LCMGenerate_inpaintv2

LCMGenerate_inpaintv2 is the "give me everything" generation node of the pack: it takes the full LCMGenerate inpaint/outpaint feature set and bolts on both T2I-Adapter and ControlNet conditioning. If a workflow needs reference-only style steering, a structural condition, and 4-step LCM inpainting at the same time, this is the node the pack author reached for - it's the biggest single-node kitchen sink in this pack's later workflows.

What's in it

Everything from LCMGenerate: mode (Inpaint/Outpaint), seed, text, steps (4), width/height, cfg, image, mask, original_image, outpaint_size/direction, pipe, batch, prompt_weighting, reference_image, style_fidelity, Reference_Only, oupaint_quality.

Then the v2 additions:

• adapter_image and adapter_weight (default 1.0) - T2I-Adapter conditioning, fed by an adapter class input from LCMT2IAdapter.
• control_image and control_weight (default 1.0) - ControlNet conditioning, which needs the pipe from LCMLoader_controlnet (that's where the ControlNet actually gets loaded).

So in one node you can steer with a reference image (Reference_Only), a T2I-Adapter, and a ControlNet, all at once. It's a lot of dials, and that's the point - it's the pack's demo workflow node.

Where it sits vs the others

The pack's generation family forms a ladder. LCMGenerate is the base. LCMGenerate_img2img / _controlnet / _IPAdapter each add one technique to the img2img line. The v2/v3 pair are the merged mega-nodes for inpaint: inpaintv2 = reference-only + T2I-Adapter + ControlNet, and inpaintv3 = reference-only + ControlNet + IP-Adapter (they share a name-space and mostly swap which adapter family they carry). If your workflow has both ControlNet and a T2I-Adapter, v2 is your node.

Setup and the honest warnings

Same pack install (ComfyUI Manager, search LCM_Inpaint-Outpaint_Comfy, or clone + pip install -r requirements.txt), and the same model requirement: LCM_Dreamshaper_v7 in ComfyUI/models/diffusers/, T2I-Adapter and/or ControlNet in subfolders under ComfyUI/models/controlnet/.

The real warning is that "everything at once" is hard to keep stable. Reference-only, a T2I-Adapter, and a ControlNet are three separate steering mechanisms all fighting over a 4-step LCM model - stack them all at full weight and you'll get mush. The sensible order: pick one primary steer (usually ControlNet), keep its weight around 0.6–0.8, and run the others near zero. And drop CFG to 1–2; the pack's 8.0 default fights the whole point of the node.

This is ambitious late-2023 code from taabata's early LCM wave, and it's a fun showpiece of what was possible back then. But honestly, for real work most people today split these techniques across native ComfyUI nodes where you can debug each one. This node is for exploring the pack's combined workflows - and it's genuinely cool at 4 steps when you get it dialed in.