Deploy InstantRetouch IP2P-BILA ZeroGPU Space

.gitattributes

@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/examples/4920_O_0_5_input.png filter=lfs diff=lfs merge=lfs -text
+assets/examples/4933_O_0_21_input.png filter=lfs diff=lfs merge=lfs -text
+assets/examples/expert116_input.png filter=lfs diff=lfs merge=lfs -text
+assets/examples/expert48_input.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -8,25 +8,20 @@ pinned: false
 license: other
 ---
 
-# InstantRetouch / BILA Space Demo
+# InstantRetouch / IP2P-BiLA Demo
 
-This Space is an isolated Hugging Face ZeroGPU/Gradio demo for direct image editing with two BILA backends:
+Public Hugging Face ZeroGPU demo for instruction-guided image retouching with the validation-selected IP2P/BiLA checkpoint.
 
-- `ip2p_bila_score1_8_104`: InstructPix2Pix base plus BILA checkpoint.
-- `flux_bila_score1_8_022`: Flux2 Klein base, task LoRA, and BILA checkpoint.
+- Model: IP2P/BiLA
+- UI: image upload, optional instruction, seed, max side, strength, and selectable examples
 
-The demo does not import or depend on the research repo's `agent/` path. It uses the validation-style direct flow:
-
-1. Load the base model.
-2. Load the selected checkpoint's `state_dict`.
-3. Generate `bila_output`.
-4. Apply the UI strength as `input + strength * (bila_output - input)`.
+This Space is isolated from the research repository. It does not import `agent/`, training scripts, or local experiment paths at runtime. Weights live in a separate Hugging Face model repo and are downloaded lazily through `BILA_WEIGHTS_REPO`.
 
 ## Required Space Variables
 
 Set one of these in the Space environment:
 
-- `BILA_WEIGHTS_REPO`: Hugging Face model repo containing the weight layout below.
+- `BILA_WEIGHTS_REPO`: Hugging Face model repo containing the IP2P weight layout below.
 - `BILA_MODEL_ROOT`: local path with the same layout, useful only for staging/debugging.
 
 Optional:
@@ -41,21 +36,14 @@ Do not commit weights into this Space repo. Put them in a separate HF model repo
 ```text
 ip2p/
   base/
+    tokenizer/
+    text_encoder/
+    vae/
+    unet/
   checkpoints/
-    epoch_5_bila_score1_8_104.pth
-flux/
-  base/
-  task_lora/
-    pytorch_lora_weights.safetensors
-  checkpoints/
-    epoch_8_bila_score1_8_022.pth
+    <bila-checkpoint>.pth
 metrics/
-  ip2p_bila_score1_8_104.json
-  flux_bila_score1_8_022.json
+  <metric-summary>.json
 ```
 
-The app lazily downloads only the selected model's allow-listed files, so it does not pull both large bases during cold start unless both models are used.
-
-## ZeroGPU Notes
-
-ZeroGPU requires the Gradio SDK and the `@spaces.GPU` decorator; this Space is configured that way. A `Dockerfile` is kept only as a fallback for standard paid GPU Spaces.
+The app follows the validation-style direct flow: load the IP2P base model, load the BiLA checkpoint named in `model_manifest.json`, generate `bila_output`, then apply strength as `input + strength * (bila_output - input)`.

app.py

@@ -52,50 +52,65 @@ from demo_runtime.manager import DemoManager
 
 
 manager = DemoManager()
+DEFAULT_MODEL = manager.default_model
+EXAMPLE_DIR = ROOT / "assets" / "examples"
+EXAMPLES = [
+    [str(EXAMPLE_DIR / "4920_O_0_5_input.png"), "Make the image feel more serene and add a subtle blue hue.", 42, 1024, 1.0],
+    [str(EXAMPLE_DIR / "4933_O_0_21_input.png"), "Improve the exposure and make the colors richer while keeping a natural photo look.", 7, 1024, 1.0],
+    [str(EXAMPLE_DIR / "expert48_input.png"), "Brighten the image and enhance clarity with balanced contrast.", 123, 1024, 0.9],
+    [str(EXAMPLE_DIR / "expert116_input.png"), "", 314, 1024, 1.0],
+]
 
 
 @spaces.GPU(duration=300, size="xlarge")
-def run_demo(image, instruction, model_key, seed, max_side, strength):
+def run_demo(image, instruction, seed, max_side, strength):
     try:
-        edited, diff, input_image, status = manager.generate(
+        edited, _diff, _input_image, status = manager.generate(
             image=image,
             instruction=instruction,
-            model_key=model_key,
+            model_key=DEFAULT_MODEL,
             seed=int(seed),
             max_side=int(max_side),
             strength=float(strength),
         )
-        comparison = [(input_image, "Input"), (diff, "Base output"), (edited, "BILA output")]
-        return edited, comparison, status
+        return edited, status
     except Exception as exc:
         raise gr.Error(str(exc))
 
 
 with gr.Blocks(title="InstantRetouch") as demo:
-    gr.Markdown("# InstantRetouch")
+    gr.Markdown(
+        """
+        # InstantRetouch
+        Instruction-guided photo retouching with the selected IP2P/BiLA checkpoint. Upload an image, enter an optional instruction, or click one of the examples below.
+
+        This public demo uses the validation-selected IP2P/BiLA model only. The strength slider blends the model output with the input for gentler or stronger edits.
+        """
+    )
     with gr.Row():
         with gr.Column(scale=1):
             image = gr.Image(type="pil", label="Input image")
-            instruction = gr.Textbox(label="Instruction", lines=3)
-            model_key = gr.Dropdown(
-                choices=manager.model_choices,
-                value=manager.default_model,
-                label="Model",
-            )
+            instruction = gr.Textbox(label="Instruction", lines=3, placeholder="Optional. Leave empty for prompt=\"\".")
             with gr.Row():
                 seed = gr.Number(value=42, precision=0, label="Seed")
                 max_side = gr.Slider(512, 2048, value=1024, step=64, label="Max side")
             strength = gr.Slider(0.0, 2.0, value=1.0, step=0.05, label="Strength")
             button = gr.Button("Run", variant="primary")
         with gr.Column(scale=1):
-            edited = gr.Image(type="pil", label="Edited")
+            edited = gr.Image(type="pil", label="Edited image")
             status = gr.Textbox(label="Status", interactive=False)
-    comparison = gr.Gallery(label="Comparison", columns=3, height="auto")
+
+    gr.Examples(
+        examples=EXAMPLES,
+        inputs=[image, instruction, seed, max_side, strength],
+        examples_per_page=4,
+        cache_examples=False,
+    )
 
     button.click(
         fn=run_demo,
-        inputs=[image, instruction, model_key, seed, max_side, strength],
-        outputs=[edited, comparison, status],
+        inputs=[image, instruction, seed, max_side, strength],
+        outputs=[edited, status],
     )
 
 

demo_runtime/manager.py

@@ -109,10 +109,6 @@ class DemoManager:
             edited_tensor = blend_strength(prepared.full_tensor, result["bila"], strength)
             edited = tensor_to_pil(edited_tensor)
             diff = tensor_to_pil(result["diff"])
-            evidence = model_cfg["evidence"]["scores_avg"]
-            status = (
-                f"{model_cfg['label']} | score_1={evidence['score_1']:.3f} | "
-                f"score_2={evidence['score_2']:.3f} | seed={int(seed)}"
-            )
+            status = f"{model_cfg['label']} | seed={int(seed)}"
             return edited, diff, prepared.full_pil, status
 

model_manifest.json

@@ -3,13 +3,16 @@
   "default_model": "ip2p_bila_score1_8_104",
   "models": {
     "ip2p_bila_score1_8_104": {
-      "label": "IP2P/BiLA (score_1 8.104)",
+      "label": "IP2P/BiLA",
       "kind": "ip2p",
       "weights": {
         "base": "ip2p/base",
         "checkpoint": "ip2p/checkpoints/epoch_5_bila_score1_8_104.pth"
       },
-      "expected_checkpoint_keys": ["state_dict.unet", "state_dict.bila"],
+      "expected_checkpoint_keys": [
+        "state_dict.unet",
+        "state_dict.bila"
+      ],
       "config": {
         "bila_grid_res": 32,
         "bila_grid_bins": 8,
@@ -28,47 +31,6 @@
           "score_2": 8.984478935698448
         }
       }
-    },
-    "flux_bila_score1_8_022": {
-      "label": "Flux/BiLA (score_1 8.022)",
-      "kind": "flux",
-      "weights": {
-        "base": "flux/base",
-        "task_lora": "flux/task_lora/pytorch_lora_weights.safetensors",
-        "checkpoint": "flux/checkpoints/epoch_8_bila_score1_8_022.pth"
-      },
-      "expected_checkpoint_keys": [
-        "state_dict.transformer_lora",
-        "state_dict.bila"
-      ],
-      "config": {
-        "bila_grid_res": 16,
-        "bila_grid_bins": 8,
-        "bila_use_flux_rgb": false,
-        "model_size": 512,
-        "cfg": false,
-        "use_t2i": false,
-        "not_scheduler_decode": true,
-        "mixed_precision": "bf16",
-        "max_sequence_length": 512,
-        "distill_strategy": "merge_then_new",
-        "distill_lora_rank": 32,
-        "distill_lora_alpha": 32,
-        "distill_lora_dropout": 0.0,
-        "task_lora_rank": 32,
-        "task_lora_alpha": 32
-      },
-      "evidence": {
-        "metric_file": "metrics/flux_bila_score1_8_022.json",
-        "source_run": "train-2026-03-11--00-29-image-all-1",
-        "model_filter": "bila",
-        "num_pairs": 89,
-        "scores_avg": {
-          "score_1": 8.02247191011236,
-          "score_2": 9.426966292134832
-        }
-      }
     }
   }
 }
-