Switch to Wan 2.2 A14B + obsxrver LightX2V distill LoRAs (high/low per transformer)

Update app.py

Update ip_adapter.py

Init pipeline inside run_inference if pipe is None — ZeroGPU workers don't share globals

Fix AttributeError in WanIPAttnProcessor — use getattr for norm_q and inner_dim

Remove separate face reference input — input image doubles as face reference

Clean up stale /data/hf_home on startup to recover from 50G storage eviction

Revert HF_HOME redirect — was causing 50G storage eviction (Wan model is in base image, not /data/)

Redirect HF cache to persistent /data/hf_home to avoid re-downloading 30GB+ model on every cold start

Fix MODEL_ID: Wan2.1-I2V-14B-720P-Diffusers

Pivot to Wan 2.1 + IP-Adapter face conditioning

Lazy LoRA download + face reference conditioning

Remove UnifiedHorusRA/Outfit_Transformations LoRA

Add warmup preload, RIFE boot fix, new LoRA repos

Add warmup preload, RIFE boot fix, new LoRA repos

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload requirements.txt with huggingface_hub

Upload app.py with huggingface_hub

Add identity-preserving BLINK LoRAs (NSFWcode, lopi999, Outfit), fix -H-/-L- pairing

Add identity-preserving BLINK LoRAs (NSFWcode, lopi999, Outfit), fix -H-/-L- pairing

Lower LoRA scale default 0.8->0.6, LightX2V low-noise fuse 1.0->0.7 for better identity preservation

Re-enable int8 T5 quant (frees VRAM), remove VAE slicing for faster decode

Enable VAE tiling+slicing to fix OOM during decode (T5 in bfloat16 uses more VRAM)

Switch to plain diffusers (load_into_transformer_2 merged upstream), drop fork + conversion patch

Switch to plain diffusers (load_into_transformer_2 merged upstream), drop fork + conversion patch

Patch _convert_non_diffusers_wan_lora_to_diffusers: make head keys optional

Add loras.txt catalogue

Add SageAttn + LightX2V distill LoRA + remove text enc quant + flow_shift=6

Add SageAttn + LightX2V distill LoRA + remove text enc quant + flow_shift=6

Add SageAttn + LightX2V distill LoRA + remove text enc quant + flow_shift=6

Increase default steps 6->20 (prompt adherence), max steps 30->50, GPU cap 300->600s

Revert guidance_scale to 1.0 and flow_shift to 3.0 (AOTI built for single-pass mode)

Fix guidance scale defaults (1→5), set_adapters per-transformer, flow_shift 3→6

Add extra LoRA repos from UnifiedHorusRA and obsxrver

Fix AOTI+LoRA: disable compiled blocks for LoRA path, adaptive duration

Patch TorchaoLoraLinear for dynamic LoRA on fp8 model

Fix gr.Video buttons kwarg - not supported in gradio 5.29

Add RIFE model/ directory files

Add RIFE model/ directory files

Add RIFE model/ directory files

Update packages.txt - r3gm base with LoRA gallery

Update requirements.txt - r3gm base with LoRA gallery

Update aoti.py - r3gm base with LoRA gallery

Update app.py - r3gm base with LoRA gallery