cfryant/blackout-ic-lora-ltx23

Blackout IC-LoRA for LTX-2.3 22B (trigger: BL4CKOUT)

A video-to-video IC-LoRA (in-context reference LoRA) for Lightricks LTX-2.3 22B dev. Give it a normal, lit clip as the reference video and it returns the same shot with every light switched off. The only light left is a flashlight beam from the camera's point of view, with deep black surroundings. Camera motion and action stay the same.

Files

file	version	notes
blackout_ic_lora_ltx23_v2_r32_step3000.safetensors	v2 (recommended)	46 pairs at 1280x704x89 @ 24 fps, rank 32, 3000 steps, official Lightricks LTX-2 trainer
blackout_ic_lora_ltx23_v1_r32_step3000_fal.safetensors	v1	21 pairs at 960x544x89 @ 24 fps, rank 32, 3000 steps, fal ltx23-trainer-v2/ic-lora/v2v

Both are bf16 safetensors, about 327 MB each. checkpoints_v2/ holds the v2 intermediate checkpoints (every 250 steps, 250-2750) for comparison.

Usage

• Base model: Lightricks/LTX-2.3, ltx-2.3-22b-dev.safetensors
• Mode: IC-LoRA / reference video-to-video. The reference (control) video is the original lit clip at full resolution (reference downscale 1).
• Suggested strength: 0.8. Try 1.0 if some lights stay on, 0.6-0.7 if the result gets too dark.
• Prompt: start with the trigger BL4CKOUT. Example: BL4CKOUT same scene, every light switched off, only a camera-mounted flashlight from the viewer's point of view, same camera motion and action.
• Negative prompt (optional): lights on, bright lighting, lit lamps, glowing ceiling lights, lit screens, daylight, sunlight, flat even lighting, overexposed, visible flashlight, flicker, jitter, warped geometry, changed camera motion
• Resolution: v2 was trained at 1280x704, so 1280x720 / 1280x704 at 24 fps works well.
• fal: works with fal-ai/ltx-2.3-22b/reference-video-to-video/lora (loras: [{path: <resolve URL of this file>, scale: 0.8}]).

Training details (v2)

• Trainer: Lightricks/LTX-2 packages/ltx-trainer, flexible strategy with a video reference condition (first-frame conditioning p = 0.1)
• Data: 46 pairs. Reference = original lit clip; target = the same clip relit as a blackout with a camera flashlight. 89 frames at 1280x704, 24 fps.
• LoRA rank 32 / alpha 32, learning rate 1.5e-4 with cosine schedule, 3000 steps, bf16, 1x H100 80GB (about 8.3 s/step)
• Captions: scene description plus a fixed blackout description; the trigger BL4CKOUT is prepended.

Limitations

• Trained on indoor and urban scenes with practical lights (corridors, bars, shops, venues). Unusual light sources or very bright exteriors may keep some light.
• The flashlight beam always comes from the camera position.

License

These LoRA weights are derivatives of LTX-2.3, so they are distributed under the LTX-2 Community License Agreement (LICENSE-2). Use is subject to its terms.