> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs-dev.ltx.io/open-source-model/vfx-post-production/alpha-gen/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs-dev.ltx.io/_mcp/server.

# Alpha Gen (Beta)

> Generate frame-aligned alpha mattes from RGB video with the LTX-2.5 Alpha Gen IC-LoRA.

Alpha Gen takes an ordinary RGB video and generates a matching grayscale alpha matte without a green screen, input mask, or text prompt. It is designed to retain hard edges as well as soft and partially transparent detail such as hair, fur, smoke, fire, sheer fabric, glass, and water.

The matte is aligned frame-for-frame with the source video. The model identifies what it interprets as the main subject; selection is semantic rather than based on which object is closest to the camera. When used as an alpha channel, white keeps the model-selected area fully opaque, black makes the unselected area fully transparent, and gray produces partial opacity. Composite the matte with the original RGB footage rather than treating the generated video as a replacement beauty pass.

**Model card:** [LTX-2.5 22B IC-LoRA Alpha Gen](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Alpha-Gen)

## Supported surfaces

| Surface | Entry point                                                                                                                                                                                                             |
| ------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| ComfyUI | [`LTX-2.5_V2V_ICLoRA_Single_Stage_Distilled.json`](https://github.com/Lightricks/ComfyUI-LTXVideo/blob/master/example_workflows/2.5/LTX-2.5_V2V_ICLoRA_Single_Stage_Distilled.json) with the Alpha Gen adapter selected |
| Python  | `python -m ltx_pipelines.ic_lora` with `--skip-stage-2`                                                                                                                                                                 |

> **Note**
>
> A dedicated Alpha Gen ComfyUI workflow will be published in a follow-up update. Until then, use the general single-stage video-to-video IC-LoRA workflow listed above.

## Requirements

* One RGB source video no larger than 1920×1088. For the most predictable results, use 121 frames or fewer. Valid `8n+1` frame counts above 121 and up to 145 may work but can be less reliable; clips longer than 145 frames are unsupported.
* Spatial dimensions divisible by 32. Pad rather than stretch clips that do not meet this requirement; for example, pad 1920×1080 to 1920×1088.
* A frame count that follows `8n+1`.
* An empty prompt. Alpha Gen uses the RGB video as its only guide and does not support prompt- or mask-based subject selection.

## Model files

| File                                                      | Role                                                                          |
| --------------------------------------------------------- | ----------------------------------------------------------------------------- |
| `ltx-2.5-22b-distilled-transformer-bf16.safetensors`      | Distilled LTX-2.5 transformer                                                 |
| `gemma4-12b-with-proj-ltx-2.5-bf16.safetensors`           | LTX-2.5 text encoder required by the supplied surfaces                        |
| `ltx-2.5-video-vae-bf16.safetensors`                      | Video VAE                                                                     |
| `ltx-2.5-audio-vae-bf16.safetensors`                      | Audio VAE                                                                     |
| `ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors` | Spatial upscaler required by the Python pipeline interface                    |
| `ltx-2.5-22b-ic-lora-alpha-gen-0.9.safetensors`           | Released Alpha Gen adapter; use strength `1.0` as specified by the model card |

## Run with ComfyUI

1. Download the Alpha Gen adapter to `ComfyUI/models/loras/`.
2. Load the LTX-2.5 single-stage video-to-video IC-LoRA workflow linked above.
3. Select `ltx-2.5-22b-ic-lora-alpha-gen-0.9.safetensors` in the IC-LoRA loader and set its strength to `1.0`.
4. Load the RGB source video as the reference video.
5. Leave the prompt empty.
6. Run the workflow and save the generated matte video.

The graph saves the matte as a standard ComfyUI video object. It does not expose separate mask, RGBA, EXR, or alpha output sockets. Export or combine the matte with the original RGB clip in your compositing application.

## Run with Python

The native IC-LoRA pipeline uses its first stage at the source resolution and skips the 2× upscale. Because `--skip-stage-2` returns half of the requested CLI dimensions, pass `--width` and `--height` at twice the source size.

The following command is the model card's example for a 1920×1088 source. The frame count and seed are example values, not required ComfyUI defaults.

```bash
python -m ltx_pipelines.ic_lora \
  --transformer-path ltx-2.5-22b-distilled-transformer-bf16.safetensors \
  --text-encoder-path gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
  --video-vae-path ltx-2.5-video-vae-bf16.safetensors \
  --audio-vae-path ltx-2.5-audio-vae-bf16.safetensors \
  --spatial-upsampler-path ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
  --prompt "" \
  --width 3840 --height 2176 --num-frames 121 --frame-rate 24 \
  --video-conditioning input_1920x1088.mp4 1.0 \
  --lora ltx-2.5-22b-ic-lora-alpha-gen-0.9.safetensors 1.0 \
  --skip-stage-2 \
  --seed 1234 \
  --output-path matte.mp4
```

For another source size, double its padded width and height for the CLI arguments. Keep the source within the 1920×1088 output limit.

## Composite the result

Use the generated matte as the alpha channel of the original RGB video:

```text
output = source_rgb × alpha + replacement_background × (1 - alpha)
```

Confirm that white covers the area you want to keep visible and black covers the area you want to replace before rendering the final composite.

## How it works

Alpha Gen VAE-encodes the RGB reference video and uses those latents as IC-LoRA conditioning with an empty text prompt. The adapter runs for the full single-stage denoise, producing a grayscale matte at the source resolution and duration.

The model decides what it treats as the main subject. It does not accept a prompt or mask for choosing a particular person or object.

## Limitations and troubleshooting

* Use 121 frames or fewer for the most predictable results. Valid `8n+1` frame counts above 121 and up to 145 may work but can be less reliable. Above 145 frames, RGB content can leak into the matte.
* The published pipeline does not automatically chunk long videos. Split longer shots into segments of 121 frames or fewer, process each segment separately, and combine the matte outputs in post-production.
* Output above 1920×1088 is not supported by the model card.
* Full-HD generation requires substantial GPU memory. The model card identifies H100/B200-class hardware for upper-bound workloads; shorten the clip or reduce the resolution if you run out of memory.
* Review fine edges, translucent materials, motion blur, occlusion, and cuts before final use.
* Do not add a prompt to choose the main subject; prompting is not supported for this adapter.

## Related pages

* [VFX & Post-Production](/open-source-model/vfx-post-production/overview)
* [Workflow & Asset Compatibility](/open-source-model/reference/workflow-asset-compatibility)
* [IC-LoRA controls and adapters](/open-source-model/integration-tools/ic-lo-ra-adapters)
* [Using ComfyUI with LTX](/open-source-model/integration-tools/comfy-ui)

> Generate frame-aligned alpha mattes from RGB video with the LTX-2.5 Alpha Gen IC-LoRA.