Framing+gaze fix: detect full-body from original Thai (translator drops it) + inject English tag, canvas 896; swap looking-at-camera->looking at viewer (#34)

- Framing+gaze fix: detect full-body from original Thai (translator drops it) + inject English tag, canvas 896; swap looking-at-camera->looking at viewer (32054e82f50e930a23cadd7623a658bb06834834)


Co-authored-by: pormungtailaw <pormungtai@users.noreply.huggingface.co>

app.py

@@ -94,6 +94,21 @@ def generate(model_id, mode, prompt, negative_prompt, ref_image,
     orig_prompt = prompt
     prompt = pm.translate_prompt(prompt, translator)
     negative_prompt = pm.translate_prompt(negative_prompt, translator)
+
+    # --- Post-translation fixes (the translator often drops/softens these) ---
+    # 1) Full-body framing: detect intent from the ORIGINAL Thai (translator tends to
+    #    drop "เต็มตัว/ทั้งตัว"); inject an explicit English tag up front so the model AND
+    #    the auto-tall-canvas logic in run_generation reliably honor it.
+    if pm.wants_full_body(orig_prompt) or pm.wants_full_body(prompt):
+        if "full body" not in prompt.lower():
+            prompt = "full body shot, head to toe, full body visible, " + prompt
+    # 2) Gaze: SD1.5 obeys the booru tag "looking at viewer" far better than "...camera".
+    if "looking at the camera" in prompt.lower() or "looking at camera" in prompt.lower():
+        prompt = (prompt.replace("looking at the camera", "looking at viewer, eye contact")
+                        .replace("looking at camera", "looking at viewer, eye contact"))
+    elif "มองกล้อง" in (orig_prompt or "") and "looking at viewer" not in prompt.lower():
+        prompt = "looking at viewer, eye contact, " + prompt
+
     if prompt != orig_prompt:
         note = f"  ·  🌐 {translator}: _{prompt[:120]}_"
 

pipeline_manager.py

@@ -491,7 +491,7 @@ def run_generation(cfg, mode, prompt, negative_prompt, ref_image,
     # Full-body framing fix: on SD1.5 a 512x768 canvas crops standing/seated subjects
     # to a portrait even when "full body" is requested. Give it more vertical room.
     if base == "sd15" and mode == "txt2img" and wants_full_body(prompt):
-        height = max(int(height), 832)
+        height = max(int(height), 896)
 
     call = dict(
         prompt=full_prompt,