AI video outpainting
Expand video beyond the original frame while preserving quality, motion, and creative intent. Built for production pipelines that reframe at scale, delivering full-length clips, native 1080p, and faithful output.
.svg)
Key capabilities
API-native architecture
Usage-based pricing, multi-GPU inference, and direct integration into your pipeline. No UI, no manual steps. Scale reframing across hundreds of videos without touching an editor.
Full-length generation
Generate up to 60 seconds of outpainted video at native 1080p. No duration caps, no resolution drops. The original clip stays intact while the model extends the canvas around it.
Aspect-ratio reformatting at scale
Reformat one master for every social platform and aspect ratio. Outpaint 16:9 into 9:16, 1:1, or 4:5 without cropping the original footage. Batch the treatment across a campaign.
Ad-tech & creative automation
One campaign video becomes every aspect ratio the channels need. Send it once to the API, get back multiple outputs programmatically. Scale reframing across hundreds of placements without manual touch.
Enterprise creative operations
Reformat long-form footage to new aspect ratios. Many competing solutions either cap clip length or reduce quality on longer videos. This API supports full-length outpainting for videos up to 60 seconds.
E-commerce & creative platforms
Extend product hero videos and lifestyle content beyond the original frame. Open up new canvas for text, graphics, or secondary elements. All at native 1080p, all ready for conversion-optimized placements.
How LTX-2.3 outpainting works
Input
Technical characteristics:
- Source video — The video to outpaint.
- Target aspect ratio — The new canvas region to generate (16:9, 9:16, 1:1, 4:5, and more).
- Prompt and model — Describe the new canvas areas (optional; most outpainting works without prompting). LTX-2.3-22B + ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors.
- Pipeline — Available via Reframe API or self-hosted IC-LoRA workflow. Two-stage: coarse generation followed by boundary refinement.
Output
Technical characteristics:
- Outpainted video — The original frame is completely preserved. The model extends the canvas and generates only the new canvas areas around the edges.
- Boundary quality — Seamless, artifact-free transition between original and generated content.
- Resolution — Native 1080p output. Full source length preserved (up to 60 seconds).
- Hardware requirements — Self-host via ComfyUI or the open-source pipeline on 80GB+ VRAM (or 32GB with FP8 quantization). API uses usage-based pricing with no hardware requirement.
Reframe
Expand a video to any aspect ratio without cropping, while preserving the original frame, motion, and quality.
- 720p output — $0.10/sec
- 1080p output — $0.20/sec
- Supports aspect ratio expansion including 16:9, 9:16, 1:1, 4:5, and more.
- Supports videos up to 60 seconds.
- Built on LTX-2.3
What users are building with LTX
FAQs
Do I need a mask?
Yes. The model is mask-conditioned: a binary mask defines the region to fill (inpainting) or the new canvas to generate (outpainting).
My removed object keeps coming back. Why?
The mask probably misses a causal cue. Dilate beyond the object and include its shadow, reflection, and contact areas. The model can reconstruct an object from a leftover shadow alone.
Should I prompt the whole scene?
No. Prompt only the masked or outpainted region. Describing content that already exists in the unmasked area can cause duplicated objects or people. Empty or minimal prompts work for most cases.
Can Reframe change aspect ratio without cropping?
Yes. That is outpainting. Extend the canvas vertically or horizontally and the model fills the new edges to match the footage.
What hardware and license do I need?
Open weights on HuggingFace under the LTX-2 Community License, requiring 80GB+ VRAM (or 32GB with FP8). The license is free under $10M ARR; enterprise licensing applies above that threshold.
Can I use Reframe via API?
Yes. The Reframe API is production-ready and available with usage-based pricing. For self-hosted deployments, use the open-source IC-LoRA workflow via ComfyUI or the LTX-2 pipeline on 80GB+ VRAM.
