Native Resolution

Native Resolution is a two-stage LTX-2.5 video-to-video workflow. The first stage resizes the source video to the selected initial canvas and establishes the composition with the full frame in context. The second stage enlarges the latent to the selected output size, attaches the source video as a guide at that size, and uses Tiled Fusion for refinement.

Tiled Fusion processes a canvas larger than an IC-LoRA’s trained window as overlapping spatial crops, then blends the stepped crops back into one shared latent canvas. It is used in the second stage of the Native 4K/8K workflow; it does not replace the initial full-frame composition pass.

The workflow is designed for a single-GPU path. Memory use and runtime still depend on the output canvas, tile size, temporal window, model stack, and decode settings, so benchmark the intended configuration on the target hardware.

This is different from independently generating tiles and stitching completed results. The tiles share one latent canvas and one noise field throughout sampling, which helps reduce discontinuities between regions. Always inspect tile boundaries and low-frequency gradients across the full clip.

When to use Native Resolution

Use this workflow when:

• You want a full-frame composition pass before refining a video at a larger output size.
• The refinement canvas is larger than the IC-LoRA’s trained spatial window.
• Full-canvas sampling would exceed available VRAM.
• You need overlapping regions to participate in the same denoising trajectory.

Do not use Tiled Fusion merely to reduce VAE decode memory. For that task, use LTXVTiledVAEDecode or LTXVSpatioTemporalTiledVAEDecode after sampling.

Example workflows

The published LTX-2.5 examples share the same subgraph layout (Load Models, Inputs, Preprocess, Generate, Decode), and a Get Tiling Sizes node in Preprocess sets the tile and output sizes.

The Native Resolution ladder (Native 4K/8K)

LTX-2.5_V2V_TiledFusion_Native_4K_8K.json runs two stages:

1. Composition — the source is resized to initial_canvas_size, and a full-frame pass establishes the composition. The published graph selects ltx-2.5-22b-ic-lora-day-to-night-0.9.safetensors for this stage.
2. Refine — the latent is enlarged with ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors (twice for the 8K path), then Tiled Fusion runs a refinement pass with spatial tiles. The published graph selects ltx-2.5-22b-ic-lora-refine-details-1.0.safetensors, uses a fresh look-and-style prompt, and attaches the original clip resized to the output size as the stage-2 guide.

The stage-1 adapter is available from LTX-2.5 Day-to-Night, the stage-2 adapter from LTX-2.5 Refine Details, and the spatial upscaler from the LTX-2.5 model repository. Use the workflow’s Model Links note for the rest of the version-matched model stack.

The source audio is encoded and held fixed during generation, then decoded and muxed into the saved output video.

Output sizes

The Get Tiling Sizes node emits model-legal canvases (dimensions in multiples of 32, frame count 8k+1) from three presets:

• tile_size — qHD, HD, or FullHD. The spatial tile the fusion sampler steps.
• initial_canvas_size — the stage-1 composition canvas.
• output_size — FullHD, 4K, or 8K.

Supported surface and version

Use only an IC-LoRA and model stack explicitly matched to the selected workflow. Do not infer LTX-2.3 support from the sampler interface or universal LTX-2.5 compatibility from one example.

Customize safely

Start with the task-specific example workflow and run it unchanged before modifying it. Keep the version-matched model and IC-LoRA together, and match the tile dimensions to the adapter’s trained spatial envelope.

The workflow contains coordinated guide, sampling, overlap, and streaming settings. Changing one in isolation can produce incompatible conditioning or visible seams. For the LTXVTiledFusionSampler interface and implementation constraints, see the node reference.

For longer clips, keep the temporal-window setting synchronized between the IC-LoRA guide and sampler so each window is encoded with the correct local context. Do not slice a whole-clip guide by hand.

Output and decoding

The sampler returns one latent canvas. Decode it with the VAE required by the version-matched workflow.

Tiled sampling and tiled VAE decoding solve different memory problems:

• Tiled Fusion sampling divides the diffusion-model work across overlapping canvas regions at every denoising step.
• Tiled VAE decoding divides only the conversion from the completed latent into pixels.

Using tiled VAE decoding does not turn a standard workflow into the Tiled Fusion workflow.

Limitations

• Full HD, 4K, and 8K are available in the size selector. Higher resolutions require more memory and runtime; benchmark the target resolution on the production hardware.
• When the second stage uses Refine Details, its model card recommends reaching 8K progressively through a 4K refinement rather than relying on one direct 8K pass; a tile at 8K sees content beyond the adapter’s trained scale and can invent texture.
• This guide documents the ComfyUI workflows, not a native Python path.
• Do not reuse the graph with another IC-LoRA unless a published workflow explicitly supports that combination with Tiled Fusion.
• Tile overlap reduces boundary risk but does not prove that every output will be seam-free.
• Large canvases still require memory for the shared latent canvas, noise field, conditioning, and decoded output.
• Tiled Fusion does not replace an untiled composition pass when an adapter must establish a new global look across a canvas far beyond its trained window.

Before production use

1. Run the selected workflow unchanged.
2. Verify that the model, IC-LoRA, and workflow target the same LTX version.
3. Test the intended output resolution on the target hardware.
4. Inspect structured edges, low-frequency gradients, motion, and tile boundaries through the full clip.
5. Record the workflow revision, adapter, sampler, tile settings, VRAM, runtime, and output resolution.

Related pages

• LTX ComfyUI Nodes
• Using ComfyUI with LTX
• IC-LoRA