Apply Qwen-Image-Edit Lightning 8-step LoRA

Apply Qwen-Image-Edit Lightning 8-step LoRA

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Simplify camera UI to coupled 2D pad + linked yaw/pitch sliders

Backbone bf16 + fp32-recast hook at head boundary; keep DA3 heads fp32 to fit pack and stay dtype-consistent

Shrink only DA3 backbone to bf16 (keep heads fp32) to fit 76GB pack while preserving correctness

Cast DA3 priors to bf16 but keep norm layers fp32 (fixes LayerNorm dtype error) to fit 76GB pack

Cast DA3 priors to bf16 so full resident set packs into the 76GB NVMe offload dir

TEMP: probe /data-nvme exact bytes

TEMP: probe per-model resident byte sizes

Point ZeroGPU offload dir at /data-nvme block device

Redirect ZeroGPU offload dir to /dev/shm to fit the 78GB tensor pack; drop diagnostics

TEMP: probe 90GB fallocate per filesystem

TEMP: probe O_DIRECT/fallocate support per filesystem

TEMP: add df -h diagnostic for offload dir sizing

Adopt standard ZeroGPU recipe: .to('cuda') at module scope, drop _ensure_cuda

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

Initial MetaView novel view synthesis demo

initial commit