ChronoEdit Upscaler LoRA (CRT AutoDL)

ChronoEdit's weak spot is resolution. As a Wan-based editor it's happiest at the resolutions it was trained on, and when you ask a 14B video model to render an edit larger than that, details get soft and structures wobble - the same ceiling you hit with any big generation model pushed past its training box. NVIDIA's answer is an upscaler LoRA, and this node fetches and applies it automatically.

CRTAutoDLChronoEditUpscalerLoRA is the twin of the pack's Distill LoRA node with a different payload. Where the distill LoRA buys speed, this one buys size. It takes model plus strength_model (default 1.0, range -100 to 100), downloads upsample_lora_diffusers.safetensors from NVIDIA's official ChronoEdit-14B-Diffusers-Upscaler-Lora repo into ComfyUI/models/loras on first run, and returns a patched MODEL with the LoRA merged in at your strength.

How to actually use it

The common pattern is to apply the upscaler LoRA for a second pass: generate the base edit at a sane resolution, then re-run through the sampler with the upscaler LoRA applied and a higher resolution target - an in-model upscale rather than a post-hoc ESRGAN pass, which is why the detail survives. The output is a single patched MODEL, so you can chain it downstream of the Distill LoRA:

ChronoEdit Model → Distill LoRA → Upscaler LoRA → sampler at higher res

Keep strength_model at 1.0 to start. If your upscaled result starts inventing grain or plastic texture, dial it back toward 0.8 before you blame the sampler.

Install

Same pack, same routine:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes
pip install -r requirements.txt

Or in ComfyUI Manager, search CRT-Nodes and install, then restart. The LoRA file downloads on first use through the node itself. Like every ChronoEdit AutoDL node here, it downloads its own file independently, so a full first-time ChronoEdit run has several progress bars - that's normal, and it's also why these nodes are worth having: zero manual model wrangling, ever.