ChronoEdit_SM_Lora

Most ComfyUI LoRA stackers exist for style sliders. This one is different: the two LoRA slots exist because ChronoEdit genuinely ships with two adapters you're supposed to use, and skipping them changes what the sampler is doing. ChronoEdit_SM_Lora sits between the model loader and the sampler, and it's where a lot of the "why is mine slow / why is mine soft" confusion starts.

What's in the slots

• lora_1 and lora_2 - combos listing everything in ComfyUI/models/loras, with a none option. Slot one is the 8-step distill LoRA (chronoedit_distill_lora.safetensors, from nvidia/ChronoEdit-14B-Diffusers); slot two is the upscale LoRA (upsample_lora_diffusers.safetensors, from ChronoEdit-14B-Diffusers-Upscaler-Lora).
• lora_scale1 / lora_scale2 - 0.1 to 1.0. The README's example runs both at 1.0; dial the upscaler down a bit if your output gets crunchy.

Output is the same ChronoEdit_SM_Model socket, one step closer to the sampler.

Why the distill LoRA is the one you reach for

ChronoEdit's base model is a full-step diffusion transformer - usable, but slow. NVIDIA shipped a distilled variant that collapses the denoising trajectory so you can sample in ~8 steps instead of dozens. That's the classic distillation tradeoff from the KB: fewer steps, less time to refine, quality still fine for editing work because the edit only has to land on a few key frames. In practice, the distill LoRA is the default path and the community treats it as part of the model - it's the same adapter Kijai ships alongside the fp16 conversion. Run without it and your sampler is back to full sampling with an 8-step expectation, which is a recipe for mushy output.

The mechanism is plain diffusers plumbing: load_lora_weights for each adapter, then set_adapters with your scales. One detail the code gets right: if you change your LoRA selection between runs, it deletes adapters from the pipeline that are no longer in your slots - so you don't get stale weights silently stacking up across runs.

The README's one tip that actually matters

The update notes are terse but worth quoting: if you use both LoRAs together, bump your steps to 8–12; if you use the upscale LoRA at 8 steps, push past 12. The upscaler wants room to work, and the distill LoRA wants to not be starved. Two lines you'll be glad you remembered the first time the output looks over-cooked or under-sampled.

Install is the pack's shared routine - ComfyUI Manager, search ComfyUI_ChronoEdit_SM, or clone into custom_nodes and pip install -r requirements.txt. The LoRA files themselves go in ComfyUI/models/loras, nothing exotic. If the combos look empty, that's the first thing to check: the README gives those two exact filenames, and the node only sees files in that folder.