ComfyUI from installation to mastery, Workstream Practice Guide (read node chain + build + save reuse)

ComfyUI from installation to mastery, Workstream Practice Guide (read node chain + build + save reuse)

Previous (Page 04)ComfyUI from installation entrance to mastery, complete guide for model selection You've chosen the base model, you've done it. This is the most critical step:How do you connect a bunch of diamonds into a visual map。

I just used it ComfyUI Sometimes, people share itWorkflowIt's like reading in the sky -- it's full of bricks and colored lines, and I don't know where to start. Until I've mastered the mind method of working as ifPlant flow lineThe material goes in from the left, the finished product comes out from the right, and the line is the conveyor belt. This article teaches you three things: to read the default workstream, to make one out of the blanks, to save it and to re-use it。

I. Psychiatry model to create node + link first

ComfyUI's workstream is essentially a node chart:

Node: a square does a specific thing, such as "read the model" writing a hint "sampling out"。
LinkLines between squares, representing data flows, are coloured by data type - green is typically a model, blue is an image, orange is a subspace. It's only possible with a colored port, which automatically prevents you from getting connected。
Add Node_Other Organiserdouble clickPops up a search box and you can add a nodal name. This is the first thing a rookie should remember。

Remember:The data flow from left to right, and you follow the line to understand the whole line。

II. UNDERSTANDING A SET OF SET OF STANDARD PRODUCTS FROM THE Default Flow

Step 1: Load the default workflow。Top menu point (fast button Ctrl+D), puddle, and there's a series of nodes on the canvas, and this is the official graph template。

Workflow

Step 2: Recognize this node chain。Here's one from left to right:

Load Checkpoint
• CLIP Text Encode is positive/negative (writing hints)
Emtty Late Image
→ KSampler
VAE Decode
Save Image

node
What's it doing
Load Checkpoint
Bring in the base model, the base of the map
CLIP Text Encode
Up-and-up and down-to-back negative
Emtty Late Image
Set a blank canvas size in the subspace
KSampler
Core engines, by parameters
VAE Decode
Decoding subspace data into colour map
Save Image
save disk, drop output folder

iii. Make one of your own

1. Plus Road CheckpointDouble-click to search Checkpoint, right-hand ckpt_name。
2. Plus two CLIP Text Encode: One fill positive, one fill negative, connect the checkpoint CLIP output to their crip mouths。
3. Add Emtty Late Image: Set up width/head (SDXL 1024x1024)。
4. Add KSampler: CHeckpoint's MODEL with its model; two CLIP's CONDITITONING companies with positive/negative; Emtty Latet company with latent_image. Sets=20, cfg=7 and sampler=euler。
5. Add VAE Decode: KSampler's Latent Company Samles, Checkpoint's VAE Company Vae。
6. Plus Save Image: VAE Decode's IMAGE Company past。
7. Run: Top Queue Prompt (Ctrl+Enter)。

Small-line skills:Moves the mouse to the port, and drags to another same port and releases. A small dot on the wrong line can be disconnected. Don't be afraid to make a mistake。

iv. Save it and return it next time

For the first time, a strong recommendation was put in. WorkFlow →Save (Ctrl+S) produces .json, remembers the layout, double-clicks the next time。

Two different formats:Save (ordinary)It's got a layout, it's always thereSave (API Format)SAVES ONLY NODES (NUMBER ID) FOR AN INTERFACE. IF YOU DON'T HAVE A NODE FOR THE WORK STREAM THAT YOU'RE LEADING IN ONLINEThe mark is wrongWe'll just have to load the missing nodes with the Manager from Book 03. When I was playing, the big guy was working so red, it was a shock, it was all green。

V. Two commonly used paths (see first step)

SET A VINCENT + HIGH CLEAR ZOOM:The first generation KSampler → VAE Decode → Latent Upscale (expanded 1.5 to 2 times) → Second KSampler (denoise 0.4 to 0.6) → VAE Decode → Save. It's better than just showing up and details。

Portrait B + LoRA:Load Checkpoint → Load LoRA (model/crip passing through Checkpoint with a stength 0.3 to 1.0) → KSampler ... LoRA puts models/loras/ (Phone 09)。

VI. 3 Steps to Understand People ' s Complex Flows

1. Start with input node: Only output without input (Load Checkpoint / Empty Late / Load Image) is the starting point。
2. Follow the line: Enter and export the data from the input to the output along the line, and see which nodes the data goes through; the line is organized with the Reroute node。
3. Watch the output push backwards: Only the input unoutputed (Save Image / Preview) is the end point, delineating the whole chain; the misnotation point becomes red/purie, focusing on it and the front。

VII. 4060 / 8G OBVIOUS ATTENTION

Load big model nodes reddish, wrong, mostlyVISIBLE SURPLUS (OOM)I don't know. Simplifying the chain approach:

Approach
clarification
Start with Parameters
-lowvram-force-fp16
 There's been a lot of decline
Quantified model
fp8 / GGF Quantification (Flux Priority Quantification)
VAE DECODED
Change VAE Decode (Tiled) block decode
Streamlined links
Delete Preview, lower resolution, Unload Model

It's gonna work, it's gonna work, but why is somebody else's drawings beautiful and clear, and you're all grayTitle 06I'll teach youAltitude- RESOLUTION, STEP, SAMPLER, CFG, HIGH-RESOLUTION MAGNIFICATION, ONE ITEM。

statement:The content of the source of public various media platforms, if the inclusion of the content violates your rights and interests, please contact the mailbox, this site will be the first time to deal with.
TutorialEncyclopedia

A.A. Electrician graphic design case, with Codex doing the master chart, the details page and the sales of the book

2026-8-13 9:39:35

Information

Economists Warn of AI Investment Crash, Bubble Far Larger Than "Internet Bubble"

2025-7-20 12:17:30

Search