Create app.py

app.py

@@ -0,0 +1,94 @@
+import gradio as gr
+import torch
+import spaces
+from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
+from PIL import Image
+
+BASE_MODEL = "Qwen/Qwen2-VL-2B-Instruct"
+BFS_MODEL = "Alissonerdx/BFS-Best-Face-Swap"
+
+# ----------- FORCE CPU FOR ZERO-GPU -----------
+DEVICE = "cpu"
+DTYPE = torch.float32
+
+print("Using device:", DEVICE)
+
+# ----------- LOAD BASE MODEL -----------
+print("Loading Qwen2-VL base model (CPU, FP32)...")
+model = Qwen2VLForConditionalGeneration.from_pretrained(
+    BASE_MODEL,
+    dtype=DTYPE,
+    device_map=None
+)
+
+model.to(DEVICE)
+
+processor = AutoProcessor.from_pretrained(BASE_MODEL)
+
+# ----------- LOAD BFS ADAPTERS -----------
+ADAPTERS = {
+    "face_v1": f"{BFS_MODEL}/bfs_face_v1.bin",
+    "head_v1": f"{BFS_MODEL}/bfs_head_v1.bin",
+    "head_v2": f"{BFS_MODEL}/bfs_head_v2.bin",
+    "head_v3": f"{BFS_MODEL}/bfs_head_v3.bin",
+}
+
+adapter_states = {
+    k: torch.load(path, map_location=DEVICE)
+    for k, path in ADAPTERS.items()
+}
+
+@spaces.GPU
+def apply_adapter(model, adapter_state):
+    print("Applying adapter:", len(adapter_state))
+    model.load_state_dict(adapter_state, strict=False)
+
+
+def bfs_swap(img1, img2, version):
+    if img1 is None or img2 is None:
+        return None
+
+    apply_adapter(model, adapter_states[version])
+
+    prompt = "swap the face of img1 onto img2, realistic, high quality"
+
+    inputs = processor(
+        text=prompt,
+        images=[img1, img2],
+        return_tensors="pt"
+    ).to(DEVICE)
+
+    out = model.generate(
+        **inputs,
+        max_new_tokens=1,
+        do_sample=False
+    )
+
+    decoded = processor.decode(out[0], skip_special_tokens=True)
+
+    if isinstance(decoded, dict) and "image" in decoded:
+        return decoded["image"]
+
+    return None
+
+
+with gr.Blocks() as demo:
+    gr.Markdown("## BFS Best Face Swap (Qwen2-VL 2B — CPU Mode for ZeroGPU)")
+
+    img1 = gr.Image(type="pil", label="Image 1")
+    img2 = gr.Image(type="pil", label="Image 2")
+
+    version = gr.Radio(
+        ["face_v1", "head_v1", "head_v2", "head_v3"],
+        label="Choose Version"
+    )
+
+    result = gr.Image(label="Output")
+
+    gr.Button("Swap!").click(
+        bfs_swap,
+        inputs=[img1, img2, version],
+        outputs=result
+    )
+
+demo.launch()