Title 11 ComfyUI from installation entrance to mastery, Decth dig deep + high-resolution magnification (supplementation)You control space with Decth, and you know the ControlNet tube position. But there's one more troublesome thing: a set of different scenes of the same role, the first round face, the second square face, the third straight hair change -- I've been staring at three pictures for half a day, not three separated twins. Then I realized:ControlNet Tube and Posture, IPAdapter Stewardship and StyleI don't know. If you want the same face, you have to rely on IPadapter, especially the FaceID series. That's what I'm talking about。
I. What the hell is IPAdapter
One sentence:Make a reference "encoding" a condition, place it in the bottom model, and leave it to the character of the reference map。
· Style Class(gypsy, poster)。
· Lockface(many faces each time): Yes FaceID SeriesIt's more specific to extracting human face features than the base version。
· DoubleOne pipe style, one pipe face, two IPadapter strings, no fights。
IT'S JUST AS IT'S NOW THE BIGGEST PAIN OF THE AI COMIC -- "SHARE A FACE" ON A HOT SEARCH. LOCAL ComfyUI + IPadapter, full, permanent free and unlimited, is the solution。
Remember:ControlNet position, IPadapter looks like you're the director。
ii. Documentation / Environmental readiness (dry in the wrong place)
Ipadapter doesn't run after the node, the model has to be put in the right place. The first thing I did was the model threw the wrong folder, the node was red, half an hour。
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Trail experience:I had only buffalo_l left Lora, and my face was still locked -- FaceID basically welded people's faces with Lora, and it wasn't enough with buffalo_l。
iii. Hand cuffing
Compared to the base chain of 05, the IPdapter is an additional "encoded reference map" link: the reference map goes past the IPdapter node and goes into KSampler with the bottom model。
Load Checkpoint
• CLIP — CLIP Text Encode (±) — KSampler.cond
• VAE → VAE Decode
Read Checkpoint
Ipadapter Unified Loader
Load Image (reference)
Advanced. MODEL
KSampler
1. Add Node: Right-click search Load Image, IPadapter United Loader, IPadapter Advanced。
2. Connect:
References
Modular MODEL first to Ipadapter United Loader, Loader Output to Advanced.Model
Advanced. MODEL KSampler.model(mode not cond
Unified Loader automatically pairs the Ipadapter model with the corresponding CLIP Vision。
3. Preset: Unified Loader chooses FACEID。
One sentence:Reference diagrams feed Advanced.image; Endimage Unified Loader to Advanced.Model; Endel out of Advanced collects Model from KSampler。Small tip: IPadapter sets the reference figure into a square to extract the feature, using as much as possible a square or a clear face。
Weld the star's face in 10 minutes
Let's start with this one. After you've got a bunch of "same face, but scenes/styles change."。
Preconditions: You've run Load Checkpoint → KSampler → VAE Decode → Save. This article adds nodes to it。
Target: Use a leading character reference to share the same face, the same style with all the following characters。
Step (sight, don't jump):
1. Down model right in position:
ipadapter models/ipadapter/
clip_vision → models/clip_vision/
buffalo_l models/insightface/models/buffalo_l/
FaceID LoRA → models/loras/。
2. Plugin:Manager search for "IPAdapter Plus" one key, restart。
3. Add Node: Load Image (reference diagram) + IPadapter United Loader + IPadapter Advanced。
4. Connect: MODEL for Load Checkpoint first to IPadapter Unified Loader, Loader output to Advanced.Model; Load Image to Advanced.image; Advanced.MODEL → KSampler.Model。
5. Participation: Unified Loader preset FACEID (lockface); Weight style 0.6~0.8, FaceID 0.9~1.0; start 0/ end 1. Queue Prompt。
Checkpoint
1 Face/style with reference diagrams, with no change in hint face = success
2 Pit 1 "face unlocked":light is too low, or Advanced is wrong (must pick up KSampler's model), or FaceID LoRA does not play models/loras/
3 Pit 2 "Facedown/Backback": FaceID does not have a buffalo_l, go back and replace section 2
4 pit 3 "significant bomb": 4060 folds –medvram with resolution down to 512。
job (selection): same reference, view 0.6, one each, and 0.9, see how much different face welding is。
IV. QUIREMENTS OF PARLIAMENTS
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Note: Model paths, node names are cross-checked by cubiq/CommyUI_IPAdapter_plus Official Document + Community Curriculum 2 Sources (validated 2026-07-21); between weights are community general sweet areas, finely adjusted upwards and downs。
V, 4060 user-specific
:: Use all models fp16don't touch fp32, double the balance。
• CLIP Vision also eats visible (about 1-2G) not to have too many reference nodes at the same time。
• FaceID's insightface reasoning is light, with a head in the bottom + crip_vision。
Resolution 512 x 768 Start;to magnify under 13。
• Multi-IPadapter serial display, run single and double; start plus –medvram.