Add info block below image: backend, loaded models, live GPU/VRAM stats, weighting note

app.py

@@ -164,6 +164,27 @@ model_management.load_models_gpu(
 # --------------------------------------------------------------------------
 import gradio as gr  # noqa: E402  # needed for gr.Progress default arg below
 
+BACKEND_INFO = (
+    "**Inference backend:** ComfyUI core (native Anima support, no custom nodes)\n"
+    "**Loaded models:** 4x Anima 2B checkpoints + Qwen-3 0.6B text encoder + Qwen-Image VAE (preloaded, ~18 GB)\n"
+    "**Prompt weighting:** supported, e.g. `(tag:1.2)`, `(tag)`, `[tag:0.8]` (per the Anima README, use higher weights than SDXL)"
+)
+
+
+# GPU stats are only real inside the ZeroGPU worker, so they are fetched at
+# the end of each generation rather than at startup.
+def gpu_stats() -> str:
+    try:
+        name = torch.cuda.get_device_name(0)
+        free, total = torch.cuda.mem_get_info(0)
+        used = total - free
+    except Exception as exc:
+        return f"**GPU:** unavailable ({exc})"
+    return (
+        f"**GPU:** {name}\n"
+        f"**VRAM:** {used / 1e9:.1f} GB used / {total / 1e9:.1f} GB total ({free / 1e9:.1f} GB free)"
+    )
+
 
 @spaces.GPU(duration=int(os.environ.get("GPU_DURATION", "30")))
 def generate_image(
@@ -181,7 +202,7 @@ def generate_image(
     lora_name: str,
     lora_strength: float,
     progress: gr.Progress = gr.Progress(track_tqdm=False),
-) -> np.ndarray:
+) -> tuple[np.ndarray, str]:
     width = max(512, min(1536, int(width) // 16 * 16))
     height = max(512, min(1536, int(height) // 16 * 16))
     seed = int(seed) if int(seed) >= 0 else random.randint(1, 2**63)
@@ -232,7 +253,7 @@ def generate_image(
     progress(1, desc="Decoding")
     decoded = vae_decode.decode(samples=sampled, vae=get_value_at_index(VAE, 0))
     image = get_value_at_index(decoded, 0)[0]
-    return (
+    img_np = (
         torch.nan_to_num(image, nan=0.0, posinf=1.0, neginf=0.0)
         .mul(255)
         .clamp_(0, 255)
@@ -240,6 +261,12 @@ def generate_image(
         .cpu()
         .numpy()
     )
+    info = (
+        BACKEND_INFO
+        + f"\n{gpu_stats()}"
+        + f"\n**Last run:** {model_name} | {width}x{height} | {steps} steps | CFG {cfg} | seed {seed}"
+    )
+    return img_np, info
 
 
 # --------------------------------------------------------------------------
@@ -261,6 +288,7 @@ SCHEDULERS = ["simple", "normal", "karras", "exponential", "sgm_uniform"]
 LORA_CHOICES = list_loras() or ["real_skin.safetensors"]
 
 output_image = gr.Image(label="Generated Image")
+info_output = gr.Markdown(BACKEND_INFO + "\n_Run a generation to see live GPU/VRAM stats._")
 
 with gr.Blocks(title="Anima Collection") as app:
     with gr.Row():
@@ -290,6 +318,7 @@ with gr.Blocks(title="Anima Collection") as app:
 
         with gr.Column(scale=1):
             output_image.render()
+            info_output.render()
 
     resolution.change(
         lambda name: list(RESOLUTIONS[name]),
@@ -303,7 +332,7 @@ with gr.Blocks(title="Anima Collection") as app:
             steps_input, cfg_input, sampler_input, scheduler_input, lora_enable, lora_name_input,
             lora_strength_input,
         ],
-        outputs=[output_image],
+        outputs=[output_image, info_output],
         show_progress=True,
     )