<p>A drag-and-drop ComfyUI workflow for LTX-2.5's IC-LoRA Pixel Spatial Upscaler. Drop the .json in and render. Six note panels inside the graph cover install, clip prep, the sizing table and an error lookup, so you should not need to leave the canvas to work out why something failed.</p><p><strong>v1.1:</strong> fixes a bug where the workflow would not queue if your transformer was not in models/diffusion_models. The prompt enhancer failed validation on a branch that never runs, so disabling it did not help. Also: the text encoder is now optional, and there is an Apple Silicon build.</p><p>It does not sharpen and it does not restore. It re-renders the frame and invents the fine detail as it goes. That makes it good on AI-generated and soft low-detail footage with nothing real to recover, and a bad idea anywhere a face has to stay the same person.</p><h3 id="why-not-just-use-the-stock-example-graph">Why not just use the stock example graph</h3><p>Its canvas width and height are wired to GetImageSize on the source, so it renders 1:1 whatever resolution you type. This one severs those links so the output size is actually settable. It also ships a placeholder first-frame image so the empty LoadImage cannot throw on your first queue.</p><h3 id="a-cli-is-included-for-real-clips">A CLI is included for real clips</h3><p><a target="_blank" rel="ugc" href="http://redetail.py">redetail.py</a> drives the same graph over ComfyUI's API and handles sizing, splitting long clips on their own scene cuts, the 8n+1 frame grid, and re-muxing your original audio:</p><pre><code>python3 redetail.py --setup
python3 redetail.py clip.mp4 --scale 1.5</code></pre><p>--setup checks your install over ComfyUI's HTTP API and names what is missing. Correction to v1.0: int8_convrot is NOT Blackwell-only, it runs on Ampere and Ada. If it will not load, check comfy-kitchen is 0.2.26 or newer before blaming your card.</p><h3 id="runs-on-24gb">Runs on 24GB</h3><p>On an RTX 4090 with the Q4_K_M GGUF transformer and tiled decode, sampling took 70s for 8 steps at 21.8GB of 24.5GB peak. The bf16 text encoder is no longer needed: a 26KB cached conditioning replaces it and saves 5.6GB of peak VRAM, bit-identical output.</p><p>1.5x is the sweet spot. On a 96GB card, 243 frames from a 768x1408 source: 1:1 took 3 min at 52GB, 1.5x took 7 min at 65GB, 2x took 17 min at 80.5GB.</p><h3 id="what-it-will-do-to-a-face">What it will do to a face</h3><p>It invents skin texture. On our test person it consistently added freckles that were not in the source. On a motocross clip it re-drew the jersey graphic and the number plate. Those stay stable across frames, but they are not the source's markings. Any logo, number or text gets re-imagined. Nothing in the model preserves identity, so check faces against your original rather than judging by sharpness.</p><p>All demo footage here is MiniMax H3 output at 640x384, 10s, upscaled 2x. Comparisons are Lanczos vs ReDetail at the same output size, so it is not just bigger-looks-better.</p><p>GitHub: <a target="_blank" rel="ugc" href="https://github.com/Bambushu/redetail">https://github.com/Bambushu/redetail</a></p><p>Licence, stated properly because this one has teeth. The CODE is MIT, but the WORKFLOW FILES and the weights are under Lightricks' LTX-2 Community Licence, which carries use restrictions that pass on to you. Anyone over $10M annual revenue needs a separate paid commercial licence from Lightricks before using LTX-2 or derivatives. A complete copy of the Agreement ships in the repo.</p>

ReDetail: LTX-2.5 generative video upscaler v1.1 (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: 65111687e3c1a880dafe9b560ff04ace21f337db5ddeede3b12f5197fcd08f87