ChronoEdit_SM_Model

ChronoEdit is NVIDIA's image editor that thinks like a video model - it rolls a clip of frames through a diffusion transformer so the edit is "reasoned about" over time instead of painted onto a flat image, then hands you back the frame where the edit actually landed. It landed in October 2025 to a crowd that had been waiting for exactly this: it's roughly on par with Qwen Edit but without the "everything drifts" degradation people kept hitting there. This node is the door to all of it - the thing that actually loads the 14B transformer into memory.

The two inputs are a fork in the road

You get exactly two combo boxes, and the whole point is that you fill one:

• dit - a single-file checkpoint in ComfyUI/models/diffusion_models. This is the fp8 route (ChronoEdit-14B-Diffusers-FP8 from the README), or a full-precision safetensors if you have the VRAM and don't mind it.
• gguf - a quantized file in ComfyUI/models/gguf. The pack auto-creates that folder and registers it, so whatever you drop in shows up here. ChronoEdit-14B-Q6_K.gguf and Q8_0 are the ones the README points at, from QuantStack/ChronoEdit-14B-GGUF.

Both default to none, and picking both is just asking the loader to prefer the GGUF. The output is a model socket of the pack's own ChronoEdit_SM_Model type, which you then feed into ChronoEdit_SM_Lora or straight into ChronoEdit_SM_KSampler.

Which one should you grab?

The 14B transformer is a Wan-architecture model, and 14B is a lot of weights. The README says 8GB+ VRAM, the project metadata is more honest at ">12G" - plan around 12GB for anything resembling comfort. That's the entire reason you have a GGUF slot at all. The GGUF Q ladder is the standard trick for big Wan-family models: Q8 is basically indistinguishable from fp16 at half the size, Q6 is the sweet spot for a 12GB card, Q4 gets you in the door with visible loss. If your card has room for Q8, just use it. Under about 12GB, fp8 stops fitting and GGUF is the only game.

The loader under the hood builds a diffusers ChronoEditPipeline - GGUF goes through WanTransformer3DModel.from_single_file with a bf16 compute dtype, the safetensors path uses init_empty_weights and fills it in. You don't need any of that to use it, but it's why the whole workflow expects this model socket rather than a normal ComfyUI MODEL.

Installing it (same for every node in this pack)

ComfyUI Manager and search ComfyUI_ChronoEdit_SM, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/smthemex/ComfyUI_ChronoEdit_SM
cd ComfyUI_ChronoEdit_SM
pip install -r requirements.txt

Then restart ComfyUI. The requirements pull in the heavy hitters - diffusers, transformers, accelerate, peft, qwen-vl-utils - and the pack also ships a big chunk of NVIDIA's original Chrono source tree and local Qwen folders, so the clone is chunky. One real gotcha: this pack is written against ComfyUI's new extension API (comfy_api.latest), so if your ComfyUI is more than a few updates old, the nodes simply won't register. Update ComfyUI before you start debugging "where are the nodes."

Download the DiT, drop it where the combo can see it, and you're one step from a working edit.