LTX-2.5 Video Upscale & Re-detail Runbook

Plan: 1080p presenter clip to 4K on a rented 4090

Mode: refine, --scale 2.0. A face has to survive and the target is 4K; pixel would invent detail on the face (freckles, lashes) and its VRAM grows with the clip. Refine's VRAM follows the 1024x576 tile, so a 24 GB card is fine (measured 23.3 GB peak at 4K on a 4090).

Output size: 3808x2144 (1.99x). Refine needs both sides divisible by 32 and the CLI picks the nearest size that keeps 16:9, so 3840x2160 is not reachable exactly.

Memory: the real constraint here

Refine at 4K is bound by system RAM, not VRAM:

RAM ~= 22 GB staged weights + 0.033 GB x frames x 8.16 MP   (with --cached-cond)

Use --budget 900. It fits 64 GB with headroom and cuts the seams from 5 to 2. Because this is one continuous take, there is no scene cut to hide a chunk boundary, so check those two frames (see QA). If a seam shows, re-run on a 128 GB-RAM box with --budget 2100 for a single chunk.

Time: about 60 minutes, scaled from the measured 21 min for 97 frames at 2176x3904 on a 4090.

Setup on the rented box

# 1. pack requirements into ComfyUI's own Python, with the image's pip constraint cleared
PIP_CONSTRAINT= /path/to/ComfyUI/venv/bin/python -m pip install -r custom_nodes/ComfyUI-LTXVideo/requirements.txt "comfy-kitchen>=0.2.26"
/path/to/ComfyUI/venv/bin/python -c "import colour, kornia, comfy_kitchen; print('ok')"

# 2. skip the 15 GB text encoder: copy ReDetail's tools/comfyui_cond_cache/ into ComfyUI/custom_nodes/, restart ComfyUI
# 3. from your machine, tunnel instead of using the provider's HTTP proxy (it blocks POST uploads)
ssh -N -L 8188:127.0.0.1:8188 -p <port> <user>@<host>

Run

python3 redetail.py --setup --model refine --comfy http://127.0.0.1:8188
python3 redetail.py presenter.mp4 --scale 2.0 --cached-cond --budget 900 --comfy http://127.0.0.1:8188

Your original audio is re-muxed by default. There is no resume: if ComfyUI dies mid-run, press Ctrl-C, fix the cause, and run it again.

QA before you upload

Face first, at 100%. Crop the face from the source and from the result, scale both to the same size with Lanczos, and compare side by side. Refine keeps the face it was given; it will not repair a soft or malformed one.
2. Fidelity numbers:

ffmpeg -i presenter_redetail.mp4 -i presenter.mp4 -lavfi "[0]scale=1920:1080:flags=lanczos[a];[a][1]ssim" -f null - 2>&1 | tail -1

Expect SSIM around 0.92-0.98 for refine. Well below that means something drifted.
3.

The two chunk boundaries (around frames 80 and 160): step through them at 100% and look for a texture change in skin and hair.
4. Motion: play it at full size; added detail can shimmer.

Then stop the rented instance.