roadready-ltx23-iclora

RoadReady — Road Cleanup IC-LoRA for LTX-2.3

Video-to-video IC-LoRA that removes road surface damage from automotive footage. Feed it a driving shot with cracked, stained, or patched asphalt and it returns the same shot with clean pavement. The car, shadows, environment, and reflections stay put. No masks, no roto.

Built by an automotive commercial editor to streamline road cleanup, one of the most common and budget-contested fixes in car spots. Entered in the LTX LoRA Jam (Utility).

Files

• roadready_iclora_ltx23_step2000_rank32.safetensors — the LoRA (rank 32, step 2000)
• comfyui_api_workflow_roadready.json — ComfyUI workflow, drag-and-drop
• LTX-2-Community-License.txt — license (this LoRA is usable only with the LTX model, per the Community License)

Base model — important

Trained on LTX-2.3 22b (dev) with the official LTX Trainer. Renders tested against ltx-2.3-22b-distilled-1.1 fp8, which fits a 24 GB card. Not trained or tested on LTX-2.5.

How to run (ComfyUI)

• Load the workflow JSON. It expects the distilled-1.1 fp8 transformer, the LTX-2.3 video VAE, and the gemma_3_12B text encoder with the ltx-2.3 text projection.
• Load this LoRA via LoraLoaderModelOnly at strength 1.0.
• Your damaged clip enters as the in-context reference through LTXVAddGuide (frame_idx 0, strength 1.0). Conform it first: 960x544, 24 fps, 8n+1 frames (49-97), e.g. ffmpeg -i IN.mov -vf "scale=960:544:force_original_aspect_ratio=increase,crop=960:544" -r 24 -frames:v 97 -c:v libx264 -crf 15 -pix_fmt yuv420p out.mp4
• Distilled sampling recipe: 8 steps, cfg 1.0, euler_ancestral, linear_quadratic.
• Caption template: ROADREADY road cleanup. {one-line scene description}. The cracked, stained, damaged asphalt is restored to a clean intact road surface while lane markings, vehicles, shadows, reflections, and surroundings stay unchanged.
• ~48 s per 97-frame clip on an RTX 4090.

Training

49 aligned damaged/clean pairs at 960x544/24fps. Most pairs built by compositing damage onto clean footage the author shot, so input and target are pixel-aligned by construction; six pairs are real winter damage with hand-restored clean targets. Rank 32, 3000 steps on one H100; step 2000 selected over step 3000 on measured lane-marking fidelity and asphalt tone.

Known behavior

• On heavily damaged surfaces the restored area can render darker than the surrounding road (reads as new asphalt rather than an artifact).
• Lane markings hold up in most shots; occasional saturation boost on double-yellows.
• Trigger word ROADREADY must be in the prompt.

License

LTX-2 Community License (included). The LoRA functions only with the LTX model and all use is subject to that license.