Fix image-to-video enhancement: render the <|image|> token by passing the user turn as content blocks, matching LTX2ImageToVideoPipeline._enhance_prompt

app.py

@@ -57,9 +57,20 @@ def enhance(prompt, image=None, progress=gr.Progress()):
         raise gr.Error("Please enter a prompt to enhance.")
     # I2V (reference-image) system prompt when an image is supplied, else T2V.
     system_prompt = LTX2_4_I2V_DEFAULT_SYSTEM_PROMPT if image is not None else LTX2_4_T2V_DEFAULT_SYSTEM_PROMPT
+    user_text = f"{CFG.user_prompt_prefix}: {prompt}"
+    # Gemma-4's processor validates that the rendered template carries exactly one <|image|> token per
+    # image passed. A plain-string user turn renders none, so supplying an image raised
+    #   ValueError: ... Found [0] <|image|> tokens and [1] images per sample
+    # on EVERY image-to-video enhancement, while text-only worked — which is why it went unnoticed.
+    # The two shapes below are exactly `LTX2Pipeline._enhance_prompt` (plain string, images=None) and
+    # `LTX2ImageToVideoPipeline._enhance_prompt` (content blocks, images=image), so this Space keeps
+    # returning what the pipelines' own enhancement returns. The text-only path is untouched.
+    user_content = (
+        [{"type": "image"}, {"type": "text", "text": user_text}] if image is not None else user_text
+    )
     messages = [
         {"role": "system", "content": system_prompt},
-        {"role": "user", "content": f"{CFG.user_prompt_prefix}: {prompt}"},
+        {"role": "user", "content": user_content},
     ]
     template = processor.tokenizer.apply_chat_template(
         messages, tokenize=False, add_generation_prompt=True