<p>I built ReDetail to upscale video in ComfyUI with LTX-2.5. Version 2.0 adds refine mode as the default. It uses Lightricks' Refine-Details IC-LoRA to rebuild missing fine detail in soft clips while keeping the framing, colour, motion and faces as they were. It works in 1024x576 tiles fused at every step, so VRAM follows the tile instead of the frame. A 4K render runs on a 24GB card; I tested it on an RTX 4090.</p><p>The 4K stills are 100% crops of 3840x2144 output from two generated clips, 1376x768 and 1664x928, each refined in one 14-minute pass on an RTX 5090. Lanczos is on the left, refine on the right. The reel shows those two and four more clips at 1:1 pixels of the 4K frames.</p><h3 id="which-file-is-which">Which file is which</h3><p>v2.0 has two files. ReDetail_LTX25_refine.json is the default refine mode. ReDetail_LTX25_upscale.json is pixel mode, the 1.x Pixel Spatial Upscaler. It re-renders the clip and invents more detail, but lets faces drift. On a Mac, use ReDetail_LTX25_upscale_MAC.json from v1.1 or GitHub. It only supports pixel mode; refine hasn't been tested on Apple Silicon.</p><h3 id="refine-or-pixel">Refine or pixel</h3><p>On seven MiniMax H3 clips (640x384, upscaled 2x), refine stayed closer to the source every time: 31.8-38.9 dB PSNR against pixel's 24.2-31.3. It flickered less too, 1.4-2.7x Lanczos against 2.2-5.1x. Pixel adds more detail (2.3-3.8x Lanczos against 1.4-2.4x) and renders in less than half the time.</p><p>I use refine when I need to preserve a face, product or framing, and for anything going to 4K. Pixel is for soft AI footage with nothing to preserve.</p><h3 id="what-it-changes-from-lightricks'-example-graph">What it changes from Lightricks' example graph</h3><ul><li><p>The tile is pinned to the 1024x576 the LoRA was trained on. The example sizes tiles from your source, making a 768x1376 clip one tile 1.8x the trained area.</p></li><li><p>The prompt-enhancer branches are disconnected. ComfyUI validates them even when switched off, and one missing file blocks the whole queue.</p></li><li><p>The Missing Models panel links the int8 files instead of 68GB of bf16.</p></li></ul><h3 id="what-you-need">What you need</h3><ul><li><p>ComfyUI-LTXVideo from 24 September 2026 or later, with its requirements installed in ComfyUI's own Python</p></li><li><p>comfy-kitchen 0.2.26 or newer, in that same Python</p></li><li><p>About 40GB of models, all linked in the Missing Models panel. The Hugging Face repos are gated, so accept the licence on each.</p></li><li><p>For 4K, a 24GB card and plenty of system RAM. On an RTX 4090 a 97-frame 4K pass took 21 minutes and peaked at about 50GB of RAM; with less, the CLI's --budget splits the clip.</p></li></ul><p>Your clip needs an audio track and a length of 8n+1 frames before it goes in. The graph's note panels have the ffmpeg commands for both.</p><h3 id="the-cli">The CLI</h3><p>The CLI on GitHub handles that prep. It supports exact scales instead of the two presets and splits long clips on their own cuts. It also puts your original audio back. With --cached-cond it never loads the text encoder.</p><pre><code>python3 redetail.py --setup
python3 redetail.py clip.mp4 --scale 1.5
python3 redetail.py clip.mp4 --scale 1.5 --model pixel</code></pre><h3 id="what-it-won't-do">What it won't do</h3><p>Refine can't fix a malformed face in the source; it rebuilds the detail of the face already there. Small text and logos can become glyph-shaped noise. Pixel re-imagines faces (it added freckles to my test person), logos, numbers and text. With either mode, compare a face against your source rather than judging overall sharpness.</p><p>GitHub: <a target="_blank" rel="ugc" href="https://github.com/Bambushu/redetail">https://github.com/Bambushu/redetail</a></p><p>The code is MIT. The workflows and weights are under Lightricks' LTX-2 Community Licence. Its use restrictions pass on to you, and any company over $10M in annual revenue needs a paid licence from Lightricks. The full text is in the repo.</p>

ReDetail 2.0: LTX-2.5 video refine + upscale to 4K (workflow + CLI)

<p>First release. Drag-and-drop ComfyUI workflow for the LTX-2.5 IC-LoRA Pixel Spatial Upscaler, plus a Python CLI that drives the same graph for longer clips.</p><p>Includes six in-graph note panels (install, clip prep, sizing, errors), a placeholder first-frame image, and the canvas links severed so the output size is actually settable. Runs on 24GB via the GGUF path.</p>

Files with SHA256: 22dd9326761e4e26b7241d7a487675a1d6c5976f91d2144dc0052047f5c025a8