Part 10 ComfyUI from installation to mastery, ControlNet, three scenesYou learned the OpenPose/Dept/Canny, and you know how screwth. But I left a pit unfilled:When will these three be used
I've stepped on my own real experience: I've thought about changing a real photo to Sabbunk style, using Canny. As a result, the wrinkles and edges of the new clothes were “welded to death” and the style changes were so hard that they were like filtration filters. Then I found out that I had to do it Decth (Depth)- It recognizes the back and forth spatial level, not the edge line. This is the one that tells Depth alone。
I. Depth. What is it
One sentence:Depth controls the "who's in the front, who's in the back, what's the size" in the picture, not the edge line。ControlNet's family is the easiest place for it to mix with Canny
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Remember:Canny Pipeline, Decth Pipe Pipe Space. Want to change skin or bones, with Deepth。
ii. Documentation / Environmental readiness (the misdirected model does not exist)
Decth is more dependent than OpenPose, with the highest probability of falling. Two things to prepare:
1 ControlNet model weight(Same as Canny/OpenPose, here):
ComfyUI/models/controlnet/
SD1.5 Bottom model is used as control_v11f1p_sd15_depth_fp16; SDXL bottom model is used as diffuser_xl_depth_full。
2 DevthAnythingV2 Pre-processing Weights(This one's the easiest to miss, node code installed, ≠ weight added):
ComfyUI/kustom_nodes/comfyui_controlnet_aux/ckpts/depth-anyth/depth_anything_v2_vitl.pth
Trail experience:For the first time, I used Depth, only the ControlNet node, went down the ControlNet model, halfway to "model not found." Half a day, it was found that the .pth pre-processing weight was missing -- node added to the canvas, and it ran short of material. The file name must be exactly the same。
III. Hand-to-hand nodes (continuous)
And the basic chain of the 10th is a "pre-treatment" link: turn the reference into a Depth and feed it to ControlNet。
Load Image →DepthAnythV2Processor Depth
Zenium
ControlNet Loader (select depth model) + Depth + Positive and Negative Plugin
Zenium
KSampler
1. Add Node: Right-key search of DevthAnythingV2Preprocessor, ControlNet Loader, ControlNet Apply。
2. Connect: IMAGE of Load Image →Preprocessor image; control_net of IMAGE →Apply of Preprocessor; CONTROL_NET →Apply of Loader _net; positive/ negative alert of Apply。
3. Select Model: Loader selects control_v11f1p_sd15_depth_fp16 (SD1.5) or diffuser_xl_depth_full (SDXL)。
4. Scope: Apply has 0.65, Start_percent 0, End_percent 1。
One sentence:The reference diagram pre-processes the depth map through Decth, then goes to Apply Advanced with the ControlNet model and connects to the main chain。
Clean-up ring: 10 minutes, change the style of the photo with Deepth
Let's start with this one. After you've got an extra "Spatial Structure Same But New Style" map in your hand, you can confirm that Depth really works。
Preconditions: You already have the 10th of the ControlNet Basic Chain. If you lose it, press this 5th node。
Target: Take a photo of your real person, use the Decth control space, and change the style to "water colour / Sabpunk / Whatever" , the gestures, the necks, the contours of the original image are preserved, but the details of the clothes follow the style。
Step (sight, don't jump):
1. Down Model: Depth ControlNet (SD1.5 with control_v11f1p_sd15_depth_fp16) dropped in Models/controlnet/。
2. Lower Preprocessing Weight: below depth_anyth_v2_vitl.pth, unnamed, comfyui_controlnet_aux/ckpts/depth-anything/。
3. Add NodePhoto by DepthAnythingV2Processor + ControlNet Loader + ControlNet Apply (Advanced)。
4. Hey, referenceLoad Image selects your photo; Loader chooses depth model; Preprocessor sets 768 (balance)。
5. Set parameters (most critical)Apply enter 0.65, start_percent0, end_percent1。
6. Write hints: positive-written "for styles" (e. g. watercolor painting) + original map master description; fixed seeds (e. g. 12345)。
7. Output: Point Queue Prompt。
Checkpoint1 Generates a pattern/ contour that is consistent with the original map, but the style changes = Depth; 2 returns “mode not found” = model or .pth path error (returns to section II); 3 produces nothing to do with the original = Apply is not connected to the title chain (returns to section III, step 2)。
job (selection): move stength from 0.65 to 0.4 and run one more -- you'll find the attitude has become loose, the original image has weakened; to 0.85, it's almost original. that's how the knob feels。
core knob: stength and pre-treatment resolution
strength (control intensity)– Depth's most understandable spin button, multi-source measurements:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
preprocessing resolution- Impact details and visibility:
512: Use when display is tight, speed is fast but depth is rough
768: Balancing, mostly adequate
1024: High detail, but visible, 4060
V. Common pits (all at once)
1. ...pth pre-processing weights are omitted— Nodal energy plus, a run-off model does not exist (first pit, back to section II)。
2. Synchronising folder_depth_anything_v2_vitl.pth cannot be mislettered and the wrong folder can not be found。
3. settings 0- It's like being uncontrollable, creating a picture and a reference。
4. Use that, Canny. Change- Welded on the sideline and the style had to be hard; change the skin, please。
5. More than one ControlNet without visible—Depth + OpenPose at the same time, 4060 recommends a resolution below 768。
VI. QUIREMENTS OF PARLIAMENTS
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
VII. 4060 / 8G SPECIALIZED & I STEPPED ON PITS
Skills: Depth pre-treatment itself takes a bit of visibility, up 4060 resolute with 768; if the OpenPose is folded at the same time, the resolution of both ControlNet is down to 512, and the visible occupancy can be pressed. Model Selection fp16 versioncould not close temporary folder: %s starts with –medvram more steady。
Five pits say all at once: 1.pth (No. 1); 2 file name/path error; 3 stength set 0 unresponsive; 4 take Canny to replace it and cause welding; 5 stacks of ControlNet without direct resolution OOM。