Batch vision checks after generation, reject inconclusive answers, recommend llava
This commit is contained in:
Justin Oros
@@ -48,7 +48,7 @@ REVIEW_MODES = ("advisory", "block", "off")
|
||||
ENGINES = ("local", "veo")
|
||||
TTS_ENGINES = ("auto", "kokoro", "piper", "say", "espeak")
|
||||
SUGGESTED_TEXT_MODELS = ["llama3.1:8b", "llama3.2:3b", "qwen2.5:7b", "mistral:7b"]
|
||||
SUGGESTED_VISION_MODELS = ["moondream", "llava:7b", "llama3.2-vision", "qwen2.5vl:7b"]
|
||||
SUGGESTED_VISION_MODELS = ["llava:7b", "qwen2.5vl:7b", "llama3.2-vision", "moondream"]
|
||||
SMALL_VISION_MODELS = ("moondream", "llava-phi3", "bakllava")
|
||||
|
||||
GLOBAL_DEFAULTS = {
|
||||
@@ -61,7 +61,7 @@ GLOBAL_DEFAULTS = {
|
||||
"video_engine": "local",
|
||||
"ollama_url": "http://localhost:11434",
|
||||
"ollama_model": "llama3.1:8b",
|
||||
"vision_model": "moondream",
|
||||
"vision_model": "llava:7b",
|
||||
"ollama_timeout": 300,
|
||||
"sd_model": "stabilityai/sdxl-turbo",
|
||||
"sd_steps": 4,
|
||||
@@ -1129,6 +1129,9 @@ def unload_pipeline():
|
||||
logger.debug("Released image model from memory")
|
||||
|
||||
|
||||
INCONCLUSIVE = ("don't know", "do not know", "not sure", "cannot", "can't", "unable", "no image", "what you see", "short explanation")
|
||||
|
||||
|
||||
def kids_vision_check(s, clip_path):
|
||||
model = s["vision_model"].strip()
|
||||
if not model:
|
||||
@@ -1142,30 +1145,35 @@ def kids_vision_check(s, clip_path):
|
||||
raise RuntimeError("Could not extract a frame for the vision check")
|
||||
images = [base64.b64encode(frame.read_bytes()).decode()]
|
||||
frame.unlink(missing_ok=True)
|
||||
if not model.split(":")[0] in SMALL_VISION_MODELS:
|
||||
unload_pipeline()
|
||||
safe, reason = True, ""
|
||||
verdict, detail = None, ""
|
||||
try:
|
||||
try:
|
||||
data = ollama_call(s, KIDS_VISION_PROMPT, temperature=0.1, model=model, images=images, num_predict=256)
|
||||
desc = str(data.get("description", ""))
|
||||
reason = str(data.get("reason", ""))
|
||||
echoed = "it is unsafe if" in reason.lower() or reason.strip().lower() in ("why", "short explanation")
|
||||
if echoed or desc.strip().lower() in ("what you see", "description"):
|
||||
raise ValueError("model echoed the prompt")
|
||||
safe = truthy(data.get("safe"))
|
||||
data = ollama_call(s, KIDS_VISION_PROMPT, temperature=0.1, model=model, images=images, num_predict=300)
|
||||
desc = str(data.get("description", "")).strip()
|
||||
reason = str(data.get("reason", "")).strip()
|
||||
low = f"{desc} {reason}".lower()
|
||||
echoed = "it is unsafe if" in low
|
||||
if desc and not echoed and not any(x in low for x in INCONCLUSIVE) and "safe" in data:
|
||||
verdict = truthy(data.get("safe"))
|
||||
detail = f"{desc} | {reason}"
|
||||
except (requests.exceptions.HTTPError, ValueError):
|
||||
text = ollama_call(s, VISION_TEXT_PROMPT, temperature=0.1, model=model, images=images, num_predict=128, as_json=False)
|
||||
pass
|
||||
if verdict is None:
|
||||
text = ollama_call(s, VISION_TEXT_PROMPT, temperature=0.1, model=model, images=images, num_predict=150, as_json=False).strip()
|
||||
low = text.lower()
|
||||
safe = "unsafe" not in low
|
||||
reason = text.strip()[:200]
|
||||
if low.startswith("unsafe") or re.search(r"\bunsafe\b", low):
|
||||
verdict, detail = False, text
|
||||
elif re.match(r"^\W*safe\b", low) and not any(x in low for x in INCONCLUSIVE):
|
||||
verdict, detail = True, text
|
||||
except Cancelled:
|
||||
raise
|
||||
except Exception as e:
|
||||
raise VisionUnavailable(f"vision model '{model}' could not run: {e}") from e
|
||||
if not safe:
|
||||
raise ValueError(f"Vision check rejected {clip_path.name}: {reason or 'no reason given'}")
|
||||
logger.info("Vision check passed for %s", clip_path.name)
|
||||
if verdict is None:
|
||||
raise VisionUnavailable(f"vision model '{model}' gave no clear answer, try llava:7b")
|
||||
if not verdict:
|
||||
raise ValueError(f"Vision check rejected {clip_path.name}: {detail[:200] or 'no reason given'}")
|
||||
logger.info("Vision check passed for %s: %s", clip_path.name, detail[:160])
|
||||
|
||||
|
||||
def load_pipeline(s):
|
||||
@@ -1497,16 +1505,18 @@ def run_job(s):
|
||||
extras = []
|
||||
vision_log = []
|
||||
n = len(script["scenes"])
|
||||
for i, scene in enumerate(script["scenes"], 1):
|
||||
check()
|
||||
set_stage(f"generating scene {i}/{n}", step=3 + (i - 1) * 2)
|
||||
path = job_dir / f"clip_{i:02d}.mp4"
|
||||
|
||||
def build_scene(i, scene, path):
|
||||
for attempt in range(1, 4):
|
||||
check()
|
||||
try:
|
||||
if engine == "local":
|
||||
extra = make_clip_local(s, scene, script, path, kids)
|
||||
else:
|
||||
extra = make_clip_veo(s, scene, script, path, kids)
|
||||
if extra and extra not in extras:
|
||||
extras.append(extra)
|
||||
return
|
||||
except Cancelled:
|
||||
raise
|
||||
except Exception as e:
|
||||
@@ -1515,32 +1525,45 @@ def run_job(s):
|
||||
if attempt == 3:
|
||||
raise
|
||||
stop_event.wait(10)
|
||||
continue
|
||||
|
||||
for i, scene in enumerate(script["scenes"], 1):
|
||||
set_stage(f"generating scene {i}/{n}", step=2 + i)
|
||||
path = job_dir / f"clip_{i:02d}.mp4"
|
||||
build_scene(i, scene, path)
|
||||
clips.append(path)
|
||||
|
||||
if kids:
|
||||
set_stage(f"vision check scene {i}/{n}", step=4 + (i - 1) * 2)
|
||||
pending = list(range(1, n + 1))
|
||||
for rnd in range(1, 4):
|
||||
unload_pipeline()
|
||||
rejected = []
|
||||
for i in pending:
|
||||
check()
|
||||
set_stage(f"vision check scene {i}/{n}", step=2 + n + i)
|
||||
try:
|
||||
kids_vision_check(s, path)
|
||||
vision_log.append({"clip": i, "attempt": attempt, "result": "passed"})
|
||||
kids_vision_check(s, clips[i - 1])
|
||||
vision_log.append({"clip": i, "round": rnd, "result": "passed"})
|
||||
except Cancelled:
|
||||
raise
|
||||
except VisionUnavailable as e:
|
||||
if s["kids_manual_review"]:
|
||||
logger.warning("Scene %d: %s. Continuing, check this video manually before publishing", i, e)
|
||||
vision_log.append({"clip": i, "attempt": attempt, "result": f"skipped: {e}"[:300]})
|
||||
vision_log.append({"clip": i, "round": rnd, "result": f"skipped: {e}"[:300]})
|
||||
else:
|
||||
raise RuntimeError(f"{e}. Turn on manual review or pick a working vision model")
|
||||
raise RuntimeError(f"{e}. Frames were not verified, so nothing was uploaded")
|
||||
except Exception as e:
|
||||
logger.warning("Scene %d attempt %d rejected: %s", i, attempt, e)
|
||||
vision_log.append({"clip": i, "attempt": attempt, "result": str(e)[:300]})
|
||||
path.unlink(missing_ok=True)
|
||||
if attempt == 3:
|
||||
raise
|
||||
stop_event.wait(5)
|
||||
continue
|
||||
if extra:
|
||||
extras.append(extra)
|
||||
logger.warning("Scene %d rejected (round %d): %s", i, rnd, e)
|
||||
vision_log.append({"clip": i, "round": rnd, "result": str(e)[:300]})
|
||||
rejected.append(i)
|
||||
if not rejected:
|
||||
break
|
||||
clips.append(path)
|
||||
if rnd == 3:
|
||||
raise RuntimeError(f"Scenes {rejected} still failed the vision check after 3 rounds")
|
||||
for i in rejected:
|
||||
set_stage(f"regenerating scene {i}/{n}", step=2 + i)
|
||||
clips[i - 1].unlink(missing_ok=True)
|
||||
build_scene(i, script["scenes"][i - 1], clips[i - 1])
|
||||
pending = rejected
|
||||
if kids:
|
||||
compliance["vision_checks"] = vision_log
|
||||
set_stage("stitching video", step=3 + n * 2)
|
||||
|
||||
+1
-1
@@ -153,7 +153,7 @@ const FIELDS=[
|
||||
["","Text and shared options","header","",""],
|
||||
["ollama_url","Ollama URL","text","","g"],
|
||||
["ollama_model","Ollama model","model:text","Writes scripts and tags. Models not installed yet are downloaded automatically","g"],
|
||||
["vision_model","Ollama vision model","model:vision","Checks generated scenes on made-for-kids profiles. moondream suits 16 GB machines","g"],
|
||||
["vision_model","Ollama vision model","model:vision","Checks every scene on made-for-kids profiles. llava:7b is recommended, moondream is too weak to judge anatomy","g"],
|
||||
["ollama_timeout","Ollama timeout (seconds)","select:120|300|600|900","Max wait per Ollama request before retrying","g"],
|
||||
["aspect_ratio","Aspect ratio","select:9:16|16:9","","g"],
|
||||
["target_seconds","Target length (seconds)","select:30|45|60|90|120|180","Split into scenes of the length below","g"],
|
||||
|
||||
Reference in New Issue
Block a user