Update app.py

app.py

@@ -1,5 +1,5 @@
-# app.py — FLUX-only with temporal chaining + Aggressive follow + Video stitching
-import os, json, uuid, re
+# app.py — FLUX-only with temporal chaining + Aggressive follow + Video stitching (lazy MoviePy)
+import os, json, uuid, re, sys, subprocess
 from datetime import datetime
 import gradio as gr
 import spaces
@@ -7,11 +7,6 @@ import torch
 from PIL import Image
 import pandas as pd
 
-# MoviePy for stitching
-from moviepy.editor import ImageClip, CompositeVideoClip, concatenate_videoclips
-from moviepy.video.io.VideoFileClip import VideoFileClip
-
-
 # =========================
 # Storage helpers
 # =========================
@@ -47,15 +42,14 @@ def ensure_project(p, suggested_name="Project"):
     name = f"{suggested_name}-{pid[:4]}"
     proj = {
         "meta": {"id": pid, "name": name, "created": now_iso(), "updated": now_iso()},
-        "shots": [],    # each shot: id,title,description,duration,fps,steps,seed,negative,image_path
+        "shots": [],    # id,title,description,duration,fps,steps,seed,negative,image_path
         "clips": [],
     }
     save_project(proj)
     return proj
 
-
 # =========================
-# LLM (ZeroGPU) — Storyboard generator (robust)
+# LLM — Storyboard generator (ZeroGPU friendly)
 # =========================
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
@@ -71,37 +65,27 @@ def _lazy_model_tok():
         return _model, _tokenizer
 
     _tokenizer = AutoTokenizer.from_pretrained(STORYBOARD_MODEL, trust_remote_code=True)
-
     use_cuda = torch.cuda.is_available()
-    preferred_dtype = torch.float16 if use_cuda else torch.float32
-
+    dtype = torch.float16 if use_cuda else torch.float32
     _model = AutoModelForCausalLM.from_pretrained(
-        STORYBOARD_MODEL,
-        device_map="auto",
-        torch_dtype=preferred_dtype,
-        trust_remote_code=True,
-        use_safetensors=True
+        STORYBOARD_MODEL, device_map="auto", torch_dtype=dtype,
+        trust_remote_code=True, use_safetensors=True
     )
-
     if _tokenizer.pad_token_id is None and _tokenizer.eos_token_id is not None:
         _tokenizer.pad_token_id = _tokenizer.eos_token_id
-
     return _model, _tokenizer
 
 def _prompt_with_tags(user_prompt: str, n_shots: int, default_fps: int, default_len: int) -> str:
     return (
         "You are a cinematographer and storyboard artist. "
-        "Given a story idea, break it into a sequence of visually DISTINCT, DETAILED shots. "
-        "For each shot, provide the objects in the scene, very specific camera placement, angle, subject position, lighting, and background details. "
-        "Imagine you're describing frames for a film storyboard, not vague events.\n\n"
-        "Return ONLY a JSON array enclosed between <JSON> and </JSON> tags.\n"
-        f"Create a storyboard of {n_shots} shots for this idea:\n\n"
-        f"'''{user_prompt}'''\n\n"
-        "Each item schema:\n"
+        "Break the idea into DISTINCT, DETAILED shots with concrete visual info: objects, camera placement/angle, subject position, lighting, background.\n\n"
+        "Return ONLY a JSON array enclosed between <JSON> and </JSON>.\n"
+        f"Create {n_shots} shots for:\n'''{user_prompt}'''\n\n"
+        "Item schema:\n"
         "{\n"
         '  "id": <int starting at 1>,\n'
         '  "title": "Short shot title",\n'
-        '  "description": "Highly specific visual description for image generation. Include camera angle, framing, time of day, subject position, lighting, mood, and background details.",\n'
+        '  "description": "Highly specific visual description (camera, framing, time of day, subject position, lighting, mood, background).",\n'
         f'  "duration": {default_len},\n'
         f'  "fps": {default_fps},\n'
         '  "steps": 30,\n'
@@ -113,7 +97,7 @@ def _prompt_with_tags(user_prompt: str, n_shots: int, default_fps: int, default_
 
 def _prompt_minimal(user_prompt: str, n_shots: int, default_fps: int, default_len: int) -> str:
     return (
-        "Reply ONLY with a JSON array starting with '[' and ending with ']'. No extra text.\n"
+        "Reply ONLY with a JSON array starting with '[' and ending with ']'.\n"
         f"Storyboard: {n_shots} shots for:\n'''{user_prompt}'''\n"
         "Item schema:\n"
         "{\n"
@@ -133,8 +117,7 @@ def _apply_chat(tok, system_msg: str, user_msg: str) -> str:
         return tok.apply_chat_template(
             [{"role": "system", "content": system_msg},
              {"role": "user", "content": user_msg}],
-            tokenize=False,
-            add_generation_prompt=True
+            tokenize=False, add_generation_prompt=True
         )
     return system_msg + "\n\n" + user_msg
 
@@ -142,46 +125,32 @@ def _generate_text(model, tok, prompt_text: str) -> str:
     inputs = tok(prompt_text, return_tensors="pt")
     inputs = {k: v.to(model.device) for k, v in inputs.items()}
     eos_id = tok.eos_token_id or tok.pad_token_id
-
     gen = model.generate(
-        **inputs,
-        max_new_tokens=HF_TASK_MAX_TOKENS,
-        do_sample=False,
-        temperature=0.0,
-        repetition_penalty=1.05,
-        eos_token_id=eos_id,
-        pad_token_id=eos_id,
+        **inputs, max_new_tokens=HF_TASK_MAX_TOKENS, do_sample=False, temperature=0.0,
+        repetition_penalty=1.05, eos_token_id=eos_id, pad_token_id=eos_id
     )
     prompt_len = inputs["input_ids"].shape[1]
     continuation_ids = gen[0][prompt_len:]
     text = tok.decode(continuation_ids, skip_special_tokens=True).strip()
     if text.startswith("```"):
-        text = re.sub(r"^```(?:json)?\s*|\s*```$", "", text, flags=re.IGNORECASE | re.DOTALL).strip()
+        text = re.sub(r"^```(?:json)?\s*|\s*```$", "", text, flags=re.I|re.S).strip()
     return text
 
 def _extract_json_array(text: str) -> str:
-    m = re.search(r"<JSON>(.*?)</JSON>", text, flags=re.DOTALL | re.IGNORECASE)
-    if m:
-        inner = m.group(1).strip()
-        if inner:
-            return inner
+    m = re.search(r"<JSON>(.*?)</JSON>", text, flags=re.S|re.I)
+    if m and m.group(1).strip():
+        return m.group(1).strip()
     start = text.find("[")
-    if start == -1:
-        return ""
-    depth = 0
-    in_str = False
-    prev = ""
+    if start == -1: return ""
+    depth = 0; in_str = False; prev = ""
     for i in range(start, len(text)):
         ch = text[i]
-        if ch == '"' and prev != '\\':
-            in_str = not in_str
+        if ch == '"' and prev != '\\': in_str = not in_str
         if not in_str:
-            if ch == "[":
-                depth += 1
+            if ch == "[": depth += 1
             elif ch == "]":
                 depth -= 1
-                if depth == 0:
-                    return text[start:i+1].strip()
+                if depth == 0: return text[start:i+1].strip()
         prev = ch
     return ""
 
@@ -205,7 +174,6 @@ def _normalize_shots(shots_raw, default_fps: int, default_len: int):
 def generate_storyboard_with_llm(user_prompt: str, n_shots: int, default_fps: int, default_len: int):
     model, tok = _lazy_model_tok()
     system = "You are a film previsualization assistant. Output must be valid JSON."
-
     p1 = _apply_chat(tok, system + " Return ONLY JSON inside <JSON> tags.",
                      _prompt_with_tags(user_prompt, n_shots, default_fps, default_len))
     out1 = _generate_text(model, tok, p1)
@@ -217,43 +185,29 @@ def generate_storyboard_with_llm(user_prompt: str, n_shots: int, default_fps: in
         out2 = _generate_text(model, tok, p2)
         json_text = _extract_json_array(out2)
         if not json_text and "[" in out2 and "]" in out2:
-            start = out2.find("["); end = out2.rfind("]")
-            if start != -1 and end != -1 and end > start:
-                json_text = out2[start:end+1].strip()
-
-    if not json_text or not json_text.strip():
-        fallback = []
-        for i in range(1, int(n_shots) + 1):
-            fallback.append({
-                "id": i,
-                "title": f"Shot {i}",
-                "description": f"Simple placeholder for: {user_prompt[:80]}",
-                "duration": default_len,
-                "fps": default_fps,
-                "steps": 30,
-                "seed": None,
-                "negative": "",
-                "image_path": None
-            })
-        return fallback
+            start, end = out2.find("["), out2.rfind("]")
+            if start != -1 and end > start: json_text = out2[start:end+1].strip()
+
+    if not json_text:
+        return [{
+            "id": i, "title": f"Shot {i}",
+            "description": f"Placeholder for: {user_prompt[:80]}",
+            "duration": default_len, "fps": default_fps,
+            "steps": 30, "seed": None, "negative": "", "image_path": None
+        } for i in range(1, int(n_shots)+1)]
 
     try:
         shots_raw = json.loads(json_text)
     except Exception:
-        json_text_clean = re.sub(r",\s*([\]\}])", r"\1", json_text)
-        shots_raw = json.loads(json_text_clean)
-
+        shots_raw = json.loads(re.sub(r",\s*([\]\}])", r"\1", json_text))
     return _normalize_shots(shots_raw, default_fps, default_len)
 
-
 # =========================
-# IMAGE GEN — FLUX only (no fallback) + Temporal chaining
+# IMAGE GEN — FLUX-only + Temporal chaining
 # =========================
 USE_CUDA = torch.cuda.is_available()
 DTYPE = torch.float16 if USE_CUDA else torch.float32
-
-# Correct, gated repo; accept access and set HF_TOKEN
-FLUX_MODEL = os.getenv("FLUX_MODEL", "black-forest-labs/FLUX.1-schnell")
+FLUX_MODEL = os.getenv("FLUX_MODEL", "black-forest-labs/FLUX.1-schnell")  # gated
 HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
 
 _flux_t2i = None
@@ -276,134 +230,96 @@ def _lazy_flux_pipes():
 
 def _flux_healthcheck():
     if not HF_TOKEN:
-        raise RuntimeError(
-            "HF_TOKEN is not set. FLUX models are gated; set a Hugging Face READ token "
-            "and accept the model terms on the repo page."
-        )
+        raise RuntimeError("HF_TOKEN is not set. Accept the model terms on HF and provide a READ token.")
     _lazy_flux_pipes()
 
 def _save_keyframe(pid: str, shot_id: int, img: Image.Image) -> str:
     pdir = project_dir(pid)
     out = os.path.join(pdir, "keyframes", f"shot_{shot_id:02d}.png")
-    img.save(out)
-    return out
-
+    img.save(out); return out
 
-# ---- Temporal prompt composer (PRIORITIZE the new shot) ----
 def _compose_temporal_prompt(shots: list, idx: int, seconds_forward: int = 5):
-    """
-    Build a prompt that continues the scene N seconds later,
-    prioritizing the NEW shot description (composition/action),
-    while keeping only identity/lighting/environment continuity.
-    Returns (composed_prompt, composed_negative).
-    """
     curr = shots[idx]
     curr_desc = (curr.get("description") or "").strip()
     curr_neg  = (curr.get("negative") or "").strip()
-
-    if idx == 0:
-        return curr_desc, curr_neg
-
-    prev = shots[idx - 1]
-    prev_desc = (prev.get("description") or "").strip()
-
+    if idx == 0: return curr_desc, curr_neg
+    prev_desc = (shots[idx-1].get("description") or "").strip()
     composed = (
         f"Continue the same scene {seconds_forward} seconds later.\n"
-        f'PRIORITIZE this new moment and its composition now: "{curr_desc}".\n'
-        "Keep continuity ONLY for subject identity, lighting palette, time of day, and general environment style.\n"
+        f'PRIORITIZE this new moment & composition: "{curr_desc}".\n'
+        "Keep continuity ONLY for subject identity, lighting palette, time of day, environment style.\n"
         f'Previous frame (context only, do not copy its framing): "{prev_desc}".\n'
-        f"Avoid replicating the previous composition; allow camera move / subject reposition consistent with {seconds_forward} seconds of natural progression."
+        f"Avoid replicating the previous composition; allow camera move / subject reposition consistent with {seconds_forward} seconds of progression."
     ).strip()
-
-    negative = (
-        curr_neg + (
-            "; identical composition as previous; exact same framing; rigid pose repeat; freeze frame; "
-            "hard scene reset; different subject identity; wildly different art style; unrelated background"
-        )
-    ).strip("; ")
-
+    negative = (curr_neg + "; identical composition as previous; exact same framing; rigid pose repeat; freeze frame; "
+                           "hard scene reset; different subject identity; wildly different art style; unrelated background").strip("; ")
     return composed, negative
 
-
 @spaces.GPU(duration=180)
 def generate_keyframe_image(
-    pid: str,
-    shot_idx: int,
-    shots: list,
-    t2i_steps: int = 18,         # FLUX: 12–22
-    i2i_steps: int = 22,         # FLUX: 16–26
-    i2i_strength: float = 0.90,  # ↑ more change toward new prompt
-    guidance_scale: float = 3.4, # ↑ stronger text pull
-    width: int = 640,
-    height: int = 640,
-    seconds_forward: int = 5,    # temporal step
-    aggressive: bool = False     # optional push
+    pid: str, shot_idx: int, shots: list,
+    t2i_steps: int = 18, i2i_steps: int = 22, i2i_strength: float = 0.90,
+    guidance_scale: float = 3.4, width: int = 640, height: int = 640,
+    seconds_forward: int = 5, aggressive: bool = False
 ):
-    """
-    Generate image for shots[shot_idx] using FLUX only.
-    - Shot 1: text2img
-    - Shot k>1: img2img from previous approved frame + temporal prompt ("N seconds later")
-    """
     try:
         t2i, i2i = _lazy_flux_pipes()
     except Exception as e:
-        raise gr.Error(
-            f"FLUX failed to load: {e}\n"
-            "Set FLUX_MODEL (e.g., 'black-forest-labs/FLUX.1-schnell') and ensure HF_TOKEN if required."
-        )
+        raise gr.Error(f"FLUX failed to load: {e}")
 
-    # Build temporal prompt
-    composed_prompt, composed_negative = _compose_temporal_prompt(shots, shot_idx, seconds_forward=seconds_forward)
+    prompt, negative = _compose_temporal_prompt(shots, shot_idx, seconds_forward=seconds_forward)
 
-    # RNG / seed
     seed = shots[shot_idx].get("seed", None)
     device = "cuda" if USE_CUDA else "cpu"
     gen = torch.Generator(device)
-    if isinstance(seed, int):
-        gen = gen.manual_seed(int(seed))
+    if isinstance(seed, int): gen = gen.manual_seed(int(seed))
 
-    # sizes
-    width  = max(256, min(1024, int(width)))
+    width = max(256, min(1024, int(width)))
     height = max(256, min(1024, int(height)))
 
-    # chaining
     prev_path = shots[shot_idx - 1].get("image_path") if shot_idx > 0 else None
     use_prev = bool(shot_idx > 0 and prev_path and os.path.exists(prev_path))
 
-    # Aggressive mode bumps
     if aggressive:
         i2i_strength = min(0.98, max(i2i_strength, 0.92))
         guidance_scale = max(guidance_scale, 3.6)
         i2i_steps = max(i2i_steps, 24)
 
-    # generate
     if not use_prev:
         out = t2i(
-            prompt=composed_prompt,
-            negative_prompt=composed_negative or None,
+            prompt=prompt, negative_prompt=(negative or None),
             num_inference_steps=int(max(10, t2i_steps)),
             guidance_scale=float(max(2.4, guidance_scale)),
-            generator=gen,
-            width=width, height=height
+            generator=gen, width=width, height=height
         ).images[0]
     else:
-        init_image = Image.open(prev_path).convert("RGB")  # previous approved frame (the "init_image")
+        init_image = Image.open(prev_path).convert("RGB")
         out = i2i(
-            prompt=composed_prompt,
-            negative_prompt=composed_negative or None,
-            image=init_image,
-            strength=float(min(max(i2i_strength, 0.70), 0.98)),
+            prompt=prompt, negative_prompt=(negative or None),
+            image=init_image, strength=float(min(max(i2i_strength, 0.70), 0.98)),
             num_inference_steps=int(max(14, i2i_steps)),
-            guidance_scale=float(max(2.4, guidance_scale)),
-            generator=gen
+            guidance_scale=float(max(2.4, guidance_scale)), generator=gen
         ).images[0]
 
-    saved_path = _save_keyframe(pid, int(shots[shot_idx]["id"]), out)
-    return saved_path
+    saved = _save_keyframe(pid, int(shots[shot_idx]["id"]), out)
+    return saved
 
+# =========================
+# MoviePy lazy install/import
+# =========================
+def _ensure_moviepy():
+    try:
+        from moviepy.editor import ImageClip, CompositeVideoClip, concatenate_videoclips
+        from moviepy.video.io.VideoFileClip import VideoFileClip
+        return ImageClip, CompositeVideoClip, concatenate_videoclips, VideoFileClip
+    except Exception:
+        subprocess.check_call([sys.executable, "-m", "pip", "install", "moviepy", "imageio", "imageio-ffmpeg"])
+        from moviepy.editor import ImageClip, CompositeVideoClip, concatenate_videoclips
+        from moviepy.video.io.VideoFileClip import VideoFileClip
+        return ImageClip, CompositeVideoClip, concatenate_videoclips, VideoFileClip
 
 # =========================
-# Video stitching helpers (pairwise dissolve + final concat)
+# Video stitching (pairwise dissolve + final concat)
 # =========================
 def _pair_clip_path(pid: str, i: int, j: int) -> str:
     return os.path.join(project_dir(pid), "clips", f"pair_{i:02d}_to_{j:02d}.mp4")
@@ -416,35 +332,18 @@ def _image_size(path: str):
         return im.width, im.height
 
 def _build_pair_clip(img_a: str, img_b: str, out_path: str, fps: int = 24, hold: float = 0.5, crossfade: float = 0.7, resize_to=None):
-    """
-    Create a dissolve transition from img_a -> img_b:
-      - show img_a for `hold` seconds
-      - dissolve for `crossfade` seconds into img_b
-      - hold img_b for `hold` seconds
-    """
+    ImageClip, CompositeVideoClip, concatenate_videoclips, VideoFileClip = _ensure_moviepy()
     ca = ImageClip(img_a).set_duration(hold + crossfade)
     cb = ImageClip(img_b).set_duration(hold + crossfade).set_start(hold)
-
     if resize_to:
         ca = ca.resize(newsize=resize_to)
         cb = cb.resize(newsize=resize_to)
-
     ca_x = ca.crossfadeout(crossfade)
     cb_x = cb.crossfadein(crossfade)
-
     total = hold + crossfade + hold
     comp = CompositeVideoClip([ca_x, cb_x]).set_duration(total)
-
-    comp.write_videofile(
-        out_path,
-        fps=fps,
-        codec="libx264",
-        audio=False,
-        preset="medium",
-        threads=os.cpu_count() or 2,
-        verbose=False,
-        logger=None
-    )
+    comp.write_videofile(out_path, fps=fps, codec="libx264", audio=False, preset="medium",
+                         threads=os.cpu_count() or 2, verbose=False, logger=None)
     comp.close(); ca.close(); cb.close()
 
 def _build_all_pair_clips(pid: str, shots: list, fps: int = 24, hold: float = 0.5, crossfade: float = 0.7, force_size=None):
@@ -460,37 +359,23 @@ def _build_all_pair_clips(pid: str, shots: list, fps: int = 24, hold: float = 0.
     for i in range(len(shots)-1):
         a = shots[i].get("image_path")
         b = shots[i+1].get("image_path")
-        if not (a and b and os.path.exists(a) and os.path.exists(b)):
-            continue
+        if not (a and b and os.path.exists(a) and os.path.exists(b)): continue
         outp = _pair_clip_path(pid, shots[i]["id"], shots[i+1]["id"])
         _build_pair_clip(a, b, outp, fps=fps, hold=hold, crossfade=crossfade, resize_to=size)
         paths.append(outp)
     return paths
 
 def _build_final_stitched_from_pairs(pair_paths: list, out_path: str, fps: int = 24):
-    if not pair_paths:
-        raise RuntimeError("No pair clips to stitch.")
-    clips = []
-    for p in pair_paths:
-        if os.path.exists(p):
-            clips.append(VideoFileClip(p))
-    if not clips:
-        raise RuntimeError("No readable pair clips on disk.")
+    ImageClip, CompositeVideoClip, concatenate_videoclips, VideoFileClip = _ensure_moviepy()
+    if not pair_paths: raise RuntimeError("No pair clips to stitch.")
+    clips = [VideoFileClip(p) for p in pair_paths if os.path.exists(p)]
+    if not clips: raise RuntimeError("No readable pair clips on disk.")
     final = concatenate_videoclips(clips, method="compose")
-    final.write_videofile(
-        out_path,
-        fps=fps,
-        codec="libx264",
-        audio=False,
-        preset="medium",
-        threads=os.cpu_count() or 2,
-        verbose=False,
-        logger=None
-    )
+    final.write_videofile(out_path, fps=fps, codec="libx264", audio=False, preset="medium",
+                          threads=os.cpu_count() or 2, verbose=False, logger=None)
     final.close()
     for c in clips: c.close()
 
-
 # =========================
 # Shots <-> DataFrame utils
 # =========================
@@ -516,23 +401,19 @@ def df_to_shots(df: pd.DataFrame) -> list:
         })
     return sorted(out, key=lambda x: x["id"])
 
-
 # =========================
 # Gradio UI
 # =========================
 with gr.Blocks() as demo:
     gr.Markdown("# 🎬 Storyboard → Keyframes → Videos → Export")
     gr.Markdown(
-        "Edit storyboard prompts, then generate keyframes.\n"
-        "**Temporal chaining**: each new shot is generated N seconds later from the previous approved frame, "
+        "Temporal chaining: each new shot is generated N seconds later from the previous approved frame, "
         "while the current shot description drives composition & action. **Model**: FLUX-only."
     )
 
-    # State
     project = gr.State(None)
     current_idx = gr.State(0)
 
-    # Header
     with gr.Row():
         with gr.Column(scale=2):
             proj_name = gr.Textbox(label="Project name", placeholder="e.g., Desert Chase")
@@ -545,21 +426,20 @@ with gr.Blocks() as demo:
             load_btn = gr.Button("Load")
     sb_status = gr.Markdown("")
 
-    # Tabs
     with gr.Tabs():
         with gr.Tab("Storyboard"):
             gr.Markdown("### 1) Storyboard")
-            sb_prompt = gr.Textbox(label="High-level prompt", lines=4, placeholder="Describe the story you want to create…")
+            sb_prompt = gr.Textbox(label="High-level prompt", lines=4, placeholder="Describe the story…")
             with gr.Row():
                 sb_target_shots = gr.Slider(1, 12, value=3, step=1, label="Target # of shots")
                 sb_default_fps  = gr.Slider(8, 60, value=24, step=1, label="Default FPS")
-                sb_default_len  = gr.Slider(1, 12, value=4, step=1, label="Default seconds per shot")
-            propose_btn = gr.Button("Propose Storyboard (LLM on ZeroGPU)")
+                sb_default_len  = gr.Slider(1, 12, value=4, step=1, label="Default seconds/shot")
+            propose_btn = gr.Button("Propose Storyboard (LLM)")
             shots_df    = gr.Dataframe(
                 headers=SHOT_COLUMNS,
                 datatype=["number","str","str","number","number","number","number","str","str"],
                 row_count=(1,"dynamic"), col_count=len(SHOT_COLUMNS),
-                label="Edit shots below (prompts & params)", wrap=True
+                label="Edit shots (prompts & params)", wrap=True
             )
             save_edits_btn = gr.Button("Save Edits ✓", variant="primary", interactive=False)
             with gr.Row():
@@ -569,18 +449,16 @@ with gr.Blocks() as demo:
         with gr.Tab("Keyframes"):
             gr.Markdown("### 2) Keyframes")
             shot_info_md = gr.Markdown("")
-            prompt_box   = gr.Textbox(label="Shot description (editable before generating)", lines=4)
+            prompt_box   = gr.Textbox(label="Shot description (editable)", lines=4)
             with gr.Row():
                 gen_btn = gr.Button("Generate / Regenerate", variant="primary")
                 approve_next_btn = gr.Button("Approve & Next →", variant="secondary")
-
             with gr.Row():
                 img_strength = gr.Slider(0.50, 0.98, value=0.90, step=0.02, label="Change vs Consistency (img2img strength)")
                 img_steps    = gr.Slider(12, 28, value=22, step=1, label="Inference Steps (img2img)")
                 guidance     = gr.Slider(2.4, 4.0, value=3.4, step=0.1, label="Guidance Scale")
                 temporal_secs = gr.Slider(1, 10, value=5, step=1, label="Temporal step (seconds later)")
                 aggressive_follow = gr.Checkbox(value=False, label="Aggressive follow prompt (more change)")
-
             with gr.Row():
                 prev_img = gr.Image(label="Previous approved image (conditioning)", type="filepath")
                 out_img  = gr.Image(label="Generated image", type="filepath")
@@ -610,31 +488,22 @@ with gr.Blocks() as demo:
 
     def on_propose(p, prompt, target_shots, fps, vlen):
         p = ensure_project(p, suggested_name=(proj_name.value if hasattr(proj_name, "value") else "Project"))
-        if not prompt or not str(prompt).strip():
+        if not str(prompt or "").strip():
             raise gr.Error("Please enter a high-level prompt.")
         shots = generate_storyboard_with_llm(str(prompt).strip(), int(target_shots), int(fps), int(vlen))
-        p = dict(p)
-        p["shots"] = shots
-        p["meta"]["updated"] = now_iso()
-        save_project(p)
+        p = dict(p); p["shots"] = shots; p["meta"]["updated"] = now_iso(); save_project(p)
         return p, shots_to_df(shots), gr.update(value="Storyboard generated (editable)."), gr.update(interactive=True)
 
-    propose_btn.click(
-        on_propose,
+    propose_btn.click(on_propose,
         inputs=[project, sb_prompt, sb_target_shots, sb_default_fps, sb_default_len],
         outputs=[project, shots_df, sb_status, save_edits_btn]
     )
 
     def on_save_edits(p, df):
-        if p is None:
-            raise gr.Error("No project in memory. Click New Project, then generate a storyboard.")
-        if df is None:
-            raise gr.Error("No storyboard table to save. Generate a storyboard first, then edit it.")
+        if p is None: raise gr.Error("No project in memory.")
+        if df is None: raise gr.Error("No storyboard table to save.")
         shots = df_to_shots(df)
-        p = dict(p)
-        p["shots"] = shots
-        p["meta"]["updated"] = now_iso()
-        save_project(p)
+        p = dict(p); p["shots"] = shots; p["meta"]["updated"] = now_iso(); save_project(p)
         return p, gr.update(value="Edits saved.")
 
     save_edits_btn.click(on_save_edits, inputs=[project, shots_df], outputs=[project, sb_status])
@@ -643,42 +512,23 @@ with gr.Blocks() as demo:
         if p is None: raise gr.Error("No project.")
         shots = df_to_shots(df)
         if not shots: raise gr.Error("Storyboard is empty.")
-
-        # lock a single seed for the project:
         proj_seed = None
-        if proj_seed_override not in [None, ""] and str(proj_seed_override).isdigit():
-            proj_seed = int(proj_seed_override)
-        if proj_seed is None:
-            proj_seed = p.get("meta", {}).get("seed", None)
+        if str(proj_seed_override or "").isdigit(): proj_seed = int(proj_seed_override)
+        if proj_seed is None: proj_seed = p.get("meta", {}).get("seed")
         if proj_seed is None:
             for s in shots:
-                if isinstance(s.get("seed"), int):
-                    proj_seed = int(s["seed"])
-                    break
-        if proj_seed is None:
-            proj_seed = int(torch.randint(0, 2**31 - 1, (1,)).item())
-
+                if isinstance(s.get("seed"), int): proj_seed = int(s["seed"]); break
+        if proj_seed is None: proj_seed = int(torch.randint(0, 2**31 - 1, (1,)).item())
         for s in shots:
-            if not isinstance(s.get("seed"), int):
-                s["seed"] = proj_seed
-
-        p = dict(p)
-        p["shots"] = shots
-        p["meta"]["seed"] = proj_seed
-        p["meta"]["updated"] = now_iso()
-        save_project(p)
-
-        idx = 0
-        prev_path = None
-        info = (
-            f"**Shot {shots[idx]['id']} — {shots[idx]['title']}**  \n"
-            f"Duration: {shots[idx]['duration']}s @ {shots[idx]['fps']} fps  \n"
-            f"Locked project seed: `{proj_seed}`"
-        )
-        return p, 0, gr.update(value=info), gr.update(value=shots[idx]["description"]), gr.update(value=prev_path), gr.update(value=None), gr.update(value=f"Ready to generate shot 1."), gr.update(value=proj_seed)
-
-    to_keyframes_btn.click(
-        on_start_keyframes,
+            if not isinstance(s.get("seed"), int): s["seed"] = proj_seed
+        p = dict(p); p["shots"] = shots; p["meta"]["seed"] = proj_seed; p["meta"]["updated"] = now_iso(); save_project(p)
+        idx = 0; prev_path = None
+        info = (f"**Shot {shots[idx]['id']} — {shots[idx]['title']}**  \n"
+                f"Duration: {shots[idx]['duration']}s @ {shots[idx]['fps']} fps  \n"
+                f"Locked project seed: `{proj_seed}`")
+        return p, 0, gr.update(value=info), gr.update(value=shots[idx]["description"]), gr.update(value=prev_path), gr.update(value=None), gr.update(value="Ready for shot 1."), gr.update(value=proj_seed)
+
+    to_keyframes_btn.click(on_start_keyframes,
         inputs=[project, shots_df, proj_seed_box],
         outputs=[project, current_idx, shot_info_md, prompt_box, prev_img, out_img, kf_status, proj_seed_box]
     )
@@ -687,124 +537,93 @@ with gr.Blocks() as demo:
         if p is None: raise gr.Error("No project.")
         shots = p["shots"]
         if idx < 0 or idx >= len(shots): raise gr.Error("Invalid shot index.")
-        shots[idx]["description"] = current_prompt  # allow tweaking
-
+        shots[idx]["description"] = current_prompt
         img_path = generate_keyframe_image(
-            p["meta"]["id"],
-            int(idx),
-            shots,
-            t2i_steps=18,
-            i2i_steps=int(i2i_steps_val),
+            p["meta"]["id"], int(idx), shots,
+            t2i_steps=18, i2i_steps=int(i2i_steps_val),
             i2i_strength=float(i2i_strength_val),
             guidance_scale=float(guidance_val),
-            width=640,
-            height=640,
+            width=640, height=640,
             seconds_forward=int(seconds_forward_val),
             aggressive=bool(aggressive_val)
         )
         prev_path = shots[idx-1]["image_path"] if idx > 0 else None
         return img_path, (prev_path or None), gr.update(value=f"Generated candidate for shot {shots[idx]['id']}.")
 
-    gen_btn.click(
-        on_generate_img,
+    gen_btn.click(on_generate_img,
         inputs=[project, current_idx, prompt_box, img_strength, img_steps, guidance, temporal_secs, aggressive_follow],
         outputs=[out_img, prev_img, kf_status]
     )
 
     def on_approve_next(p, idx, current_prompt, latest_img_path):
         if p is None: raise gr.Error("No project.")
-        shots = p["shots"]
-        i = int(idx)
+        shots = p["shots"]; i = int(idx)
         if i < 0 or i >= len(shots): raise gr.Error("Invalid shot index.")
         if not latest_img_path: raise gr.Error("Generate an image first.")
-
-        # commit
         shots[i]["description"] = current_prompt
         shots[i]["image_path"] = latest_img_path
-        p["shots"] = shots
-        p["meta"]["updated"] = now_iso()
-        save_project(p)
-
-        # next
+        p["shots"] = shots; p["meta"]["updated"] = now_iso(); save_project(p)
         if i + 1 < len(shots):
             ni = i + 1
-            info = (
-                f"**Shot {shots[ni]['id']} — {shots[ni]['title']}**  \n"
-                f"Duration: {shots[ni]['duration']}s @ {shots[ni]['fps']} fps  \n"
-                f"Locked project seed: `{p['meta'].get('seed')}`"
-            )
+            info = (f"**Shot {shots[ni]['id']} — {shots[ni]['title']}**  \n"
+                    f"Duration: {shots[ni]['duration']}s @ {shots[ni]['fps']} fps  \n"
+                    f"Locked project seed: `{p['meta'].get('seed')}`")
             prev_path = shots[ni-1]["image_path"]
             return p, ni, gr.update(value=info), gr.update(value=shots[ni]["description"]), gr.update(value=prev_path), gr.update(value=None), gr.update(value=f"Approved shot {shots[i]['id']}. On to shot {shots[ni]['id']}.")
         else:
             return p, i, gr.update(value="**All keyframes approved.** Proceed to Videos tab."), gr.update(value=""), gr.update(value=shots[i]["image_path"]), gr.update(value=None), gr.update(value="All shots approved ✅")
 
-    approve_next_btn.click(on_approve_next, inputs=[project, current_idx, prompt_box, out_img], outputs=[project, current_idx, shot_info_md, prompt_box, prev_img, out_img, kf_status])
+    approve_next_btn.click(on_approve_next,
+        inputs=[project, current_idx, prompt_box, out_img],
+        outputs=[project, current_idx, shot_info_md, prompt_box, prev_img, out_img, kf_status]
+    )
 
-    # ---- Videos tab handlers
+    # ---- Videos tab
     def on_build_pairs(p, fps, hold, xfade):
-        if p is None:
-            raise gr.Error("No project.")
+        if p is None: raise gr.Error("No project.")
         shots = p.get("shots", [])
-        if len(shots) < 2:
-            raise gr.Error("Need at least 2 approved images to build pair clips.")
-        if not any(s.get("image_path") for s in shots):
-            raise gr.Error("No approved images yet. Approve keyframes first.")
-
+        if len(shots) < 2: raise gr.Error("Need at least 2 approved images.")
+        if not any(s.get("image_path") for s in shots): raise gr.Error("No approved images yet.")
         pair_paths = _build_all_pair_clips(
             p["meta"]["id"], shots,
             fps=int(fps), hold=float(hold), crossfade=float(xfade),
-            force_size=None  # or (640, 640) to force uniform size
+            force_size=None
         )
-        if not pair_paths:
-            raise gr.Error("Could not find any consecutive pairs with images.")
+        if not pair_paths: raise gr.Error("No consecutive pairs with images found.")
         return {"pair_clips": pair_paths, "final": None}
 
-    build_pairs_btn.click(
-        on_build_pairs,
-        inputs=[project, v_fps, v_hold, v_xfade],
-        outputs=[vd_table]
-    )
+    build_pairs_btn.click(on_build_pairs, inputs=[project, v_fps, v_hold, v_xfade], outputs=[vd_table])
 
     def on_build_final(p, fps):
-        if p is None:
-            raise gr.Error("No project.")
+        if p is None: raise gr.Error("No project.")
         pid = p["meta"]["id"]
         clips_dir = os.path.join(project_dir(pid), "clips")
-        pair_paths = sorted(
-            [os.path.join(clips_dir, f) for f in os.listdir(clips_dir) if f.startswith("pair_") and f.endswith(".mp4")]
-        )
-        if not pair_paths:
-            raise gr.Error("No pair clips found. Click 'Build pair clips' first.")
+        pair_paths = sorted([os.path.join(clips_dir, f) for f in os.listdir(clips_dir)
+                             if f.startswith("pair_") and f.endswith(".mp4")])
+        if not pair_paths: raise gr.Error("No pair clips found. Build pair clips first.")
         outp = _final_stitched_path(pid)
         _build_final_stitched_from_pairs(pair_paths, outp, fps=int(fps))
         return {"pair_clips": pair_paths, "final": outp}
 
-    build_final_btn.click(
-        on_build_final,
-        inputs=[project, v_fps],
-        outputs=[vd_table]
-    )
+    build_final_btn.click(on_build_final, inputs=[project, v_fps], outputs=[vd_table])
 
+    # save/load
     def on_save(p):
-        if p is None:
-            raise gr.Error("No project in memory.")
-        path = save_project(p)
-        return gr.update(value=f"Saved to `{path}`")
+        if p is None: raise gr.Error("No project in memory.")
+        path = save_project(p); return gr.update(value=f"Saved to `{path}`")
 
     save_btn.click(on_save, inputs=[project], outputs=[sb_status])
 
     def on_load(file_obj):
         p = load_project_file(file_obj)
         seed_val = p.get("meta", {}).get("seed", None)
-        return (
-            p,
-            gr.update(value=f"Loaded project `{p['meta']['name']}` (id: `{p['meta']['id']}`)"),
-            shots_to_df(p.get("shots", [])),
-            gr.update(value=seed_val)
-        )
+        return (p,
+                gr.update(value=f"Loaded `{p['meta']['name']}` (id: `{p['meta']['id']}`)"),
+                shots_to_df(p.get("shots", [])),
+                gr.update(value=seed_val))
 
     load_btn.click(on_load, inputs=[load_file], outputs=[project, sb_status, shots_df, proj_seed_box])
 
 if __name__ == "__main__":
-    _flux_healthcheck()  # fail fast with clear error if FLUX isn't available
+    _flux_healthcheck()
     demo.launch()