WepeNerd/ltx-ripple-demo

Space

37

stars

17

commits

Sep 3, 2026

updated

gradio
mcp-server

README

LTX Ripple — First-Frame Edit Propagation

LTX Ripple is a First Frame All Frames (FFAF) IC-LoRA for LTX-2.5 by WepeNerd.

How it works

  1. Upload a video — its first frame is extracted and shown to you automatically.
  2. Optionally, describe an edit for that first frame (add effects, change style, replace objects, etc.) right below the frame preview. The frame is edited with conradlocke/krea2-identity-edit and the edited frame is shown back for confirmation before it is used as the first-frame conditioning input. Leave it empty (or upload your own edited frame) to use the plain extracted frame.
  3. LTX Ripple propagates the edit through the entire video while maintaining temporal consistency.

You can optionally provide a text description of the edit to help guide the model, but the LoRA can work without any prompt at all.

These mirror the reference ComfyUI workflow shipped with the LoRA (Workflow_example/LTX_Ripple_FFAF_Edit.json):

  • LoRA strength: 1.35 (increase if the edit isn't carried through, decrease if the output changes too much)
  • Distilled inference: 8 steps, euler, no CFG (guidance_scale=1.0)
  • Resolution: 1152×768 (single-stage distilled), lanczos resize + centre crop
  • Duration: 4 seconds by default → 97 frames at 24fps (up to 5 s / 121 frames)
  • Output: h.264 / yuv420p / crf 12 at 24fps

Credits

Distributed under the LTX-2.x Community License Agreement.

Contributors

multimodalart

14 commits

WepeNerd

3 commits

WepeNerd/ltx-ripple-demo

Space

37

stars

17

commits

Sep 3, 2026

updated

gradio
mcp-server

README

LTX Ripple — First-Frame Edit Propagation

LTX Ripple is a First Frame All Frames (FFAF) IC-LoRA for LTX-2.5 by WepeNerd.

How it works

  1. Upload a video — its first frame is extracted and shown to you automatically.
  2. Optionally, describe an edit for that first frame (add effects, change style, replace objects, etc.) right below the frame preview. The frame is edited with conradlocke/krea2-identity-edit and the edited frame is shown back for confirmation before it is used as the first-frame conditioning input. Leave it empty (or upload your own edited frame) to use the plain extracted frame.
  3. LTX Ripple propagates the edit through the entire video while maintaining temporal consistency.

You can optionally provide a text description of the edit to help guide the model, but the LoRA can work without any prompt at all.

These mirror the reference ComfyUI workflow shipped with the LoRA (Workflow_example/LTX_Ripple_FFAF_Edit.json):

  • LoRA strength: 1.35 (increase if the edit isn't carried through, decrease if the output changes too much)
  • Distilled inference: 8 steps, euler, no CFG (guidance_scale=1.0)
  • Resolution: 1152×768 (single-stage distilled), lanczos resize + centre crop
  • Duration: 4 seconds by default → 97 frames at 24fps (up to 5 s / 121 frames)
  • Output: h.264 / yuv420p / crf 12 at 24fps

Credits

Distributed under the LTX-2.x Community License Agreement.

Contributors

multimodalart

14 commits

WepeNerd

3 commits