Update inpainting_module.py

inpainting_module.py

@@ -923,7 +923,7 @@ class InpaintingModule:
                 if is_spaces:
                     # On Spaces, use minimal preview steps
                     preview_steps = min(preview_steps, 8)
-                    logger.info(f"Spaces environment - using {preview_steps} preview steps")
+                    logger.debug(f"Spaces environment - using {preview_steps} preview steps")
 
                 preview_result = self._generate_inpaint(
                     image=image,
@@ -937,7 +937,7 @@ class InpaintingModule:
                     generator=generator
                 )
             else:
-                logger.info("Spaces environment - skipping preview to fit 300s limit")
+                logger.debug("Spaces environment - skipping preview to fit 300s limit")
 
             if preview_only:
                 generation_time = time.time() - start_time
@@ -968,44 +968,23 @@ class InpaintingModule:
 
             # Optimize for Hugging Face Spaces ZeroGPU (stateless, 300s hard limit)
             if is_spaces:
-                # ZeroGPU timing breakdown (stateless architecture):
-                # - First-time model loading: ~150-180s (download + CPU load + GPU transfer)
-                # - Subsequent GPU allocation: ~40-60s (CPU to GPU transfer only)
-                # - Inference: ~15s/step (slower on shared H200 vs dedicated L4)
-                # - Platform limit: 300s hard limit (Pro tier, cannot be overridden)
+                # ZeroGPU timing breakdown with model caching:
+                # - Model loading from cache: ~60s (cached models, CPU to GPU transfer)
+                # - Inference: ~15s/step (on shared H200)
+                # - Blending & overhead: ~20s
+                # - Platform limit: 300s hard limit (Pro tier)
                 #
-                # Strategy:
-                # 1. Skip preview stage to maximize quality budget
-                # 2. Detect if this is first load (just initialized pipeline)
-                # 3. Use fewer steps on first load, more steps on subsequent requests
-
-                # Check if this is likely the first request after model loading
-                # (models just loaded will have generation_time ≈ 0)
-                is_first_load = not hasattr(self, '_first_generation_done')
-
-                if is_first_load:
-                    # First request after Space startup or model reload
-                    # Time budget: 300s - 180s (loading) - 20s (overhead) = 100s
-                    # Safe steps: 100s / 15s = 6-7 steps
-                    spaces_max_steps = 7  # Conservative for first-time loading
-                    logger.warning("⚠️  First inpainting request - using reduced steps to avoid timeout")
-                    logger.warning(f"   Model loading + 7 steps ≈ 280s (safe)")
-                    logger.warning(f"   💡 Tip: Re-run for higher quality (11-12 steps, ~245s)")
-                    self._first_generation_done = True
-                else:
-                    # Subsequent requests (model already in CPU memory)
-                    # Time budget: 300s - 60s (CPU→GPU) - 20s (overhead) = 220s
-                    # Safe steps: 220s / 15s = 14 steps (use 11-12 for safety margin)
-                    spaces_max_steps = 11  # Balanced quality for subsequent requests
+                # Strategy with unified 12-step approach:
+                # - Skip preview completely (done above)
+                # - Use consistent 12 steps for quality and color accuracy
+                # - Time budget: 60s (load) + 180s (12 steps) + 20s (blend) = 260s
+                # - Safety margin: 40s buffer for platform variability
+
+                spaces_max_steps = 12  # Unified step count for all templates
 
                 if num_steps > spaces_max_steps:
-                    original_steps = num_steps
                     num_steps = spaces_max_steps
-                    logger.info(f"Spaces ZeroGPU - reduced steps: {original_steps} → {num_steps}")
-                    if is_first_load:
-                        logger.info(f"   (First load mode: prioritizing completion over quality)")
-                    else:
-                        logger.info(f"   (Subsequent request: ~{60 + num_steps * 15}s estimated)")
+                    logger.debug(f"Spaces deployment: using {num_steps} steps (optimized for 300s limit)")
 
             full_result = self._generate_inpaint(
                 image=image,