Upload app.py with huggingface_hub

app.py

@@ -366,81 +366,64 @@ def update_dimensions_on_upload(image, target_long=1024):
     return max(16, (nw // 16) * 16), max(16, (nh // 16) * 16)
 
 
-# duration=120 gives headroom so high-res (1536-2048) runs don't hit the default
-# ~60s ZeroGPU wall limit ("GPU task aborted"). Billing is on ACTUAL seconds, so a
-# fast 1024 run still costs little; this only raises the admission reservation.
-@spaces.GPU(size="xlarge", duration=120)
-def infer(
-    images_b64_json,
-    prompt,
-    lora_adapter,
-    seed,
-    randomize_seed,
-    guidance_scale,
-    steps,
-    preserve_identity=True,
-    identity_strength=65,
-    output_size=1280,
-    fix_hands=False,
-    progress=gr.Progress(track_tqdm=True),
-):
+def _run_edit(pil_images, prompt, lora_adapter, seed, randomize_seed,
+              guidance_scale, steps, preserve_identity, identity_strength,
+              output_size, fix_hands):
+    """Core image edit shared by the UI (`infer`) and the API (`api_predict`).
+    Runs inside a @spaces.GPU entry. Returns (PIL result image, seed_used)."""
     gc.collect()
     torch.cuda.empty_cache()
 
-    pil_images = b64_to_pil_list(images_b64_json)
+    pil_images = [im for im in (pil_images or []) if im is not None]
     if not pil_images:
-        raise gr.Error("Please upload at least one image to edit.")
-    if not prompt or prompt.strip() == "":
+        raise gr.Error("Please provide at least one image to edit.")
+    if not prompt or str(prompt).strip() == "":
         raise gr.Error("Please enter an edit prompt.")
 
     # Identity lock (pose-friendly): append a strength-scaled "same person, pose may
-    # change" instruction so the model re-poses the head while preserving the
-    # face/identity. No masking/compositing — the head is edited so it can follow the
-    # new pose. `preserve_identity` is the master on/off (uncheck for anime/3D);
-    # `identity_strength` (0-100) sets how forcefully identity is locked. Applies on
-    # top of ANY selected LoRA.
+    # change" instruction. `preserve_identity` master on/off; `identity_strength`
+    # (0-100). Applies on top of ANY selected LoRA.
     eff_strength = int(identity_strength) if bool(preserve_identity) else 0
-    prompt_used = _identity_prompt(prompt, eff_strength)
-    # Anatomy lock + realism/HD boost on person edits (positive prompt — works even
-    # at Guidance=1).
+    prompt_used = _identity_prompt(str(prompt), eff_strength)
+    # Anatomy lock + realism/HD boost on person edits (positive prompt, works at CFG=1).
     if eff_strength > 0:
         prompt_used = prompt_used + ANATOMY_POS + REALISM_POS
 
-    spec = ADAPTER_SPECS.get(lora_adapter)
-    if not spec:
-        raise gr.Error(f"Configuration not found for: {lora_adapter}")
-
-    adapter_name = spec["adapter_name"]
-    if adapter_name not in LOADED_ADAPTERS:
-        print(f"--- Downloading and Loading Adapter: {lora_adapter} ---")
-        try:
-            pipe.load_lora_weights(spec["repo"], weight_name=spec["weights"], adapter_name=adapter_name)
-            LOADED_ADAPTERS.add(adapter_name)
-        except Exception as e:
-            raise gr.Error(f"Failed to load adapter {lora_adapter}: {e}")
+    # Optional style LoRA. Empty / unknown adapter -> plain base-model edit.
+    spec = ADAPTER_SPECS.get(lora_adapter) if lora_adapter else None
+    if spec:
+        adapter_name = spec["adapter_name"]
+        if adapter_name not in LOADED_ADAPTERS:
+            print(f"--- Downloading and Loading Adapter: {lora_adapter} ---")
+            try:
+                pipe.load_lora_weights(spec["repo"], weight_name=spec["weights"], adapter_name=adapter_name)
+                LOADED_ADAPTERS.add(adapter_name)
+            except Exception as e:
+                raise gr.Error(f"Failed to load adapter {lora_adapter}: {e}")
+        pipe.set_adapters([adapter_name], adapter_weights=[1.0])
     else:
-        print(f"--- Adapter {lora_adapter} already loaded. ---")
-
-    pipe.set_adapters([adapter_name], adapter_weights=[1.0])
+        try:
+            pipe.disable_lora()   # no style requested — clear any adapter from a prior call
+        except Exception:
+            pass
 
-    # "Fix hands": the anti-extra-finger terms live in the NEGATIVE prompt, which is
-    # ignored at Guidance=1 (CFG off on this distilled model). Raising Guidance turns
-    # true-CFG on so those negatives actually suppress extra/fused fingers; a couple
-    # more steps give CFG room to work. Costs ~2x time — hence it's opt-in.
+    # "Fix hands": the anti-extra-finger terms live in the NEGATIVE prompt, ignored at
+    # Guidance=1 (CFG off on this distilled model). Raising Guidance turns true-CFG on
+    # so those negatives suppress extra/fused fingers; +steps give CFG room. ~2x time.
     if bool(fix_hands):
         guidance_scale = max(float(guidance_scale), float(os.getenv("FIXHANDS_CFG", "2.5")))
         steps = max(int(steps), int(os.getenv("FIXHANDS_STEPS", "6")))
 
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
+    seed = int(seed)
 
     generator = torch.Generator(device=device).manual_seed(seed)
     negative_prompt = (
         "worst quality, low quality, bad anatomy, bad hands, text, error, missing fingers, "
         "extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry"
     )
-    # Identity + anatomy negatives (active when Guidance > 1). Reinforce face lock
-    # and suppress extra hands / >5 fingers / duplicated limbs.
+    # Identity + anatomy negatives (active when Guidance > 1).
     if eff_strength > 0:
         negative_prompt = negative_prompt + ", " + IDENTITY_NEG + ", " + ANATOMY_NEG
     width, height = update_dimensions_on_upload(pil_images[0], output_size)
@@ -452,18 +435,78 @@ def infer(
             negative_prompt=negative_prompt,
             height=height,
             width=width,
-            num_inference_steps=steps,
+            num_inference_steps=int(steps),
             generator=generator,
-            true_cfg_scale=guidance_scale,
+            true_cfg_scale=float(guidance_scale),
         ).images[0]
         return result_image, seed
-    except Exception as e:
-        raise e
     finally:
         gc.collect()
         torch.cuda.empty_cache()
 
 
+# duration=120 gives headroom so high-res (1536-2048) runs don't hit the default
+# ~60s ZeroGPU wall limit ("GPU task aborted"). Billing is on ACTUAL seconds, so a
+# fast 1024 run still costs little; this only raises the admission reservation.
+@spaces.GPU(size="xlarge", duration=120)
+def infer(
+    images_b64_json,
+    prompt,
+    lora_adapter,
+    seed,
+    randomize_seed,
+    guidance_scale,
+    steps,
+    preserve_identity=True,
+    identity_strength=65,
+    output_size=1280,
+    fix_hands=False,
+    progress=gr.Progress(track_tqdm=True),
+):
+    pil_images = b64_to_pil_list(images_b64_json)
+    return _run_edit(pil_images, prompt, lora_adapter, seed, randomize_seed,
+                     guidance_scale, steps, preserve_identity, identity_strength,
+                     output_size, fix_hands)
+
+
+def _as_pil(x):
+    if x is None:
+        return None
+    if isinstance(x, Image.Image):
+        return x.convert("RGB")
+    try:
+        return Image.fromarray(x).convert("RGB")
+    except Exception:
+        return None
+
+
+@spaces.GPU(size="xlarge", duration=120)
+def api_predict(image, prompt, lora_adapter="", image2=None, seed=0,
+                randomize_seed=True, guidance_scale=1.0, steps=4,
+                preserve_identity=True, identity_strength=65, output_size=1280,
+                fix_hands=False, progress=gr.Progress(track_tqdm=True)):
+    """Edit an image from a text prompt (Qwen-Image-Edit-2511). Returns [result_image, seed_used].
+
+    image           : the image to edit (the person/subject whose identity is kept).
+    prompt          : what to change, e.g. "change her outfit to a red dress; turn to look left".
+    lora_adapter    : optional style name from the adapter list ("" = no style LoRA / base edit).
+    image2          : optional second reference image (e.g. a style or lighting reference).
+    seed            : int seed (ignored if randomize_seed=True).
+    randomize_seed  : pick a new random seed each call.
+    guidance_scale  : true-CFG. 1.0 = fastest; >1 (e.g. 2.5) activates the negative prompt
+                      that suppresses extra fingers / identity drift (slower).
+    steps           : denoise steps. 4 is native for this distilled model.
+    preserve_identity : keep the person's face while allowing pose/head to change.
+    identity_strength : 0-100, how forcefully identity is locked.
+    output_size     : long-side pixels 1024-2048. Higher = more HD detail, slower.
+    fix_hands       : raise CFG/steps to suppress extra/fused fingers (slower).
+    """
+    imgs = [im for im in (_as_pil(image), _as_pil(image2)) if im is not None]
+    return _run_edit(imgs, prompt, lora_adapter, seed, randomize_seed,
+                     guidance_scale, steps, preserve_identity, identity_strength,
+                     output_size, fix_hands)
+
+
 css = r"""
 @import url('https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;600;700;800&family=JetBrains+Mono:wght@400;500;600&display=swap');
 *{box-sizing:border-box;margin:0;padding:0}
@@ -1636,6 +1679,33 @@ with gr.Blocks() as demo:
 
     run_btn = gr.Button("Run", elem_id="gradio-run-btn")
 
+    # ── Documented public API endpoint: /edit ──────────────────────────────────
+    # Typed inputs so API/MCP callers pass a real image + prompt (not the UI's
+    # internal base64 blob). Hidden from the UI; fully functional over the API.
+    with gr.Row(visible=False):
+        api_image   = gr.Image(type="pil", label="image")
+        api_image2  = gr.Image(type="pil", label="image2")
+        api_prompt  = gr.Textbox(label="prompt")
+        api_lora    = gr.Textbox(label="lora_adapter", value="")
+        api_seed    = gr.Number(label="seed", value=0, precision=0)
+        api_rand    = gr.Checkbox(label="randomize_seed", value=True)
+        api_cfg     = gr.Number(label="guidance_scale", value=1.0)
+        api_steps   = gr.Number(label="steps", value=4, precision=0)
+        api_pres    = gr.Checkbox(label="preserve_identity", value=True)
+        api_idst    = gr.Number(label="identity_strength", value=65, precision=0)
+        api_size    = gr.Number(label="output_size", value=1280, precision=0)
+        api_fix     = gr.Checkbox(label="fix_hands", value=False)
+        api_out     = gr.Image(label="result_image", type="pil", format="png")
+        api_seed_o  = gr.Number(label="seed_used", precision=0)
+        api_btn     = gr.Button("api_edit")
+    api_btn.click(
+        fn=api_predict,
+        inputs=[api_image, api_prompt, api_lora, api_image2, api_seed, api_rand,
+                api_cfg, api_steps, api_pres, api_idst, api_size, api_fix],
+        outputs=[api_out, api_seed_o],
+        api_name="edit",
+    )
+
     demo.load(fn=None, js=gallery_js)
     demo.load(fn=None, js=wire_outputs_js)
 
@@ -1661,6 +1731,7 @@ with gr.Blocks() as demo:
             const fhVal     = fhEl     ? fhEl.checked : fh;
             return [imgsJson, promptVal, loraVal, s, rs, gs, st, presVal, idstVal, oszVal, fhVal];
         }""",
+        show_api=False,   # internal UI path; the public API is /edit (api_predict)
     )
 
     example_load_btn.click(
@@ -1668,6 +1739,7 @@ with gr.Blocks() as demo:
         inputs=[example_idx],
         outputs=[example_result],
         queue=False,
+        show_api=False,
     )
 
 if __name__ == "__main__":