FuzzPuppy/LTX-2.3-Black-Magic-LoRA

Black Magic: LTX-2.3 22B Shadow Reconstruction IC-LoRA

Black Magic is a video-to-video IC-LoRA for LTX-2.3 22B that transforms dark or underexposed footage into a plausible, visually rich interpretation of the hidden scene. It doesn’t simply brighten the video. It interprets what might be hidden in the dark.

Prompt

Restore this underexposed video to a natural, well-exposed version with realistic colors and recovered shadow detail. A woman performs poi fire dancing. She has her back to the camera and is facing an audience.

Prompt

Restore this underexposed video to a natural, well-exposed version with realistic colors and recovered shadow detail. A man and woman dance at a nightclub.

Prompt

Restore this underexposed video to a natural, well-exposed version with realistic colors and recovered shadow detail. A fire pit in front of a house in the forest.

Prompt

Restore this underexposed video to a natural, well-exposed version with realistic colors and recovered shadow detail. Three men in a snowy field.

Prompt

Restore this underexposed video to a natural, well-exposed version with realistic colors and recovered shadow detail. A monkey eats corn on the cob.

Prompt

Restore this underexposed video to a natural, well-exposed version with realistic colors and recovered shadow detail. An owl in the forest.

Prompt

Restore this underexposed video to a natural, well-exposed version with realistic colors and recovered shadow detail.

Prompt

Restore this underexposed video to a natural, well-exposed version with realistic colors and recovered shadow detail. A porcupine under a rock.

Prompt

Restore this underexposed video to a natural, well-exposed version with realistic colors and recovered shadow detail. A tiger in the forest.

Black Magic is a generative VFX model, not conventional low-light enhancement. It does not claim to reveal the original signal faithfully. Where the input contains too little information, it creates a temporally coherent interpretation of what could be in the dark.

Model File

• black-magic-ic-lora-450.safetensors

Model Details

• Base model: LTX-2.3 22B
• Training type: IC-LoRA (video-to-video)
• Reference input: dark video
• Reference downscale factor: 1
• Recommended LoRA strength: 1.0–1.25

Intended Use & Out-of-Scope

Intended use: Generative shadow reconstruction for short creative and VFX shots. The model is designed to preserve visible subjects, composition, motion, and light placement while imagining plausible detail in crushed shadows. This makes Black Magic especially suited to dark concerts, festivals, animals in the wild at night, nighttime footage, and other shots where a compelling reconstruction matters more than accuracy.

Out of scope: Faithful photographic restoration, scientific enhancement, surveillance, forensics, or any use that requires hidden details to match the original scene. Indiscernible content is generated, not recovered.

Usage

ComfyUI

A ready-to-run, two-phase ComfyUI graph is included: ltx23-black-magic-lora-workflow.json. It uses ComfyUI core nodes and the official Lightricks ComfyUI-LTXVideo package. No Black Magic custom nodes are required.

1. Put black-magic-ic-lora-450.safetensors in ComfyUI/models/loras.
2. Install or update the official Lightricks ComfyUI-LTXVideo nodes.
3. Load ltx23-black-magic-lora-workflow.json.
4. Select the dark source in the Load Video node.
5. Update the positive prompt with a short scene description that guides the reconstruction (see below).
6. Queue the workflow.

Prompting

The following is the instruction used during training and is the recommended prompt prefix:

Restore this underexposed video to a natural, well-exposed version with realistic colors and recovered shadow detail.

Append a short description of what the scene contains, or what Black Magic should plausibly construct in the shadows:

A tiger in a forest.

The description acts as creative direction. Keep it concise if you want the visible reference to remain dominant; add more detail when the shadows are nearly empty and you want to steer the reimagined content. Long or strongly stylized prompts can intentionally pull the result farther from the source and may produce a painted look.

Recommended Settings

• Black Magic strength: 1.0–1.25
• Distilled LoRA strength: 0.5
• Phase one: 8 distilled Euler steps at approximately 640×352 for 16:9 video
• Phase-one guidance: video CFG 3, audio CFG 1, modality guidance 2, rescale 0.7
• Phase two: CFG 1, seed 43, sigmas 0.85, 0.7250, 0.4219, 0
• Output: 1280×704 for the included 16:9 workflow

The included workflow safely decodes up to 129 output frames in one LTX VAE temporal tile. For a longer valid LTX sequence, set:

minimum temporal_size = output frame count + 7

Use frame counts that are one more than a multiple of eight, such as 81, 121, or 129. Keep input dimensions divisible by 32.

Tips and Limitations

• Visible evidence of subjects, composition, or motion gives the model stronger anchors. Fully black regions leave more room for invention.
• A concise scene description steers what the model imagines in ambiguous shadows.
• Re-running with another seed can produce a different but still plausible interpretation of the same darkness.
• This LoRA reconstructs video appearance; it was not trained as an audio model.

Dataset

The model was trained on video from Pexels and BVI-RLV. Pexels videos were synthetically darkened to create aligned reference and target pairs and are subject to the Pexels license.

Real low-light pairs came from BVI-Lowlight: Fully registered datasets for low-light image and video enhancement by P. Anantrasirichai, A. Malyugina, R. Lin, and D. R. Bull (2023), available under CC BY 4.0 from IEEE DataPort. The clips were cropped, resized, and temporally sampled for training.

Training

• Technique: IC-LoRA (rank 32, alpha 32) on the LTX-2.3 22B video transformer
• Checkpoint: step 450
• Infrastructure: LTX-2 Community Trainer

License

The model weights are released under the LTX-2 Community License. The source datasets remain subject to their respective licenses described in the Dataset section above.

Acknowledgments

• Lightricks for LTX-2.3, the official ComfyUI nodes, and the LTX-2 Community Trainer.
• The creators whose videos from Pexels and Mixkit were used as the reference videos to produce the model-card examples.
• The Pexels creators and the BVI-RLV authors for the source training material.