LTX 梳理 2026 年 AI VFX 工具：LTX-2.5 九套开源工作流及适用场景

The most useful AI VFX tools in 2026 are the ones whose output keeps moving through a pipeline: an alpha matte you can composite, an ACEScg EXR sequence you can grade, a finished shot that still matches your layout camera. The LTX open-weights stack covers nine workflows built for that, most on LTX-2.5, from Alpha Gen for AI rotoscoping to SDR to HDR for gradeable masters. Below: each tool's job, inputs, and limits.

Prerequisites: These are open-source ComfyUI workflows, several with native Python paths. Running LTX-2.5 locally takes an NVIDIA GPU with 32GB+ VRAM (A100 80GB or H100 recommended) and CUDA 12.7 or higher per the LTX system requirements page; the LTX trainer docs recommend CUDA 13+ for best fine-tuning performance. Frame counts follow the VAE rule (F-1) % 8 == 0 (written 8k+1 in the VFX docs), so 121 frames is valid and 120 is not.

Which LTX-2.5 tool fits which job?

LTX-2.5 groups its AI VFX tools and VFX capabilities under the VFX & Post-Production docs, plus three IC-LoRA feature guides. Quick reference, from those pages:

• Alpha Gen (Beta): RGB video in, grayscale alpha matte out. ComfyUI and Python.

• Layout to Render: clay render or playblast plus an art-directed first frame in, finished shot out. ComfyUI and Python.

• SDR to HDR: 8-bit SDR video in, EXR sequence plus BT.2020/HLG 10-bit master out. ComfyUI and Python.

• Native HDR: EXR still or frame sequence in, color-managed EXR plus HLG master out. ComfyUI (image-to-video, inpainting) and Python.

• Refine & Restore: soft, compressed, or damaged footage in, rebuilt detail or restored color out. ComfyUI only.

• Native Resolution: any workflow-matched IC-LoRA task on a canvas up to 8K via Tiled Fusion. ComfyUI only.

• Relight: exterior footage in, footage relit to a chosen sun direction out. Validated on LTX-2.3, not yet on LTX-2.5.

• In/Outpainting: source video plus mask in, filled or extended frame out. LTX-2.3 IC-LoRA on the LTX-2.5 stack.

• Union Control: depth, canny, or pose from a reference video in, new content on the same structure out. LTX-2.3 IC-LoRA on the LTX-2.5 stack.

AI rotoscoping with Alpha Gen

Alpha Gen generates a frame-aligned matte from ordinary RGB video with no green screen, input mask, or prompt, and it is built to hold soft and semi-transparent detail like hair, fur, smoke, fire, sheer fabric, glass, and water. Use the matte as the alpha of your original plate (source_rgb × alpha + background × (1 - alpha)), not as a replacement beauty pass.

Source size tops out at 1920×1088, so pad 1080p plates to 1088 rather than stretching. Clips of 121 frames or fewer behave predictably, and above 145 frames RGB content can leak into the matte; the pipeline doesn't chunk for you, so split long shots into 121-frame segments and join the mattes in comp. And subject selection is semantic: the model picks the main subject, and you cannot redirect it with a prompt. If the shot has two candidate subjects, check frame one before committing the batch.

Layout to Render for previs

Layout to Render turns a Blender or Unreal viewport animation or blocky playblast into a finished shot. The layout drives camera path, framing, and object placement; an art-directed first frame painted over the layout's first frame drives materials, lighting, and palette.

Prompt wording: the docs warn against words like "clay", "3D", or "Unreal", which pull the output toward the blockout look. Write one or two sentences describing the finished shot. Use a 24 fps layout, keep width and height divisible by 64 (1920×1088 is the known-good size), and expect the workflow to snap length down to 8k+1, so a 72-frame layout returns 65 frames. Plan audio separately: ComfyUI output is silent and Python generates new audio rather than passing the source through.

SDR to HDR and Native HDR

SDR to HDR and Native HDR solve different problems, and picking the wrong one is the most common HDR mistake. SDR to HDR takes 8-bit SDR video and reconstructs highlight range that the SDR grade compressed out. It synthesizes that range; it does not recover original scene data. Output is half-float EXR by default (float32 optional) in scene-linear ACEScg, switchable to ACEScct, plus a BT.2020/HLG 10-bit master. No prompt, no mask: it applies to the whole frame.

Native HDR is for plates that are already HDR. It accepts EXR stills or folders of EXR frames declared as ACEScg, scene-linear sRGB/Rec.709, or ACEScct, and returns color-managed EXR plus an HLG master. HLG, PQ, or HDR10 .mp4 and .mov files are not accepted; convert them to EXR first. One gotcha for compositors: the ComfyUI loader keeps only the first three EXR channels, so alpha, extra channels, and header metadata do not round-trip. Keep any Alpha Gen matte as a separate element.

Refine & Restore for archive plates

Refine & Restore pairs two video-to-video adapters in one Tiled Fusion Upscale workflow. Restore cleans and colorizes damaged archive footage. Refine Details rebuilds texture, edges, and grain in soft, compressed, or upscaled footage. For archive plates, run Restore first and Refine Details second; refining first can sharpen the damage and reduce color.

Both are generative. They produce plausible detail, not pixel-accurate recovery, so treat rebuilt faces, lettering, and period color as interpretations. Two practical notes from the docs: Refine Details can denoise original grain, so plan to reintroduce grain if it matters, and 8K works best as 4K refine, Lanczos resize to 8K, then an optional gentle second pass.

Relight and inpaint plate fixes

Relight is an IC-LoRA that relights exterior footage to a chosen sun direction and hardness. A shaded "light-direction ball" composited into the top-right corner of every frame is the control signal, with a short caption naming one of 12 trained looks and one of eight directions. It was trained on exteriors only and is validated on LTX-2.3, so run it on the LTX-2.3 stack at its trained 1280×704, 121 frames, 24 fps. The Relight guide covers the ball geometry.

In/Outpainting uses one IC-LoRA for both jobs: a mask marks what to generate (white = generate, black = keep). Describe the full scene in the prompt, not the edit ("a horse walking down an empty country road", not "replace the car with a horse"), and size replacement masks for the new object. Green traces at mask edges come from the green composited under the mask; raise blend dilation or re-encode the mask losslessly.

Union Control for locked structure

Union Control is a single IC-LoRA that follows depth maps, canny edges, or pose skeletons extracted from a reference video, so output keeps the reference layout and motion while the prompt sets the look. Depth suits environments with clear depth layers, canny suits architecture and hard-surface work, and pose suits character motion. Strength 1.0 locks structure; 0.5–0.8 gives the model more freedom. Keep CFG between 1.0 and 1.5, since the distilled model bakes guidance in.

Which tool should you run first?

Tool choice depends on what you already have, not on which name sounds closest. Synthesizing across the docs, these AI VFX tools sort like this:

1. EXR plates already in ACES? Native HDR. SDR to HDR is for SDR sources.

2. SDR footage headed for an HDR grade? SDR to HDR last, after any Restore or Refine pass, so the reconstruction sees the final detail.

3. Need a matte? Alpha Gen on the RGB plate, kept as its own element because Native HDR won't carry it.

4. Blockout or previs? Layout to Render, with an art-directed first frame ready before you start.

5. Relight, inpaint, or structure transfer? Check the version first. Relight is LTX-2.3 only; In/Outpainting and Union Control reuse LTX-2.3 adapters on the LTX-2.5 stack as documented exceptions. Don't assume other LTX-2.3 adapters work on LTX-2.5.

Item 2's ordering is our recommendation from the documented inputs, not a documented pipeline.

What hardware do AI VFX tools need?

Hardware needs for AI VFX tools start at the LTX-2.5 floor of 32GB+ VRAM and climb with canvas size. The Alpha Gen model card points to H100/B200-class GPUs for full-HD upper-bound workloads. Native Resolution lists Full HD, 4K, and 8K in its selector but calls that "not a performance guarantee", so benchmark peak VRAM and runtime on your production card before committing a full clip. For sizing a farm, see the hardware guide for AI video models.

Summary

LTX-2.5's AI VFX tools are pipeline-shaped: Alpha Gen returns frame-aligned mattes, Layout to Render keeps the layout camera, SDR to HDR and Native HDR return EXR plus HLG masters, and Refine & Restore rebuilds archive and soft footage. Relight is still LTX-2.3 only, while In/Outpainting and Union Control run LTX-2.3 adapters on the LTX-2.5 stack. All of them are generative, so validate edges, color handling, and tile boundaries on representative frames first.

Ready to test one? Start with the tool that matches your next shot and run its published workflow unchanged on a 121-frame trim. For wiring generated elements into an existing pipeline, read integrating AI video into a VFX pipeline.