Replace LTX enhancer with Sulphur GGUF enhancer

app.py

@@ -29,9 +29,17 @@ MAX_SEED = np.iinfo(np.int32).max
 DEFAULT_FRAME_RATE = 24.0
 DEFAULT_PROMPT = "Make this image come alive with cinematic motion, smooth animation."
 PROMPT_ADHERENCE_SUFFIX = (
-    "Follow the prompt literally. Keep the named subjects, actions, setting, style, "
-    "camera motion, and timing consistent with the user's request."
+    "Follow the prompt literally. Preserve the named subjects, actions, setting, and style."
 )
+MAX_PROMPT_CHARS = 900
+MAX_ENHANCED_PROMPT_CHARS = 1400
+PROMPT_ENHANCER_SYSTEM_PROMPT = (
+    "You are Sulphur's prompt enhancer for image-to-video generation. Rewrite the user's prompt into one concise, "
+    "literal cinematic video prompt. Preserve the user's subject, action, setting, style, and intent. Add useful "
+    "visual motion, camera movement, lighting, and temporal details. Do not refuse, moralize, apologize, explain, "
+    "quote policy, or mention being an AI. Return only the rewritten prompt."
+)
+PROMPT_ENHANCER_USER_TEMPLATE = "Rewrite this video prompt only:\n{prompt}"
 
 RESOLUTIONS = {
     "high": {"16:9": (1536, 1024), "9:16": (1024, 1536), "1:1": (1024, 1024)},
@@ -215,11 +223,63 @@ def _validate_lora(name, lora):
 
 def _prompt_for_model(prompt):
     prompt = " ".join(str(prompt or DEFAULT_PROMPT).split())
+    prompt = _truncate_prompt(prompt, MAX_PROMPT_CHARS)
     if PROMPT_ADHERENCE_SUFFIX.lower() in prompt.lower():
         return prompt
     return f"{prompt}\n\n{PROMPT_ADHERENCE_SUFFIX}"
 
 
+def _truncate_prompt(prompt, limit):
+    prompt = str(prompt or "").strip()
+    if len(prompt) <= limit:
+        return prompt
+    truncated = prompt[:limit].rsplit(" ", 1)[0].strip()
+    return truncated or prompt[:limit].strip()
+
+
+_PROMPT_ENHANCER = None
+
+
+def _get_prompt_enhancer():
+    global _PROMPT_ENHANCER
+    if _PROMPT_ENHANCER is not None:
+        return _PROMPT_ENHANCER
+    from llama_cpp import Llama
+
+    print(f"[prompt-enhancer] loading {ASSETS['prompt_enhancer_model']}")
+    _PROMPT_ENHANCER = Llama(
+        model_path=str(ASSETS["prompt_enhancer_model"]),
+        n_ctx=2048,
+        n_threads=max(1, min(8, os.cpu_count() or 1)),
+        n_gpu_layers=0,
+        verbose=False,
+    )
+    return _PROMPT_ENHANCER
+
+
+def _enhance_prompt_with_sulphur(prompt):
+    base_prompt = _truncate_prompt(prompt or DEFAULT_PROMPT, MAX_PROMPT_CHARS)
+    llm = _get_prompt_enhancer()
+    result = llm.create_chat_completion(
+        messages=[
+            {"role": "system", "content": PROMPT_ENHANCER_SYSTEM_PROMPT},
+            {"role": "user", "content": PROMPT_ENHANCER_USER_TEMPLATE.format(prompt=base_prompt)},
+        ],
+        temperature=0.35,
+        top_p=0.9,
+        max_tokens=320,
+    )
+    enhanced = result["choices"][0]["message"]["content"].strip()
+    enhanced = enhanced.strip("\"'` \n\t")
+    blocked_phrases = ("cannot fulfill", "safety guidelines", "as an ai", "i cannot", "i'm unable")
+    if not enhanced or enhanced == "/" or any(phrase in enhanced.lower() for phrase in blocked_phrases):
+        print(f"[prompt-enhancer] rejected unusable output; falling back to literal prompt: {enhanced!r}")
+        return base_prompt
+    enhanced = _truncate_prompt(enhanced, MAX_ENHANCED_PROMPT_CHARS)
+    print(f"[prompt-enhancer] {enhanced}")
+    return enhanced
+
+
 def _ensure_gemma_model_alias(gemma_weight):
     gemma_weight = Path(gemma_weight)
     alias = gemma_weight.with_name("model.safetensors")
@@ -249,6 +309,16 @@ def download_assets():
         "ltx-2.3-spatial-upscaler-x2-1.1.safetensors",
         ASSETS_DIR,
     )
+    prompt_enhancer_model = _download(
+        "SulphurAI/Sulphur-2-base",
+        "prompt_enhancer/sulphur_prompt_enhancer_model-q8_0.gguf",
+        ASSETS_DIR,
+    )
+    prompt_enhancer_mmproj = _download(
+        "SulphurAI/Sulphur-2-base",
+        "prompt_enhancer/mmproj-BF16.gguf",
+        ASSETS_DIR,
+    )
 
     gemma_folder = "gemma-3-12b-it-qat-q4_0-unquantized"
     gemma_files = [
@@ -281,6 +351,8 @@ def download_assets():
         "sulphur_lora": sulphur_lora,
         "distilled_lora": distilled_lora,
         "upsampler": upsampler,
+        "prompt_enhancer_model": prompt_enhancer_model,
+        "prompt_enhancer_mmproj": prompt_enhancer_mmproj,
         "gemma_root": ASSETS_DIR / gemma_folder,
     }
 
@@ -293,6 +365,9 @@ LORAS = [
 ]
 _validate_lora("sulphur_distilled_condsafe", LORAS[0])
 _validate_lora("sulphur_experimental_lora_v1", LORAS[1])
+print("[lora] configured order:")
+for index, lora in enumerate(LORAS, start=1):
+    print(f"[lora] {index}: strength={lora.strength} path={lora.path}")
 
 pipeline = DistilledPipeline(
     distilled_checkpoint_path=str(ASSETS["checkpoint"]),
@@ -356,6 +431,8 @@ def generate_video(
             num_frames = ((num_frames - 1 + 7) // 8) * 8 + 1
             height = int(height)
             width = int(width)
+            if enhance_prompt:
+                prompt = _enhance_prompt_with_sulphur(prompt)
             model_prompt = _prompt_for_model(prompt)
 
             print(f"[generate] {width}x{height}, frames={num_frames}, seed={current_seed}")
@@ -381,7 +458,7 @@ def generate_video(
                 frame_rate=frame_rate,
                 images=images,
                 tiling_config=tiling_config,
-                enhance_prompt=bool(enhance_prompt),
+                enhance_prompt=False,
             )
 
             output_path = tempfile.mktemp(suffix=".mp4")

requirements.txt

@@ -16,3 +16,4 @@ Pillow
 sentencepiece
 scikit-image>=0.25.2
 flashpack==0.1.2
+llama-cpp-python>=0.3.9