falVideo to videoComing soon

fal-ai/ltx-2.3-quality/reference-video-to-video/lora

Name: fal-ai/ltx-2.3-quality/reference-video-to-video/lora
Brand: fal

fal-ai/ltx-2.3-quality/reference-video-to-video/lora is a video to video tool from fal that runs in the background and returns video. Coming soon: listed with its price, not yet open to runs.

This tool is listed but not yet available to run.

Price pending

Pricing

Price pending.

Limits

Runs: In the background
Returns: video
Longest queue wait: 3600s
Run deadline: 1800s
Largest inline input: 1024 KB
Largest input file: 500 MB
Largest output: 4768 MB
Outputs per run: 16

Input

Field	Kind	Default	Description
`prompt` required	text	—	The prompt to guide generation.
`video_url` required	URL or file	—	The URL of the reference video that supplies motion/structure.
`control_video_url`	text	—	Optional pre-computed control video (e.g. an already-rendered depth / edge / pose composite). When provided together with skip_control_preprocess, the built-in control estimation is skipped and this video is used directly as the control signal, resampled to the output resolution and frame count.
`skip_control_preprocess`	true or false	false	Skip the built-in control estimation (depth / edge / pose) and use control_video_url directly as the control signal. Requires control_video_url; ignored if it is not set. With resolution='auto', the output follows the source video size/aspect up to the LTX limits. Explicit resolutions are honored and 64-aligned so the RGB video and control video stay matched.
`preserve_original_video`	true or false	false	True video-to-video: the base video_url is used as the generation's starting point, while control_video_url still drives the structure. The amount of the original preserved is controlled by `strength` (denoise): lower `strength` keeps more of the original video's pixels. Requires skip_control_preprocess, video_url and control_video_url. Default off leaves the standard control path unchanged.
`image_url`	URL or file	—	Optional reference image for style/character anchoring. This is the FIRST-frame keyframe.
`mid_image_url`	text	—	Optional middle-frame keyframe image. When set, the look/appearance is anchored to this image around the middle of the clip (in addition to image_url at the start), so the appearance follows a moving camera instead of drifting. Requires image_url.
`end_image_url`	URL or file	—	Optional last-frame keyframe image. When set, the look/appearance is anchored to this image at the end of the clip. Useful for shots where the camera travels to a different place than the start image. Requires image_url.
`video_strength`	number 0 to 1	0.6	Video conditioning strength. Lower values give the model more freedom to change the reference video motion/structure.
`strength`	number 0 to 1	1	Sampler denoise strength for reference-video-to-video. With preserve_original_video on, this is the video-to-video amount: lower values keep more of the original video's pixels (e.g. 0.5 = keep ~50%), 1.0 fully regenerates. Without preserve_original_video it only trims the denoise schedule.
`num_frames`	number (whole) 9 to 481	121	The number of output frames to generate. This IC-LoRA endpoint caps the resolution x frames volume for stability (~150 frames at 720p-class; more at smaller sizes). A request above the budget returns a 422 — reduce num_frames or use a smaller resolution.
`resolution`	JSON	auto	The size of the generated video. In direct-control mode (skip_control_preprocess with control_video_url), 'auto' follows the source video size/aspect up to the LTX limits and explicit sizes are honored with 64px alignment. In the built-in control estimation mode, the output stays near the official control workflow size. Higher resolutions reduce the maximum stable frame count; lower resolutions can run longer.
`frames_per_second`	number 1 to 60	24	Frames per second of the generated video.
`num_inference_steps`	number (whole) 8 to 30	15	Number of inference steps. Defaults to 15 and can be increased up to 30.
`guidance_scale`	number 1 to 20	1	Classifier-free guidance scale. The default is tuned for fast, high-quality generation.
`generate_audio`	true or false	true	Whether to include audio in the returned video. When disabled, the final MP4 is returned without an audio track.
`negative_prompt`	text	color distortion, overexposure, static, blurry details, subtitles, style, artwork, painting, frame, still, dim overall tone, worst quality, low quality, JPEG compression artifacts, ugly, mutilated, extra fingers, poorly drawn hands, poorly drawn face, deformed, disfigured, malformed limbs, fused fingers, motionless frame, cluttered background, three legs, crowded background, walking backwards	The negative prompt to steer generation away from.
`seed`	number (whole)	—	Random seed for reproducibility. If None, a random seed is chosen.
`enable_prompt_expansion`	true or false	true	Whether to enable prompt expansion.
`video_quality`	one of low, medium, high, maximum	high	The quality preset of the generated video.
`video_write_mode`	one of fast, balanced, small	balanced	The write mode of the generated video.
`loras` required	JSON	—	Up to 3 LoRAs to apply on top of LTX-2.3. Each path is downloaded through the registry SSRF-safe downloader before it is loaded. Max size: 3 GB per LoRA.

Same request in code

curl https://api.spiral.ai/v1/runs \
  -H "Authorization: Bearer $SPIRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{
    "provider": "fal",
    "model": "fal-ai/ltx-2.3-quality/reference-video-to-video/lora",
    "input": {
      "prompt": "A dancer in a flowing red dress against a black backdrop, studio lighting, cinematic.",
      "video_url": "https://v3b.fal.media/files/b/0a8824b1/sdm0KfmenrlywesfzY1Y1_if6euPp1.mp4",
      "loras": {}
    }
  }'