Spaces:
Running on Zero
Running on Zero
Commit History
-
Fix resolution rounding: derive required multiple from named constants Running
The IC-LoRA reference-conditioning video must be divisible by STAGE_1_RESOLUTION_DIVISOR(2) * reference_downscale_factor(4) * VAE_SPATIAL_COMPRESSION(32) = 256, not the 128 the previous fix used. VAE_SPATIAL_COMPRESSION(32) is verified against the checkpoint's embedded VAE config (encoder_blocks in the safetensors metadata). The factor of 2 comes from ic_lora.py running stage 1 at half the requested output resolution before applying reference_downscale_factor to build the reference conditioning. RES_MULTIPLE is now computed from these named constants and the live pipeline's reference_downscale_factor, instead of a hardcoded literal, so a future LoRA with a different downscale factor won't silently reintroduce this crash. Also fixes the default width/height (1024x640 -> 1536x768) to match what on_video_upload actually computes for the bundled 1280x720 example clips.2c57808
-
Fix resolution rounding: use multiples of 128 for two-stage half-res VAE constraint
Verified signature2537fc3 -
Upload folder using huggingface_hub
Verified signature125b505 -
Upload requirements.txt with huggingface_hub
Verified signature4e74d1d -
Delete example_lowres.mp4 with huggingface_hub
Verified signature5d89a9f -
Upload folder using huggingface_hub
Verified signature8c86c53 -
Upload app.py with huggingface_hub
Verified signature5aa089a -
Upload app.py with huggingface_hub
Verified signature8a43e99 -
Upload app.py with huggingface_hub
Verified signature7df16e2 -
Upload app.py with huggingface_hub
Verified signature169354e -
Upload app.py with huggingface_hub
Verified signaturef2f2a4d -
Upload app.py with huggingface_hub
Verified signature50f04af -
Upload app.py with huggingface_hub
Verified signaturebcfe1f4 -
Add torchvision (Gemma-3 image processor dependency)
Verified signature15b6ed7 -
Initial LTX-2.3 IC-LoRA spatial upscaler app
Verified signature677b78a -
initial commit
Verified signaturee45c7c3
