Lightricks / LTX-2.5-22b-LoRA-Cinemagraph

Cinemagraph

Transforms still images into smooth, looping cinemagraphs with selective motion while keeping the rest of the scene frozen.

It is based on the LTX-2.5 foundation model.

Example Outputs

Model Files

ltx-2.5-22b-lora-cinemagraph-0.9.safetensors

Model Details

• Base Model: LTX-2.5-22B Video
• Training Type: LoRA
• Modality: image-to-video

Intended Use & Out-of-Scope

Intended use: Creating cinemagraph videos from still images — producing looping videos with selective motion (water, light, elements) while keeping the rest of the scene static. Ideal for creating animated wallpapers, social media content, and artistic motion pieces.

Out of scope: Scenes with camera movement, character/body movement, complex dynamic scenes, videos with multiple moving subjects. The LoRA is optimized for static-camera, single-element-motion scenarios.

How It Works

This LoRA fine-tunes LTX-2.5 to generate cinemagraph-style videos from still images. It excels at creating smooth, continuous-loop animations where only specific elements move (water flowing, neon lights pulsing, rain falling, reflections shimmering) while the rest of the frame remains completely frozen. The model learns to respect spatial isolation — keeping backgrounds, people, and static objects completely still while animating only the designated element.

The training dataset consisted of "no moving people" scenarios with locked-off cameras, so the LoRA strongly respects composition and avoids unwanted motion outside the intended subject.

Usage

🔌 ComfyUI

1. Copy the LoRA weights into models/loras .
2. Load the LTX-2.5-22B base model.
3. Add ltx-2.5-22b-lora-cinemagraph-0.9.safetensors as the LoRA.
4. Use the T2V/I2V workflow from the ComfyUI-LTXVideo repository .
5. Start with LoRA strength 1.0–1.2 (the sweet spot) and adjust within 0.5–1.5 per generation.
6. Include the trigger word CINEMAGRAPH_MOTION in the positive prompt.

Recommended Settings

• LoRA strength / weight: 0.5–1.5 (sweet spot: 1.0–1.2 )
• Spatial Guidance (STG): Use stg_v mode, scale 1.0 , targeting block 29 for enhanced motion control and boundary preservation.
• Inference steps: 30 (from training validation)
• Guidance scale: 4.0
• Prompting: See examples below. The model responds well to explicit descriptions of what should move and what should remain still. Emphasize "locked-off static camera" and "seamless natural loop" for best results.

Example positive prompt:

CINEMAGRAPH_MOTION, tripod locked-off static camera, zero camera movement, only the neon sign flickers, the starburst, Desert, MOTEL, and NO VACANCY neon tubes subtly pulse and flicker like a real vintage neon sign, everything else stays perfectly still, seamless natural loop

Example negative prompt:

cars moving, camera movement, pan, tilt, zoom, parallax, whole image moving, background sliding, person moving, flicker on entire image, noisy texture, crawling texture, distorted, blurry, low quality

Tips & Troubleshooting

• Explicit motion boundaries: Be very specific about what moves and what doesn't. "Only the [element] moves; everything else frozen" works better than vague descriptions.
• Avoid camera language: The LoRA was trained on static-camera footage. Omit phrases like "camera movement," "pan," "zoom," and "parallax."
• Seamless loops: Use "seamless natural loop" or "continuous looping motion" in the prompt to encourage cyclical, repeating movement.
• Strength tuning: The usable range is 0.5–1.5 . 0.5 is conservative; 1.0–1.2 is the sweet spot; pushing toward 1.5 drives selective motion more aggressively. Test within the range for your use case.
• Resolution: Best results at 512×704 pixels, 25 frames (the training resolution).

Dataset

Cinemagraph (No Moving People) — A carefully curated collection of still images and video references designed for cinemagraph creation. The dataset emphasizes scenes with selective motion elements (flowing water, flickering neon, falling leaves, shimmering light reflections) while keeping characters, backgrounds, and static objects completely frozen. All scenes were shot with locked-off, tripod-mounted static cameras to eliminate parallax and camera drift.

Training

• Technique: LoRA fine-tuning on LTX-2.5-22B using the LTX-2 Community Trainer .
• Hyperparameters: learning_rate=0.0001, batch_size=1, gradient_accumulation_steps=1, optimizer=adamw8bit, scheduler=cosine, rank=32, alpha=32, dropout=0.0
• Steps: 1000
• Infrastructure: LTX-2 Community Trainer .

References

• Code: GitHub Repository
• ComfyUI: ComfyUI-LTXVideo

License

See the LTX-2-community-license for full terms.

Acknowledgments

• Base model by Lightricks
• Training infrastructure: LTX-2 Community Trainer