LCMGenerate_inpaintv3

LCMGenerate_inpaintv3 is the other mega-node in the pack's inpaint family, and the direct sibling of inpaintv2. Same idea - reference-only steering plus structural conditioning plus 4-step LCM inpainting in one node - but it trades the T2I-Adapter for an IP-Adapter. If your workflow wants image-prompting (style from a reference picture) combined with ControlNet structure during inpaint, this is the one.

What's in it

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

Then the v3 additions, which is where it differs from v2:

• control_image and control_weight (default 1.0) - ControlNet conditioning. The pipe must come from LCMLoader_controlnet (or LCM_IPAdapter_inpaint's mode-loaded ControlNet) for this to actually do anything.
• ip_model - the IP-Adapter, loaded by LCM_IPAdapter or LCM_IPAdapter_inpaint.
• pil_image - the reference image for the IP-Adapter (yes, that's just the image input name).
• scale (default 1.0) - how hard the IP-Adapter pulls.

So compared to v2's adapter_image/adapter_weight/adapter trio, v3 carries ip_model/pil_image/scale. v2 = T2I-Adapter track; v3 = IP-Adapter track. The ControlNet input is shared by both.

How to wire it

LCM_IPAdapter_inpaint is the natural loader pairing here, because it loads the ControlNet and the IP-Adapter into the same pipeline - one loader, one generation node, and you've got the pack's full "reference + structure + image-prompt + inpaint at 4 steps" showpiece. Or combine LCMLoader_controlnet with LCM_IPAdapter if you prefer the pieces separate.

Setup and honesty

Standard pack install (ComfyUI Manager → search LCM_Inpaint-Outpaint_Comfy, or clone + pip install -r requirements.txt), then LCM_Dreamshaper_v7 in models/diffusers/, ControlNet folder under models/controlnet/, IP-Adapter file under models/controlnet/, CLIP vision encoder in models/clip_vision/.

The honest truth about v3 (and v2, and the pack's other combined nodes): stacking reference-only + ControlNet + IP-Adapter onto a 4-step distilled model is the hardest thing this pack asks of the model. Three steering signals, almost no denoising time to reconcile them. Dial scale and control_weight down - start both around 0.5–0.6 - and keep CFG at 1–2, or you'll get oversaturated incoherence. It works, and when you land it the speed is wild, but it's fiddly. This is very much late-2023 "look what I can do with LCM" code, and it's best treated as a fun experiment rather than a daily driver - today's separated native nodes are easier to reason about.