ChronoEdit_SM_Enhance_Loader

Here's the thing nobody tells you when you first run ChronoEdit locally: the NVIDIA demo you saw was not prompting the model with your sentence. It was running your sentence through a Qwen vision-language model that stared at your image, thought about it in a chain-of-reasoning sort of way, and wrote the actual, much longer instruction. That's why people who reproduced the demo locally found it "hard to prompt" - the same prompt that worked on the HF Space failed in ComfyUI, because the Space was quietly doing the prompt engineering for them. This node is the local version of that step.

What it loads

ChronoEdit_SM_Enhance_Loader pulls in the prompt enhancer, a Qwen VL (vision-language) model, and hands it over as a ChronoEdit_SM_Model_En socket for the ChronoEdit_SM_Enhance node to use:

• repo - a HuggingFace model id, defaulting to Qwen/Qwen2.5-VL-7B-Instruct. Leave the default unless you know you want something else.
• clip - a combo of files in ComfyUI/models/clip. This is the alternate path: drop a local qwen_2.5_vl_7b.safetensors (or a Qwen3 variant) in there and it loads from disk instead of pulling from the hub. Handy if you're offline or already have the file.

One of the two needs to be valid; with repo set and clip at none, it downloads from HuggingFace on first use.

Why it's optional, and why you still want it

Two honest facts. First, it's genuinely optional - the README flags it as 可选 ("optional") and notes it's 太慢 ("too slow"). You can absolutely run a full ChronoEdit workflow with just the core CLIPTextEncode nodes and skip this entire branch. Second, skipping it is the most common reason ChronoEdit results underwhelm. The model was tuned around detailed instructions, and a good enhancer turns "put sunglasses on the cat" into a multi-sentence brief that pins down pose, lighting, and what not to change.

The cost is real, though. Qwen2.5-VL-7B is a 7-billion-parameter language model sitting in your VRAM next to a 14B diffusion transformer. You're not going to load both on a 12GB card at the same time - this is a "offload or run on the demo-grade machine" situation, and the pack's own example workflow chains it right into the pipeline anyway, so budget for it.

The plumbing

Its model output feeds ChronoEdit_SM_Enhance, which returns the polished prompt as a plain string. That string goes into a CLIPTextEncode as your positive - not into the sampler directly. Install is the pack's shared recipe (Manager: search ComfyUI_ChronoEdit_SM; or clone + pip install -r requirements.txt); note that the requirements include qwen-vl-utils and transformers, which this node is the main reason for.