfg_input = gr.Slider(minimum=1.0, maximum=1.5, value=1.0, step=0.1, label="CFG (1.0 best, max 1.5)")
+                    seed_input = gr.Number(value=-1, label="Seed (-1=random)", precision=0)
+                gen_btn = gr.Button("Generate (4 steps)", variant="primary", size="lg")
+            with gr.Column():
+                output_img = gr.Image(type="pil", label="Output")
+                status_box = gr.Textbox(label="Status", interactive=False)
+
+        gen_btn.click(fn=gui_generate,
+            inputs=[prompt_input, neg_input, res_input, cfg_input, seed_input],
+            outputs=[output_img, status_box],
+            concurrency_limit=1,
+            api_name="generate")
+
+        gr.Examples(
+            examples=[
+                ["anime girl with silver hair, fantasy armor, dramatic lighting, beautiful background art, professional anime illustration quality, cinematic composition, detailed shading, high quality anime art."],
+            ],
+            inputs=[prompt_input],
+            cache_examples=True,
+            cache_mode="lazy",
+        )
+
+        def _on_unload():
+            with _proc_lock:
+                proc = _active_proc
+            if proc and proc.poll() is None:
+                print("[cleanup] User disconnected, killing sd-cli process")
+                proc.kill()
+
+        demo.unload(_on_unload)
+
+    demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True,
+                theme="NoCrypt/miku", mcp_server=True)
+
+
+# ---------------------------------------------------------------------------
+# Entry point
+# ---------------------------------------------------------------------------
 
-demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True, theme="NoCrypt/miku", mcp_server=True)
+if __name__ == "__main__":
+    if len(sys.argv) > 1 and sys.argv[1] != "--gradio":
+        cli_main()
+    else:
+        gradio_main()