ChronoEdit Distill LoRA (CRT AutoDL)

The reason ChronoEdit is actually usable on a consumer GPU is this LoRA. NVIDIA shipped ChronoEdit-14B as a 14-billion-parameter Wan-based editor, and running that at full step count is a slog. The fix is a distillation LoRA - a small adapter that was trained to collapse the model's many denoising steps into few. This node fetches that LoRA automatically and patches your model with it.

It's the same lineage as the Lightning/Turbo/LCM adapters that have been doing this to image models for years: distillation trains a student that predicts where the teacher would land after dozens of steps, in a couple of big jumps. Fast, and with the usual honest tradeoff - fewer steps means less refinement time, and pushed too hard, quality thins out. For ChronoEdit the distill LoRA is the intended path, not a hack; community workflows treat it as part of the standard stack.

How it works

CRTAutoDLChronoEditDistillLoRA takes two things:

• model - a MODEL, almost always the output of the pack's ChronoEdit Model node.
• strength_model (default 1.0, range -100 to 100) - how hard to apply the LoRA.

On first run it downloads chronoedit_distill_lora.safetensors from NVIDIA's official ChronoEdit-14B-Diffusers repo into ComfyUI/models/loras, then merges it into your model with comfy.sd.load_lora_for_models. Output is a single patched MODEL that you feed straight into your sampler.

The input that matters

strength_model is the only dial, and it behaves like every LoRA strength you've met: 1.0 is the tested default, lower softens the effect, and the negative range exists if you want to go against the distillation for more detailed (slower) generations - an experiment worth trying when your edit needs finer texture than the distilled path gives you. Start at 1.0 and only move if the output looks oversmoothed.

Wire the patched model to your KSampler, and you'll typically want to drop your step count hard compared to an undistilled run - that's the whole point. Pair it with the ChronoEdit CLIP - WAN, ChronoEdit VAE, and (for reference edits) ChronoEdit CLIP Vision siblings, all of which this pack also auto-downloads.

Install

The pack again. ComfyUI Manager → search CRT-Nodes, or:

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

Restart ComfyUI. The LoRA file downloads on first use through the node - no manual HuggingFace trip. Heads-up: this is one of several ChronoEdit nodes that each download their own file on first run, so a fresh workflow will show multiple progress bars before your first image appears.