Add 10Eros v1 as selectable model (FP8 full checkpoint, not LoRA)

Add 10Eros v1 FP8 mixed LoRA to LTX_ASSETS

Update app.py

Update app.py

Update app.py

Update app.py

Disable Triton in worker to fix ZeroGPU _cuda_init block during generation

Fix ZeroGPU GPU access: switch from subprocess.Popen to multiprocessing.Process

Switch to ZeroGPU POC pattern: duration=1 billing, subprocess retains CUDA context

Update app.py

Update app.py

Upload 3 files

Set GPU duration to 120 seconds

Run prompt enhancer before GPU generation

Load Sulphur prompt enhancer via pretrained GGUF

Replace LTX enhancer with Sulphur GGUF enhancer

Use Sulphur LoRAs and strengthen prompt handling

Load raw Gemma weights with Sulphur key mapping

Defer LTX model loading to GPU worker

Alias quantized Gemma weights for LTX loader

Align LTX Space runtime dependencies

Avoid xformers Dynamo import in LTX app

Avoid torch dynamo import under ZeroGPU

Set Sulphur LoRA strength to one

Replace Wan2GP wrapper with direct LTX pipeline

Stream WanGP events and cache uploads

Remove ZeroGPU duration and log WanGP startup

Prefetch LTX distilled LoRA at startup

Prefetch WanGP runtime assets at startup

Fix WanGP LTX model selection

Fix ZeroGPU startup CUDA initialization

Warm WanGP session at startup

Capture stderr (tqdm/progress) in log stream alongside stdout

Persist WanGPSession across generations; eliminate per-call model reload

Replace Frames slider with Seconds (1-5s); convert to frames internally at 25fps

Single-quote 'Experimental' in subtitle

Update subtitle to reflect experimental LoRA v1

Pre-download ltx-2.3-22b-distilled-lora-384 required by experimental LoRA

Switch to experimental Sulphur LoRA v1

Pre-download Gemma at startup; bump guidance to 5.0 for prompt adherence

Reduce GPU duration to 120s (H200)

Flatten hf_hub_download subfolder so LoRA lands where Wan2GP expects it

Use correct Sulphur distill LoRA from SulphurAI/Sulphur-2-base/distill_loras/

Stream tqdm progress lines by reading chunks and splitting on \r

Stream subprocess output line-by-line; increase GPU duration to 360s

Fix: don't use rank-768 LoRA alongside full distil model; capture stderr

Use generate.py subprocess approach like the original — no direct WanGP imports in app

All deps in requirements.txt; remove all runtime pip install and patch logic

Patch Wan2GP source to make rembg import optional after clone

Install Wan2GP deps via filtered requirements in setup() — no more whack-a-mole