LCM_IPAdapter_inpaint

LCM_IPAdapter_inpaint is the maxed-out loader of the pack: it builds the LCM_Dreamshaper_v7 pipeline, loads a ControlNet into it, and wraps an IP-Adapter around the whole thing. This is the backend for the pack's fanciest workflows - the ones that combine a reference image for style, structural control from a ControlNet, and 4-step LCM speed in a single pass. If you were confused about why there are two IP-Adapter loaders, this is the one to pick when inpaint + control + image-prompting all need to happen together.

What's different from LCM_IPAdapter

Same three inputs as the plain loader - device (cpu/cuda), ip_adapter (file dropdown from models/controlnet), ip_adapter_full_path (manual path override) - plus one more: mode, a dropdown of folders in ComfyUI/models/controlnet. That folder is loaded as a ControlNetModel into an LatentConsistencyModelPipeline_inpaintV3, and the IP-Adapter then wraps the finished pipe. So while the plain loader gives you a pipe with IP-Adapter only, this one gives you a pipe that can be driven by both control_image/control_weight and ip_model/pil_image/scale inputs on the generation side (LCMGenerate_inpaintv3).

Same IP-Adapter rules as the plain loader: filename containing plus switches you to IPAdapterPlus with 16 tokens, and a CLIP vision encoder is expected in ComfyUI/models/clip_vision.

Setup

The full stack, in order:

1. Install the pack - ComfyUI Manager (search LCM_Inpaint-Outpaint_Comfy) or git clone https://github.com/taabata/LCM_Inpaint-Outpaint_Comfy into custom_nodes, then pip install -r requirements.txt.
2. LCM_Dreamshaper_v7 diffusers model → ComfyUI/models/diffusers/.
3. IP-Adapter checkpoint → ComfyUI/models/controlnet/ (as a file).
4. ControlNet → a subfolder in ComfyUI/models/controlnet/ (as a folder, for the mode dropdown).
5. CLIP vision encoder → ComfyUI/models/clip_vision/.

The honest part

This is the heaviest loader in the pack, and it shows. You're loading the full diffusers LCM model, a ControlNet, and an IP-Adapter - all on top of whatever ComfyUI already has cached - so 8 GB VRAM cards will sweat. And there's a real tension baked into the design: IP-Adapter wants to steer the image toward a reference while ControlNet wants to enforce structure, and a 4-step LCM model has very little room to reconcile a fight between them. Keep control_weight and the IP-Adapter scale modest (0.5–0.8) and expect to iterate.

It's a genuinely ambitious node for its era (late 2023, from taabata's early LCM wave), and when it works the speed-to-possibility ratio is impressive. Just don't expect it to behave like today's cleanly separated native IP-Adapter and ControlNet nodes - those are easier to debug. This one's for people who want the whole stack in a single loader and don't mind the 2023-era rough edges.