LTX-2.5 Restore and Refine Details: Free AI Video Restoration for Old and Soft Footage (Plus a No-GPU Way to Try It)

Lightricks just added two of the most practical tools yet to its LTX-2.5 Creative Lab: Restore and Refine Details. Both are free, open-weight IC-LoRA adapters for AI video restoration. Restore cleans up and colorizes damaged archive footage, from sepia film scans to mushy broadcast transfers. Refine Details rebuilds texture, edges and grain in soft, compressed or AI-generated clips.

If you have a box of family VHS rips, a documentary project full of public-domain newsreels, or AI video that looks plasticky at 1080p, this is aimed squarely at you. The catch is hardware: the official workflow wants a 32 GB GPU. Below is exactly what shipped, what it costs to run, and the free no-GPU path you can try in a browser today.

At a Glance

• What it is: two video-to-video IC-LoRAs for LTX-2.5: Restore (archive cleanup plus colorization) and Refine Details (texture and sharpness rebuild).
• What’s free: the weights (1.71 GB for Restore, 1.31 GB for Refine Details), the ComfyUI workflow and the docs. Hugging Face asks you to accept the license before downloading.
• License: LTX-2.x Community License. Commercial use is free for companies under $10 million in annual revenue.
• Hardware: Lightricks lists 32 GB VRAM, 32 GB RAM and 100 GB of disk as the minimum for LTX-2.5.
• The catch: these are generative tools. They invent plausible color and detail, so they do not recover what was really there.

What Lightricks Shipped in the LTX-2.5 Creative Lab

The LTX-2.5 Creative Lab collection is Lightricks’ shelf of task-specific adapters for its 22-billion-parameter video model. It already held Day-to-Night, Clean Plate, Deblur, Decompression, Colorization and a Pixel Spatial Upscaler. In late September it gained three more: Restore, Refine Details and SDR-to-HDR.

An IC-LoRA (in-context LoRA) takes your existing video as a guide and re-renders it, so it doesn’t generate from scratch. The base model does the heavy lifting, and the small adapter teaches it one specific job. That’s why each download is under 2 GB, even though the full LTX-2.5 stack we covered in August is a much heavier lift. The SDR-to-HDR adapter covers similar ground to Netflix Eyeline’s DiffHDR, so this post focuses on the two restoration tools.

Restore vs. Refine Details: Which AI Video Restoration Tool Do You Need?

The official Refine & Restore guide spells out the decision clearly:

• Run Restore on damaged, low-resolution, sepia or black-and-white archive footage. It cleans up the damage, reconstructs structure and infers color. Describe the period, place, lighting, materials and clothing in your prompt.
• Run Refine Details on footage whose color and composition are already fine but look soft. That covers compressed web video, conventional upscales and AI generations.
• Run both, in order, for archive footage you want at 4K: Restore first, then Refine Details. Lightricks warns that refining first can sharpen the damage and wash out color.
• Be careful with clean cinema footage. Refine Details can scrub away original film grain, so plan to add grain back afterward if it matters.

How the Tiled Fusion Workflow Works

Both adapters run through one ComfyUI graph, LTX-2.5_V2V_TiledFusion_Upscale.json. Your clip gets resized to the output canvas and attached as the guide. The sampler then works on overlapping tiles inside one shared latent and fuses them after every denoising step. That’s how a model trained on small tiles can output Full HD, 4K or even 8K.

The numbers to remember:

• Refine Details is trained on a 1024 × 576 tile. Use the tiled workflow even for 1080p.
• For Restore, Lightricks recommends a 960 × 544 tile, with a 1440 × 1056 or 1440 × 1088 canvas for 4:3 film and 1920 × 1088 for 16:9.
• Canvas sides must be multiples of 32, and frame counts must follow 8n+1 (97, 121 and so on).
• Your original audio track passes through untouched. Neither adapter generates new sound.

One prompting rule surprised us. With Refine Details, describe only the rendering (sharpness, grain, lighting, grade) and never name objects or people. Every tile gets the full prompt, so “a red car” can sprout a red car in every tile.

How It Compares to Other Free Video Restoration Tools

Try It Free: Three Paths, From Zero GPU to Full Workflow

Path 1: no GPU, five minutes. No hosted demo exists for Restore or Refine Details yet. The closest free option is Lightricks’ own LTX-2.3 Pixel Spatial Upscaler Space, which runs on free ZeroGPU hardware. Drop a short low-res clip into the left panel, add a prompt describing the scene, and run it. You’ll get a 4× generative re-render, close in spirit to Refine Details. ZeroGPU quotas and queues apply, so keep clips short.

Path 2: the official ComfyUI workflow (32 GB GPU). In ComfyUI, open Manager → Install Custom Nodes, search “LTXVideo,” install it and restart. Or install it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Lightricks/ComfyUI-LTXVideo.git
pip install -r ComfyUI-LTXVideo/requirements.txt

Next, accept the license on the LTX-2.5 model page and on each adapter’s page while logged into Hugging Face. Then put these files in place:

• ltx-2.5-22b-distilled-transformer-bf16.safetensors → models/diffusion_models/
• gemma4-12b-with-proj-ltx-2.5-bf16.safetensors → models/text_encoders/
• ltx-2.5-video-vae-bf16.safetensors and ltx-2.5-audio-vae-bf16.safetensors → models/vae/
• ltx-2.5-22b-ic-lora-restore-1.0.safetensors (1.71 GB) and ltx-2.5-22b-ic-lora-refine-details-1.0.safetensors (1.31 GB) → models/loras/

Load the Tiled Fusion Upscale workflow and pick your clip in Load Video. In Load Models, swap in the Restore adapter for archive footage (Refine Details is the default). Set the size in Preprocess → Get Tiling Sizes, then run. If memory is tight, launch ComfyUI with python main.py --reserve-vram 5 and try the low-VRAM loader nodes that ship with the pack.

Path 3: smaller GPUs (unofficial). Community GGUF builds such as Abiray/LTX-2.5-Distilled-GGUF shrink the transformer to 15.7 GB at Q4_K_M or 12.9 GB at Q3_K_M. Lightricks hasn’t validated the Restore or Refine Details workflow with quantized weights, so expect some quality loss and some tinkering. If you’re on 8–12 GB, SeedVR2 from the table above is the more realistic free upscaler, and pairing it with SageAttention in ComfyUI helps with speed.

Honest Trade-Offs

• It invents history. Lightricks says it directly: period-plausible color isn’t recovered historical truth. Don’t present a Restore pass as archival fact in a documentary without a disclaimer.
• Faces and lettering drift. Neither adapter guarantees that identities, signs or text survive intact. Check every shot with faces or titles.
• The hardware bill is real. The official floor is a 32 GB card, and the recommended setup is an 80 GB A100 or an H100. Most budget creators will be renting a GPU or using quantized workarounds.
• Restore wasn’t trained on LTX-2.5. It was trained on LTX-2.3 and tested unchanged with the 2.5 stack, according to the docs.
• 8K is a trap. A direct 8K pass can rearrange texture. Lightricks recommends refining to 4K, upscaling to 8K with Lanczos, then running an optional gentle second pass.
• No independent benchmarks yet. Every quality claim so far comes from Lightricks’ own examples.

Workflow Tips for Better Restorations

1. Don’t pre-clean your source. Skip denoising and sharpening before Restore. The adapter needs to see the damage to fix it.
2. Never stretch 4:3 to 16:9. Use the 1440 × 1088 canvas for old film and keep the black bars, or crop afterward.
3. Use the negative prompt as a period guard. List modern objects, logos and signage so Restore doesn’t put a smartphone in a 1937 street scene.
4. Test on 97 frames first. Lock your prompt and canvas on about four seconds before you commit hours of GPU time to the full reel.
5. Add grain back last. After Refine Details, put film grain back in your editor so the result doesn’t look waxy next to untouched footage.

Want to see how LTX handles 4K detail in practice? Watch our LTX 2.3 Pro 4K scene tests, and subscribe to the Creative AI Show on YouTube for hands-on comparisons. For frame-accurate edits rather than restoration, our guide to LTX Ripple is the companion read.

FAQ

Is LTX-2.5 Restore free?

Yes. The weights are a free download from Hugging Face once you accept the license, and the ComfyUI workflow is free on GitHub. The only cost is the hardware you run it on.

Can I use LTX-2.5 Restore commercially?

Yes, if your company (including affiliates) earns under $10 million a year. Above that, the LTX-2.x Community License requires a paid license from Lightricks. This isn’t legal advice, so read the license before client work.

What are the system requirements for LTX-2.5 Refine Details?

Lightricks’ official minimum for LTX-2.5 is an NVIDIA GPU with 32 GB VRAM, 32 GB system RAM, 100 GB of free storage, CUDA 12.7+ and Python 3.12+. Higher output resolutions use more memory and take longer.

Can AI really restore old black-and-white film?

It can make old film look clean and colorized, but the color and fine detail are educated guesses. Treat the result as a creative interpretation, not a historical record.

Is there a free online demo of LTX Restore?

Not yet. The closest free browser option is Lightricks’ LTX-2.3 Pixel Spatial Upscaler Space on Hugging Face, which does generative 4× upscaling but no colorization.

Accurate as of October 1, 2026. Model files, hardware requirements and free-tier quotas change fast, so check the linked model cards before you download.