Improve generated image sharpness

app.py

@@ -80,8 +80,8 @@ FOREGROUND_MODEL_REVISION = os.environ.get(
     "FOREGROUND_MODEL_REVISION",
     "e2bf8e4460fc8fa32bba5ea4d94b3233d367b0e4",
 )
-PRIMARY_IMAGE_STEPS = int(os.environ.get("PRIMARY_IMAGE_STEPS", "4"))
-PRIMARY_SPRITE_STEPS = int(os.environ.get("PRIMARY_SPRITE_STEPS", "6"))
+PRIMARY_IMAGE_STEPS = int(os.environ.get("PRIMARY_IMAGE_STEPS", "6"))
+PRIMARY_SPRITE_STEPS = int(os.environ.get("PRIMARY_SPRITE_STEPS", "8"))
 PRIMARY_SPRITE_ATTEMPTS = max(1, int(os.environ.get("PRIMARY_SPRITE_ATTEMPTS", "3")))
 PRIMARY_GUIDANCE_SCALE = float(os.environ.get("PRIMARY_GUIDANCE_SCALE", "1.5"))
 USE_PRIMARY_IMAGE_MODEL = os.environ.get("USE_PRIMARY_IMAGE_MODEL", "1") == "1"
@@ -1764,17 +1764,17 @@ def diffusion_negative_prompt(spec: AssetSpec) -> str:
     }.get(spec.camera, "")
     if spec.composition in {"single_subject", "icon", "animation_frame"}:
         return (
-            "multiple subjects, duplicate character, duplicate subject, repeated subject, character sheet, model sheet, turnaround, "
-            "lineup, alternate views, multiple poses, multiple views, pair, group, cropped subject, scenery, landscape, "
-            "environment, environmental framing, scenic background, surrounding landscape, ground plane, floor plane, "
-            "background border, background props, drop shadow, text, watermark"
+            "multiple subjects, duplicate character, repeated subject, character sheet, turnaround, lineup, alternate views, "
+            "multiple poses, cropped subject, scenery, landscape, environment, environmental framing, ground plane, "
+            "background props, drop shadow, text, watermark, blurry, soft focus, out of focus, low detail, smeared details, "
+            "compression artifacts"
             + camera_negative
         )
     if spec.composition == "seamless":
-        return "visible seams, borders, frame, perspective mockup, text, watermark" + camera_negative
+        return "visible seams, borders, frame, perspective mockup, text, watermark, blurry, soft focus, low detail" + camera_negative
     if spec.composition == "sprite_sheet":
-        return "irregular grid, overlapping cells, labels, text, watermark" + camera_negative
-    return "text, labels, watermark, presentation mockup" + camera_negative
+        return "irregular grid, overlapping cells, labels, text, watermark, blurry, soft focus, low detail" + camera_negative
+    return "text, labels, watermark, presentation mockup, blurry, soft focus, out of focus, low detail" + camera_negative
 
 
 def build_camera_control_image(spec: AssetSpec, width: int = 1024, height: int = 576) -> Image.Image:
@@ -1924,19 +1924,20 @@ def primary_diffusion_prompt(spec: AssetSpec) -> str:
         variation = f"Variation: {compact_prompt_words(spec.variation, 4)}. " if spec.variation else ""
         subject = (spec.group or spec.role).replace("_", " ")
         return (
-            f"{camera} Single isolated {subject} subject. One complete body, one pose, centered and fully visible. "
-            "Fill about eighty percent of the frame with modest even margins. "
+            f"{camera} Single isolated {subject} subject. One complete centered body in one pose. "
+            "Fill eighty percent of the frame with even margins. "
+            "Crisp game-ready details. "
             "Plain uniform white field only; no environment, scenery, ground plane, or surrounding border. "
             f"{variation}{description}"
         )
     description = compact_prompt_words(spec.prompt, 26 if variation else 34)
     composition = {
-        "icon": "Exactly one centered readable game icon, uniform white field.",
-        "animation_frame": "Exactly one isolated animation frame subject at stable scale, uniform white field.",
-        "full_frame": "One complete edge-to-edge game image filling the canvas.",
-        "seamless": "One seamless edge-to-edge repeating game-art tile.",
-        "sprite_sheet": "One deliberately organized regular-grid sprite sheet.",
-        "freeform": "One user-defined game image following the requested composition exactly.",
+        "icon": "Exactly one centered readable game icon with crisp details, uniform white field.",
+        "animation_frame": "Exactly one isolated sharp animation frame subject at stable scale, uniform white field.",
+        "full_frame": "One complete edge-to-edge game image with crisp readable details filling the canvas.",
+        "seamless": "One sharp seamless edge-to-edge repeating game-art tile.",
+        "sprite_sheet": "One crisp deliberately organized regular-grid sprite sheet.",
+        "freeform": "One sharp user-defined game image following the requested composition exactly.",
     }[spec.composition]
     return f"{camera} {composition} {variation}{description}"
 
@@ -2131,12 +2132,16 @@ def polish_diffusion_asset(image: Image.Image, spec: AssetSpec) -> bytes:
 def polish_neural_cutout(image: Image.Image, spec: AssetSpec) -> bytes:
     """Resize and normalize a neural alpha cutout without changing its generated RGB art."""
     contained = image.convert("RGBA")
-    contained.thumbnail((spec.width, spec.height), Image.LANCZOS)
-    canvas = Image.new("RGBA", (spec.width, spec.height), (0, 0, 0, 0))
-    offset = ((spec.width - contained.width) // 2, (spec.height - contained.height) // 2)
-    canvas.alpha_composite(contained, dest=offset)
     if spec.composition in {"single_subject", "icon", "animation_frame"}:
-        canvas = normalize_sprite_foreground(canvas, spec)
+        # Crop the high-resolution model output first, then resize the subject
+        # once. The previous thumbnail-then-normalize path resampled RGB twice
+        # and visibly softened small game sprites.
+        canvas = normalize_sprite_foreground(contained, spec)
+    else:
+        contained.thumbnail((spec.width, spec.height), Image.LANCZOS)
+        canvas = Image.new("RGBA", (spec.width, spec.height), (0, 0, 0, 0))
+        offset = ((spec.width - contained.width) // 2, (spec.height - contained.height) // 2)
+        canvas.alpha_composite(contained, dest=offset)
     out = io.BytesIO()
     canvas.save(out, format="PNG")
     return out.getvalue()