Reject scene-contaminated transparent assets

README.md

@@ -67,7 +67,7 @@ If a requested asset has no deterministic hook, images are still generated for r
 
 ## Contract-driven generation and validation
 
-Filenames and IDs do not decide whether an image is a scene or sprite. Dimensions, transparency, composition, camera, quantity, subject count, and silhouette requirements come from the manifest. Transparent single-subject outputs receive alpha extraction, occupancy normalization, and subject-count checks. Opaque outputs are checked for unwanted transparency. Seamless requests receive an opposite-edge similarity check. Other compositions are not incorrectly rejected for containing multiple subjects.
+Filenames and IDs do not decide whether an image is a scene or sprite. Dimensions, transparency, composition, camera, quantity, subject count, and silhouette requirements come from the manifest. Transparent single-subject outputs receive alpha extraction, occupancy normalization, subject-count checks, and a retained-scene check that rejects environmental foreground spread around an isolated subject. Opaque outputs are checked for unwanted transparency. Seamless requests receive an opposite-edge similarity check. Other compositions are not incorrectly rejected for containing multiple subjects.
 
 The image model can still violate these instructions, so every result requires visual review. Single-subject outputs that fail automated checks are regenerated with a fresh seed rather than cropped into a guessed subject. The review UI can regenerate one output or a complete multi-output asset set.
 
@@ -89,7 +89,7 @@ segmind/SSD-1B + latent-consistency/lcm-lora-ssd-1b
 
 SSD-1B is an Apache-2.0 distilled SDXL model. The LCM adapter enables low-step generation. The weights run inside the Space's ZeroGPU allocation, so the app does not consume Hugging Face Inference Provider credits and does not require a paid image API.
 
-Sprites and backgrounds are generated directly from their written prompts. No procedurally drawn sprite, palette, silhouette, camera guide, or background layout is passed into the production model. Sprite prompts use a portrait canvas and positive single-subject composition language; the negative prompt separately rejects sheets, lineups, repeated subjects, and multiple views. After transparency extraction, sprites with zero or multiple significant foreground components are regenerated with a new seed up to `PRIMARY_SPRITE_ATTEMPTS` times. If every attempt fails, the deployed Space returns an explicit model-generation error rather than returning a contact sheet. Background negative prompts reject characters and generic particle overlays.
+Sprites and backgrounds are generated directly from their written prompts. No procedurally drawn sprite, palette, silhouette, camera guide, or background layout is passed into the production model. Sprite prompts use a portrait canvas and positive single-subject composition language; the negative prompt separately rejects sheets, lineups, repeated subjects, multiple views, and surrounding environments. After transparency extraction, sprites with zero or multiple significant foreground components, retained scene-like alpha, implausible silhouettes, or inadequate scale are regenerated with a new seed up to `PRIMARY_SPRITE_ATTEMPTS` times. If every attempt fails, the deployed Space returns an explicit model-generation error rather than returning a contact sheet or environment-contaminated cutout. Background negative prompts reject characters and generic particle overlays.
 
 SSD-1B's CLIP text encoder has a 77-token context window. Production prompts therefore use a conservative word budget, place the camera and single-subject contract first, and compact the free-form description. This prevents those critical constraints from being silently truncated behind verbose style text.
 

app.py

@@ -1851,7 +1851,8 @@ def diffusion_negative_prompt(spec: AssetSpec) -> str:
         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, "
-            "room, ground, floor, background props, drop shadow, text, watermark"
+            "environment, environmental framing, scenic background, surrounding landscape, ground plane, floor plane, "
+            "background border, background props, drop shadow, text, watermark"
             + camera_negative
         )
     if spec.composition == "seamless":
@@ -1954,11 +1955,13 @@ def primary_diffusion_prompt(spec: AssetSpec) -> str:
         camera = f"User-defined camera: {compact_prompt_words(spec.camera_instruction, 10)}."
     variation = f"Distinct variation: {compact_prompt_words(spec.variation, 8)}. " if spec.variation else ""
     if spec.composition == "single_subject":
-        description = compact_prompt_words(spec.prompt, 12 if variation else 18)
+        description = compact_prompt_words(spec.prompt, 6 if variation else 10)
+        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. "
-            f"Fill about eighty percent of the frame with modest even margins on a uniform white field. "
+            "Fill about eighty percent of the frame with modest even margins. "
+            "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)
@@ -2024,6 +2027,9 @@ def primary_diffusion_png(spec: AssetSpec, index: int, run_id: int) -> tuple[byt
                 else:
                     last_failure_detail = "the last output did not contain one significant foreground subject"
                 continue
+            if has_scene_like_foreground_frame(content):
+                last_failure_detail = "the last output retained a surrounding scene instead of an isolated subject"
+                continue
             if has_implausibly_thin_foreground_subject(content, spec):
                 last_failure_detail = "the last output contained an implausibly thin foreground silhouette"
                 continue
@@ -2339,6 +2345,46 @@ def has_undersized_foreground_subject(content: bytes) -> bool:
     return max(width_ratio, height_ratio) < 0.68
 
 
+def has_scene_like_foreground_frame(content: bytes) -> bool:
+    """Detect retained scenery spread around a nominally isolated transparent subject.
+
+    Alpha extraction can turn a model-generated environment into one connected
+    foreground component, so subject counting alone cannot identify it. A clean
+    cutout may reach one or two sides of its canvas, but retained scenery fills
+    nearly every outer region around the center. The conservative 3x3 test avoids
+    rejecting ordinary tall, wide, or circular silhouettes.
+    """
+    image = Image.open(io.BytesIO(content)).convert("RGBA")
+    alpha = image.getchannel("A")
+    width, height = image.size
+    if width < 9 or height < 9:
+        return False
+
+    occupied_regions = 0
+    total_foreground = 0
+    perimeter_regions = 0
+    for row in range(3):
+        y0 = round(row * height / 3)
+        y1 = round((row + 1) * height / 3)
+        for column in range(3):
+            x0 = round(column * width / 3)
+            x1 = round((column + 1) * width / 3)
+            area = max(1, (x1 - x0) * (y1 - y0))
+            foreground = sum(
+                alpha.getpixel((x, y)) >= 64
+                for y in range(y0, y1)
+                for x in range(x0, x1)
+            )
+            total_foreground += foreground
+            if row != 1 or column != 1:
+                perimeter_regions += 1
+                if foreground / area >= 0.28:
+                    occupied_regions += 1
+
+    foreground_ratio = total_foreground / max(1, width * height)
+    return perimeter_regions == 8 and foreground_ratio >= 0.38 and occupied_regions >= 7
+
+
 def free_diffusion_png(spec: AssetSpec, index: int, run_id: int) -> tuple[bytes | None, str | None]:
     global FREE_DIFFUSION_PIPE, FREE_DIFFUSION_ERROR
     if FREE_DIFFUSION_ERROR:
@@ -2589,6 +2635,8 @@ def validate_asset_png(content: bytes, spec: AssetSpec) -> list[str]:
         )
         if any(value > 16 for value in corners):
             warnings.append("transparent output has opaque corner pixels")
+        if spec.composition in {"single_subject", "icon", "animation_frame"} and has_scene_like_foreground_frame(content):
+            warnings.append("transparent cutout retains a surrounding scene instead of an isolated subject")
         if spec.expected_subjects is not None:
             subject_count = estimated_foreground_subject_count(content)
             if subject_count != spec.expected_subjects:
@@ -2763,7 +2811,7 @@ def model_configuration_summary() -> str:
         "Every production image is generated directly from its written prompt by the primary text-to-image model. "
         "The user-defined composition and camera contract is placed first inside a conservative CLIP prompt budget. "
         "Transparent single-subject foregrounds are normalized to the requested canvas, then pass configured subject-count, "
-        "silhouette, and scale validation; failures are regenerated with a new seed up to "
+        "retained-scene, silhouette, and scale validation; failures are regenerated with a new seed up to "
         f"{PRIMARY_SPRITE_ATTEMPTS} times. No procedural guide is supplied to the model. The procedural renderer is "
         "development-only and is blocked in the deployed Space when the primary model fails."
     )