LTX Refine and Restore: ComfyUI setup and tests

This tutorial tests LTX Refine and Restore in ComfyUI's Native Resolution workflow. The speaker calls them Refiner and DeArchive or de-aging LoRA; the description supplies their release names. Examples include aerial footage reduced to 512 pixels and processed to 4K, an archive commercial, and a 1K AI-generated clip.

The setup covers model replacement, quantization and tile settings. The speaker uses an RTX 5090 with a three-year-old Threadripper and 128 GB of system RAM. He uses a quantized model and integer-eight text encoders instead of the suggested BF16 setup, while noting that even his GPU is under pressure. The workflow has full HD, 4K and 8K output options. He recommends testing at full HD first and allowing much longer for 8K.

Overlapping tiles let the workflow process portions of a frame. Prompts therefore need to describe what each portion could contain: a specific subject can otherwise appear in an unrelated tile. The local AI examples also show the limits of inferred detail. Two people merge into one character, helmets resemble civilian hats, and smoke trails become lights.

The advanced section unpacks the subgraph to expose the IC-LoRA loaders and adjust their strengths. For atomic bomb footage, the speaker settles on Restore with 30% Refine. These are his experimental results, rather than evidence that restoration recovers the original scene accurately.