Swap Wan2.1 -> HunyuanVideo-1.5, 544x960 portrait 9:16, 49 frames @ 24fps

Fix: use WanPipeline(Wan2.1-T2V-1.3B) directly, drop VACE GGUF to avoid layer mismatch

Swap LTX-Video -> Wan2.1-VACE-1.3B GGUF for better video quality

Increase LTX-Video steps 4->25 for better quality, extend GPU duration to 120s

Fix spaces import order, upgrade LLM to Qwen2.5-7B-Instruct on ZeroGPU

Add imageio and imageio-ffmpeg for export_to_video

Add sentencepiece and tiktoken for LTX-Video tokenizer

Remove debug instrumentation

Fix LLM: use local Qwen2.5-0.5B-Instruct via transformers pipeline, no external API

Fix LLM: switch to direct requests call on api-inference.huggingface.co, bypass router

Fix: switch chat_completion to text_generation with Mistral [INST] template

Fix LLM: use Mistral-7B (available on hf-inference), drop response_format, add JSON fallback

Fix: use provider=hf-inference to bypass router, no third-party provider needed

Restart Space

Switch LLM to Llama-3.2-3B-Instruct (available on HF serverless)

Swap to Qwen2.5-1.5B LLM and LTX-Video; output mp4; slim requirements

Perf: background preload RAG, thread-safe lazy loading, declare models in README for HF caching

Fix: lazy-load models and RAG dataset on first use to avoid init-phase network failures

Fix: load HF_TOKEN from Space secret for InferenceClient

Fix encoding: remove non-ASCII comments, ensure clean UTF-8 app.py with two tabs

Add PE tab with RAG+CoT generation alongside Basic LLMPopcorn

Fix: use hf_hub_download + load_file for AnimateDiff-Lightning

Initial commit for LLMPopcorn Demo

initial commit