ChronoEdit_SM_Enhance

The default prompt sitting in this node is a giveaway for what it does: "inside the floral teacup of steaming tea, a small cute mouse is sitting and taking a bath, a tiny white bath towel draped over its head." It's a perfectly fine sentence - and nowhere near enough for ChronoEdit. This node is the translator between the sentence you actually want and the detailed, image-aware instruction the model edits well. It's the second half of the pack's prompt-enhancer pair, and it only does anything useful if ChronoEdit_SM_Enhance_Loader fed it a Qwen VL model first.

What it takes

• image - the same reference image you're editing. The enhancer looks at it. This is the crucial bit: it's not generic prompt-polishing, it's an instruction written about this specific photo.
• prompt - your intent, in plain words. Keep it short; the whole point is that it gets expanded.
• model - the ChronoEdit_SM_Model_En output from ChronoEdit_SM_Enhance_Loader.

Output is a single STRING: the enhanced prompt.

How it works

Mechanically it runs the Qwen vision model on your image plus your line of text and asks it to think through the edit - the pack literally calls the underlying function prompt_enhance, built on the CoT (chain-of-thought) prompt logic NVIDIA shipped with ChronoEdit. The model reasons about what's in the frame, what your instruction implies about the scene, and what should stay untouched, then writes the verbose brief. The output reads like a paragraph of art direction: scene, subject, what to preserve, style notes.

This is why the example workflow's positive prompt is a wall of text about a kimono figure becoming a PVC collectible - that's not someone being dramatic, that's the enhancer's handiwork. And it's why the community's first complaint about local ChronoEdit was "the same prompt works in the demo but not in ComfyUI": the HF Space ran this step by default, and anyone reproducing it locally without this pair was missing the whole prompt-engineering layer.

Wiring it and the honest caveats

The STRING output feeds a CLIPTextEncode (as the positive), which produces the CONDITIONING that goes into ChronoEdit_SM_KSampler. It does not plug into the sampler directly.

Caveats, straight: this is a 7B VL model doing real generation, so each call takes a few seconds to tens of seconds and hogs VRAM next to the 14B DiT. The README flags it optional and slow. If your VRAM is tight, skip this pair and write detailed prompts yourself - a good manual prompt gets you 90% there. But if you're chasing the demo's behavior and your prompts feel weak, this is the missing ingredient, and it's worth the wait.

Install is the pack's shared recipe: ComfyUI Manager, search ComfyUI_ChronoEdit_SM, or clone into custom_nodes and pip install -r requirements.txt. Keep ComfyUI current - the pack targets the new extension API, and stale ComfyUI just won't show the node.