Introduction

If you have been waiting for LTX-2.5 to stretch past “nice HD plate” into real VFX delivery sizes, Lightricks just opened the ladder. On 2026-09-29, ComfyUI-LTXVideo PR #570 merged two example graphs under example_workflows/2.5/: Native 4K/8K and Tiled Fusion Upscale. Official docs call the pattern Native Resolution — a two-stage video-to-video path designed for a single GPU.

This is not a rehash of the August open-weights drop. Those foundation weights got you into the world model. Native Resolution is the VFX Week day-one toolkit: shared-latent tiled refine, Get Tiling Sizes presets, and the new Refine Details / Restore IC-LoRAs.

What Native Resolution Actually Does

Native Resolution is a two-stage ladder:

  1. Composition — resize the source to initial_canvas_size and run a full-frame pass so the whole frame stays in context. The published Native graph uses the Day-to-Night IC-LoRA for this stage.
  2. Refine — enlarge the latent with ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors (twice on the 8K path), then run Tiled Fusion with the Refine Details IC-LoRA, a fresh look-and-style prompt, and the original clip resized to the output size as the stage-2 guide.

The companion graph, LTX-2.5_V2V_TiledFusion_Upscale.json, is the single-pass detail-refine / upscale path used for Refine & Restore. Same tiling family, different job: polish soft footage, or clean archive plates before a delivery refine.

Tiled Fusion ≠ Old Tiled Sampler

Warm warning sticker for your node graph: LTXVTiledFusionSampler is not LTXVTiledSampler. You cannot swap them and call it a day.

  • The older tiled sampler samples and blends tiles more independently.
  • Tiled Fusion keeps one shared latent canvas and one noise field, steps overlapping spatial crops at every denoise step, and fuses them back so regions share the same trajectory.

That shared trajectory is why Full HD / 4K / 8K can stay selectable without stitching finished tiles like a quilt. Inspect tile boundaries and low-frequency gradients anyway — overlap reduces seam risk; it does not issue seam insurance.

Also keep the jobs straight: Tiled Fusion is for diffusion sampling memory. For decode-only pressure, use tiled VAE decode nodes after sampling. One does not magically become the other.

Get Tiling Sizes: FullHD / 4K / 8K

Both published graphs share the familiar LTX-2.5 subgraph layout (Load Models → Inputs → Preprocess → Generate → Decode). Inside Preprocess, Get Tiling Sizes emits model-legal canvases (multiples of 32, frame count 8n+1) from three levers:

  • tile_size — qHD / HD / FullHD sampling tile
  • initial_canvas_size — stage-1 composition canvas
  • output_size — FullHD / 4K / 8K

Those presets are not a VRAM guarantee. Peak memory still tracks tile size, temporal window, model stack, and decode settings. Benchmark the intended resolution on your box before you commit a hero clip.

Refine Details, Restore, and Progressive 8K

Two new adapters live on Hugging Face under Lightricks and drop into ComfyUI/models/loras/:

  • Refine Details — rebuilds fine texture, edges, and grain on soft, compressed, generated, or already-upscaled footage. Trained for 1024×576 tiles; use the tiled workflow even at full HD.
  • Restore — cleans and colorizes damaged archive material (low-res transfers, tape, sepia, B&W scans). For archive, Restore first, then Refine. Refining first can sharpen the damage and mute the inferred color.

These are generative tools. They synthesize plausible detail and color; they do not resurrect lost pixels. Treat rebuilt faces, lettering, and objects as interpretations.

For 8K, Lightricks recommends a progressive path: refine to 4K first, Lanczos resize toward 8K, then an optional gentle second pass. A direct 8K refine can invent texture because each tile sees content beyond the adapter’s trained scale.

Source audio is encoded, held fixed during generation, then decoded and muxed back into the saved video — so your dialogue and beds survive the upscale.

How to Open It Today

# Update the official LTX video node pack, then load either:
# example_workflows/2.5/LTX-2.5_V2V_TiledFusion_Native_4K_8K.json
# example_workflows/2.5/LTX-2.5_V2V_TiledFusion_Upscale.json

Pull the version-matched distilled transformer, Gemma-4 text encoder, video/audio VAEs, latent spatial upscaler, and the IC-LoRA you need from the workflow’s Model Links note (Lightricks HF repos are gated — request access first). Run the published graph unchanged before you start twiddling overlap, streaming windows, or sampler family.

For a community walkthrough of the same drop, see the ComfyUI Wiki Native Resolution note.

Original Source

https://github.com/Lightricks/ComfyUI-LTXVideo/commit/5722b53a08daec57f5442347b235cb84df0afbe7

https://docs.ltx.io/open-source-model/vfx-post-production/native-resolution

https://docs.ltx.io/open-source-model/vfx-post-production/refine-and-restore

Conclusion

Native Resolution is the practical bridge from “I can edit with LTX-2.5” to “I can aim Full HD, 4K, or 8K V2V on one GPU without stitching nightmares.” Update ComfyUI-LTXVideo, load the 2.5 example graphs, keep Fusion and the old tiled sampler mentally separate, and climb the ladder progressively when you chase 8K. Archive plates get Restore → Refine; soft modern plates can jump straight to Refine Details. Happy tiling — may your seams stay shy and your canvases stay shared.

—Aurelia ♡