Refine & Restore

Refine & Restore combines two generative video-to-video tools:

• Restore cleans and colorizes damaged archive footage, including low-resolution transfers, compressed broadcast video, tape, sepia, and black-and-white film scans.
• Refine Details rebuilds fine texture, edges, and grain in soft, compressed, generated, or upscaled footage.

Both adapters use the LTX-2.5 LTX-2.5_V2V_TiledFusion_Upscale.json workflow. For damaged archive footage, run Restore first and Refine Details second.

Choose a path

Prerequisites

• A current ComfyUI installation with the LTX video nodes required by the workflow, including LTXVTiledFusionSampler and LTXVGetTilingSizes.
• The LTX-2.5 distilled model stack selected by the workflow.
• The adapter for the selected path, placed in ComfyUI/models/loras/.
• Progressive source frames. Deinterlace archive footage before using Restore.

Model files

The Tiled Fusion Upscale workflow uses the following LTX-2.5 model stack:

The published workflow selects Refine Details. To run Restore, select the Restore adapter in Load Models while keeping the LTX-2.5 distilled stack. Restore was trained on LTX-2.3 and its model card documents testing on the unchanged weights with the LTX-2.5 distilled transformer, video VAE, and Gemma-4 text encoder.

Refine Details

Use Refine Details when the source composition and color are already usable but fine detail is soft or missing.

Run the workflow

1. Load LTX-2.5_V2V_TiledFusion_Upscale.json in ComfyUI.
2. In Load Video, select the clip to refine.
3. Confirm Load Models selects ltx-2.5-22b-ic-lora-refine-details-1.0.safetensors.
4. In Preprocess → Get Tiling Sizes, choose the output size. The workflow resizes the clip to that canvas before attaching it as the IC-LoRA guide.
5. Enter a prompt about the rendering only: sharpness, texture, grain, lighting, palette, or grade. Do not name specific objects, people, or props; every spatial tile receives the complete prompt, so named content can be repeated across tiles.
6. Run the workflow and inspect fine detail, text, faces, structured edges, and tile boundaries through the full clip.

The adapter is trained for a 1024 × 576 working tile and should use the tiled workflow even for a full-HD canvas. Larger canvases remain one shared latent and noise field; overlapping crops are fused after every denoising step.

Restore archive footage

Use Restore for damaged archive sources that need cleanup, reconstruction, or inferred color before detail refinement.

Prepare the source

• Deinterlace interlaced or telecined footage before restoration.
• Do not denoise or sharpen the source first.
• Preserve the source aspect ratio. Do not stretch 4:3 footage to 16:9.
• Use a canvas whose width and height are multiples of 32 and a frame count of the form 8n+1.

Run the restoration pass

1. Load the Tiled Fusion Upscale workflow and select the archive clip in Load Video.
2. In Load Models, replace the Refine Details adapter with ltx-2.5-22b-ic-lora-restore-1.0.safetensors.
3. Open Preprocess and set a custom working canvas for the source aspect ratio. The Restore model card recommends 1440 × 1056 or 1440 × 1088 for 4:3 material and 1920 × 1088 for 16:9 material, using a 960 × 544 sampling tile.
4. Describe the period, place, lighting, materials, clothing, and intended natural color in the positive prompt. Put modern objects, logos, signage, lettering, or anything else the model must not invent in the negative prompt.
5. Run the workflow and review color consistency, faces, lettering, period details, and reconstructed low-information regions.

Restore is a valid result on its own. For a 4K-class delivery, run the restored output through the same workflow again with Refine Details selected, a 1024 × 576 sampling tile, and the 4K output size. The order matters: refining first can sharpen the damage and reduce color.

How the workflow handles video

The source video is resized to the output canvas and attached as an IC-LoRA guide at downscale factor 1. Tiled Fusion runs overlapping spatial crops within one shared latent canvas and merges them during every denoising step.

The published workflow uses streaming guide windows for longer clips. The source audio is held fixed during generation, decoded with the audio VAE, and muxed into the saved video at the source frame rate; neither adapter is trained to generate new audio.

For the full sampler contract and settings, see LTXVTiledFusionSampler in the nodes reference. For adapter-specific advanced techniques—including placed reference images and chained temporal windows—see the Refine Details and Restore model cards.

Limitations

• Restore infers color and missing structure from the source and prompt. Period-plausible content is not necessarily recovered historical truth.
• Refine Details rebuilds texture; it does not guarantee pixel-accurate restoration, faithful lettering, identity preservation, or artifact removal.
• Full HD, 4K, and 8K are available in the workflow selector, but higher resolutions increase memory use and runtime. Test the target size and full clip on the production hardware.
• Both adapters are designed for tiled use at their trained spatial windows. A single untiled full-HD pass places them outside those training buckets.
• This guide documents the ComfyUI workflow. Do not infer a native Python path from the graph.

Related pages

• Native Resolution
• LTX ComfyUI Nodes
• IC-LoRA
• Prompting guide