CoolFace
Apppublic

WaveCut/ltx-2.3-upscale-ic-lora

sourceHugging Faceupdated 3mo agoView on Hugging Face
0likes
App README

LTX 2.3 Distilled Upscale IC-LoRA Space

Upload a short video and upscale/refine it using the Zlikwid/LTX_2.3_Upscale_IC_Lora In-Context LoRA applied to diffusers/LTX-2.3-Distilled-Diffusers.

How it works

  1. 1.The input video is inspected before inference so oversized jobs fail early.
  2. 2.The app samples an 8k+1 frame count instead of decoding the whole upload blindly.
  3. 3.Frames are optionally center-cropped with Zoom.
  4. 4.Frames are bicubic-upscaled to the target working resolution.
  5. 5.The bicubic-upscaled video is passed both as an IC-LoRA reference condition and as an anchored video condition. The video condition keeps the generated latents tied to the source frames.
  6. 6.The output is encoded at an FPS that preserves source duration when metadata is available.
  7. 7.Source audio is muxed back into the output when Preserve Source Audio is enabled.

This follows the upstream LoRA instructions from Zlikwid/LTX_2.3_Upscale_IC_Lora:

  • —prompt is fixed to upscale
  • —LoRA weight is ltx2.3_upscale_ic-lora_06250.safetensors
  • —the reference video is bicubic-upscaled before the LTX-2.3 pass
  • —the LoRA metadata reference_downscale_factor=2 is preserved

The distilled LTX checkpoint requires:

  • —num_inference_steps=8
  • —guidance_scale=1.0
  • —sigmas=DISTILLED_SIGMA_VALUES

Those values are fixed by the app instead of being exposed as free sliders.

Tested ZeroGPU Limits

  • —Input duration: up to 12 seconds
  • —Upload size: up to 200 MB
  • —Output frame pixels: up to about 1024x576 / 768x768
  • —Sampled frames: up to 65
  • —Queue concurrency: 1 GPU job at a time

Controls

  • —Processing Preset: Sets safe defaults for resolution, sampled frames, and LoRA strength.
  • —Output Resolution: Choose a target resolution. Use Source × Bicubic Factor to derive the output size from the cropped/source video and the bicubic factor.
  • —Bicubic Pre-Upscale Factor: Multiplier for Source × Bicubic Factor mode.
  • —Zoom: Center crop before upscaling.
  • —Refinement Strength (LoRA): Controls the IC-LoRA adapter weight.
  • —Input Preservation Strength: Anchors the output to the bicubic input. Raise it if the result drifts or corrupts the source; lower it if you want more generative detail.
  • —Sampled Frames (8k+1): Number of frames sent to LTX. The output FPS is derived from source duration.
  • —Preserve Source Audio: Reuses the input audio track in the final MP4.
This is a generative refinement LoRA, not a simple deterministic upscaler. The output may differ from the input depending on settings.