Bigger fast-swap time window (400/500s)

Revert to CPU onnxruntime (gpu variant didn't win over insightface's onnxruntime dep on ZeroGPU)

Fast swap: detection+recognition only at 320px (5-8x CPU speedup) to fit the GPU window

Raise fast-swap GPU time budget (GPU inswapper + first-run downloads)

onnxruntime -> onnxruntime-gpu for CUDA InsightFace

Fast swap: run InsightFace on GPU (CUDA EP) with CPU fallback — was hardcoded CPU, blew the ZeroGPU time window

Add Quick Start scenario picker with per-scenario guide-video requirement hints

Add Fast Swap mode (InsightFace inswapper) as default — deterministic, promptless, actually swaps (fastswap.py)

Add Fast Swap mode (InsightFace inswapper) as default — deterministic, promptless, actually swaps (app.py)

Expose guide-adherence (condition strength) slider; defaults 0.7/1.0 for actual head swap (pipeline.py)

Expose guide-adherence (condition strength) slider; defaults 0.7/1.0 for actual head swap (app.py)

Fix ZeroGPU lifecycle + V2V guide-video conditioning (README.md)

Fix ZeroGPU lifecycle + V2V guide-video conditioning (pipeline.py)

Fix ZeroGPU lifecycle + V2V guide-video conditioning (app.py)

Bump diffusers to >=0.34.0 for LTX 0.9.8 support

Fix: restore lora_strength default to 1.0 (was 0.0 from diagnostic run)

Fix: load full pipeline from distilled repo to fix VAE/scheduler version mismatch

diagnostic: default lora_strength to 0.0 to test base pipeline

Diagnostic: set BFS LoRA weight=0 to isolate base pipeline vs LoRA corruption

Fix scrambled output: switch output_type pt→np (no tensor permute guessing)

Add peft (required for load_lora_weights)

Add tiktoken + sentencepiece for LTX-Video tokenizer

Fix BLIP-2 auto-describe: decode only new tokens, not full prompt+answer sequence

Switch to LTX-Video-0.9.8-13B-distilled + ltx-2 BFS LoRA; fix 22B architecture mismatch

Wrap generate in try/except to surface real error in status box

Handle Gradio 5.x VideoData dict in generate(); extract plain path before use

Switch auto-describe from SmolVLM to BLIP-2 (fixes ImportError)

Fix auto-describe: use batch_decode instead of decode, add multi-marker assistant reply extraction

Switch auto-describe from Florence-2 to SmolVLM-Instruct (fixes forced_bos_token_id crash)

Add gfpgan, insightface, onnxruntime, timm dependencies

Add auto-describe (Florence-2), face alignment (InsightFace), GFPGAN enhancement, trim note

Update app title to FaceOff-FaceSwapper

Rename to FaceOff-FaceSwapper

Add gr.Examples with local face image (fixes Wikipedia 403)

Add example face image for gr.Examples

Upgrade Gradio sdk_version 5.7.1 → 5.50.0 (fixes VideoData schema validation error)

Enable show_error=True to expose runtime errors

Fix: monkey-patch gradio_client boolean schema bug causing /gradio_api/info 500

Fix build: remove gradio pins (let HF own), fix pydantic v2, fix progress param

Fix: update gradio to 5.7.1, gradio_client, pydantic v2

Update requirements.txt

Update app.py

Update app.py

Update requirements.txt

Update requirements.txt

Update app.py

Update app.py

Update app.py

Update app.py

Update app.py