Generative AI Full Course Lecture 6 Stack Depth Pose for Structural Control

Generative AI Full Course - Lecture 6: Stack Depth + Pose for Structural Control

Full tutorial link > https://www.youtube.com/watch?v=h4-F5Jx-HQo

Run ComfyUI in Colab, then direct structure and composition with ControlNet, depth, edges, pose and masks.

🔗 Links:

Windows AI requirements guide: [ https://www.patreon.com/posts/111553210 ]

📦 Full course resource repository: [ https://github.com/FurkanGozukara/Generative-AI-Tools-And-Tecniques-2026-2027-Fall ]

📺 Full course playlist: [ https://www.youtube.com/playlist?list=PLUMEJUep1hiI ]

SECourses Discord: [ https://discord.com/invite/software-engineering-courses-secourses-772774097734074388 ]

🎯 Learn cloud setup, uploads, downloads and backups, then compare controls and place subjects independently.

🧩 Qwen-Image, InstantX Union, Depth Anything 3, Canny, AnyLine, SDPose and SAM3.1.

Use the chapters to revisit each operation and comparison.

⏱️ Chapters:

00:00:00 ComfyUI in Colab and structural control

00:00:58 Upload, rename and save the notebook

00:01:30 Notebook saving versus temporary runtime files

00:02:30 Select a GPU and connect the runtime

00:03:03 Notebook forms, code and cell controls

00:03:44 Install ComfyUI and inspect detected hardware

00:04:19 Match diffusion model, encoder and VAE

00:04:49 Download models and inspect automatic choices

00:06:05 System RAM and GPU memory are separate

00:06:20 Choose direct model links and destination folders

00:07:02 Run an optional download and inspect its file

00:07:26 Start ComfyUI and open the temporary URL

00:08:05 Open the matched starter workflow and prompt

00:08:32 Generate remotely and watch the live progress

00:09:05 Inspect and download the completed image

00:09:44 Save and export the editable workflow JSON

00:10:16 Upload a reference image to the runtime

00:10:59 Upload a video into the input folder

00:11:17 Refresh the loader and inspect video playback

00:11:39 Reopen a workflow and reconnect its inputs

00:11:57 Stop the server before exporting a folder

00:12:13 Choose output subfolders and include inputs

00:12:46 Create and download the complete result ZIP

00:13:18 Inspect ZIP outputs, graphs and the manifest

00:14:20 Restart ComfyUI on the same runtime

00:15:35 Understand reset, disconnect and file persistence

00:16:03 Select and verify a paid A100 allocation

00:16:46 Compare automatic model choices across runtimes

00:17:02 Terminate the session and verify it ends

00:17:40 Update the local environment and restart ComfyUI

00:18:51 From references and masks to structural control

00:19:24 Trace ControlNet through the conditioning path

00:19:41 Distinguish model patches, LoRAs and references

00:20:48 Match original Qwen with InstantX Union

00:21:05 Download matched model and control weights

00:21:44 Install Depth Anything 3 and refresh loaders

00:22:35 Fix seed, steps and sampling conditions

00:23:17 Inspect the source and align its working size

00:24:05 Connect Depth Anything 3 and inference

00:25:12 Render grayscale depth before connecting generation

00:25:54 Inspect spatial layers and save the map

00:26:40 Load the compatible Qwen model family

00:26:56 Write the prompt and conditioning branches

00:28:02 Connect the union loader and apply node

00:29:20 Connect the matching VAE and output path

00:30:20 Generate the baseline without structural control

00:31:07 Restore depth control with unchanged settings

00:31:33 Compare source, depth and both generated images

00:32:31 Understand the VAE input and diagnose failures

00:33:45 Compare control strengths 0.5, 0.8 and 1

00:35:05 Inspect geometry and prompt adherence separately

00:36:03 Separate control strength from start and end

00:36:47 Generate end 0.3 and end 0.65

00:37:51 Compare early release with full-schedule control

00:38:49 Choose edges, line detail, depth or pose

00:39:50 Install the auxiliary package and run AnyLine

00:41:22 Set up SDPose and choose joint drawing

00:43:37 Inspect all four maps at the same scale

00:43:57 Generate with each structural map type

00:46:12 Compare the results against their actual maps

00:47:13 Enlarge Canny and adjust its thresholds

00:48:24 Compare detailed and simplified edge maps

00:48:59 Inspect pose joints against the source photograph

00:49:57 Chain depth and pose conditioning correctly

00:50:56 Test both controls at full strength

00:51:40 Balance depth 0.5 with pose 0.6

00:52:23 Compare results and diagnose competing constraints

00:53:45 Prompt an explorer and robot in one scene

00:54:40 Inspect placement, attributes and editability

00:55:50 Generate separate subjects and inspect their quality

00:57:16 Extract masks and place the subjects independently

00:59:13 Match scale, perspective and contact points

01:01:14 Repair the visible contact-layer rectangle

01:01:29 Pad the soft mask with empty space

01:02:35 Match image and mask sizes before repositioning

01:03:56 Move the robot while keeping the explorer fixed

01:04:41 Repair the boot contact and inspect grounding

01:05:25 Compare single generation with editable composition

01:06:09 Export subjects, masks and the final graph

01:07:11 Reopen saved projects and use the learner files

Materials include the notebook, graphs, inputs, maps and results. Download model weights separately.

💬 Ask in the comments or Discord. Support SECourses on Patreon for more guides and resources.

#ComfyUI #ControlNet #GoogleColab

Video Transcription

• 00:00:08 You can follow this course without owning a  powerful graphics card. We will run a familiar  

• 00:00:14 ComfyUI workflow in Google Colab, preserve its  files, then use structural controls to direct  

• 00:00:19 shape and composition. Colab provides a remote  computer. Your browser shows its notebook and  

• 00:00:29 the ComfyUI interface, while the model runs on  the allocated GPU. Uploading an image sends it  

• 00:00:35 to that remote machine. The course materials  include this notebook and a practical guide.  

• 00:00:46 Keep a local copy as a clean starting point  for future runtimes. The guide explains the  

• 00:00:51 setup order and which files need a backup. Open  Google Colab and choose Upload notebook. Browse  

• 00:01:01 to your local notebook file and open it. Colab  imports the document; it has not yet installed  

• 00:01:07 ComfyUI or downloaded the model weights. Give  your copy a recognizable name. This copy belongs  

• 00:01:19 to the selected Google account, so check the  account before opening or saving notebooks,  

• 00:01:24 especially when you switch between free and  paid accounts. The document and the computer  

• 00:01:31 have different lifetimes. Saving the notebook  preserves its instructions and settings. Files  

• 00:01:37 created under the runtime's content directory need  their own backup before that machine disappears.  

• 00:01:58 We will first use a free account with a T4,  then show the paid account's A100 selection.  

• 00:02:04 Availability can change, and the GPU you request  still needs to be allocated. Google restricts  

• 00:02:14 some external web interface use on free managed  runtimes. A successful session does not guarantee  

• 00:02:19 future access. If Colab stops it, keep your  exported files and use a permitted runtime or  

• 00:02:25 the local route. Open Runtime, then Change runtime  type. Select the T4 GPU and save the choice. Leave  

• 00:02:36 High RAM off for this first pass so the displayed  resources describe the ordinary configuration.  

• 00:02:42 Connect the runtime and wait for its connection  state. A GPU name in a menu is a request; the  

• 00:02:48 hardware report from the running machine is what  confirms the device we received. These sections  

• 00:03:03 contain code cells with convenient forms. The form  controls are settings for that cell. Revealing the  

• 00:03:09 code lets you inspect what will run without  needing to rewrite the installation yourself.  

• 00:03:16 Change the optional link field once, then click  outside it. The corresponding value in the source  

• 00:03:22 changes with the form. Save the notebook after  editing, and leave unwanted download fields empty.  

• 00:03:32 Run all is less useful here because the server  cell stays running while you use ComfyUI. It also  

• 00:03:38 includes optional downloads. We will deliberately  run each required stage and inspect its result.  

• 00:03:44 Keep the latest stable release selection and run  the installation cell. It detects the hardware,  

• 00:03:50 installs ComfyUI and Manager, and prepares  the tunnel used to reach the interface from  

• 00:03:55 this browser. Read the detected GPU and the  installed version in this output. These are  

• 00:04:07 the conditions of our session. Installation time  depends on downloads and the machine, so wait for  

• 00:04:13 the completion message before continuing. This  starter uses Qwen-Image 2.1. Its diffusion model,  

• 00:04:23 text encoder and autoencoder have separate jobs  and separate files. Keep their exact compatible  

• 00:04:29 filenames together when you share a working  recipe. The smaller text encoder option saves  

• 00:04:35 memory through quantization. That is a storage  and computation choice, not a promise that every  

• 00:04:41 prompt produces an identical image to the larger  encoder's output. Run the required model cell and  

• 00:04:51 inspect its completion. If a download fails, read  which file failed before retrying. A partial file  

• 00:04:57 or a missing model is not fixed by changing the  image prompt. The model cell has an automatic  

• 00:05:03 selection. It considers both system memory and GPU  memory, then prints the actual files it chooses.  

• 00:05:09 That printed choice is more informative than  guessing from a GPU's name. Run cells individually  

• 00:05:16 from top to bottom. The running cell shows its  activity and output; the square stops that cell.  

• 00:05:22 Interrupting an installation midway may leave  work that its error message needs you to repair.  

• 00:05:33 The downloaded weights live inside this runtime.  Reusing them during the same session avoids  

• 00:05:39 another transfer, but a new machine may need the  downloads again. Saving the notebook alone does  

• 00:05:45 not preserve them. The runtime uses Linux paths.  Its content folder is on the remote computer,  

• 00:05:54 while the notebook we downloaded is on our  own computer. A local drive letter cannot  

• 00:06:00 identify a remote input file. System RAM  and GPU memory are separate resources. A  

• 00:06:08 large amount of system RAM does not turn into the  same amount of fast GPU memory, although a model  

• 00:06:14 may move some data between them. The optional  downloader accepts direct model file links.  

• 00:06:24 Open the model's official file page and use its  download link, rather than copying the address of  

• 00:06:29 a page that only describes the repository. Here  the example is an upscaler, so its destination  

• 00:06:36 is the upscale models folder. A diffusion model  belongs in its own loader's folder. Moving every  

• 00:06:42 model into checkpoints would hide some from the  intended loaders. Leave the token field empty  

• 00:06:48 for these public downloads. An access token  is only relevant when the chosen repository  

• 00:06:53 requires authorization. Do not save a private  token in a notebook you intend to share. Run  

• 00:07:02 this optional example and inspect its destination  in the Files panel. The file should be inside the  

• 00:07:08 expected model category. Refresh ComfyUI's model  list after adding a file to a running session.  

• 00:07:26 Start the server cell and wait for the printed  web address. That address opens the ComfyUI  

• 00:07:31 instance on this runtime. Keep the cell running  while we work in its separate browser tab.  

• 00:07:51 The tunnel connects this browser to the remote  server. It is temporary and may change after a  

• 00:07:56 restart. Bookmarking an old tunnel address  does not provide a permanent hosted ComfyUI  

• 00:08:02 installation. Open the supplied starter workflow  from Workflows. Its model selections should match  

• 00:08:10 the completed downloads. If a loader shows  a missing file, return to the dependency and  

• 00:08:16 destination before trying to run it. Replace  the starter prompt with our workshop scene:  

• 00:08:22 1 male explorer, a workbench in the foreground,  and a bright window farther away. A clearly  

• 00:08:27 arranged scene gives us useful structure to  discuss later. Keep the demonstrated size and  

• 00:08:33 sampling settings, then run the workflow. We  already studied these graph controls locally,  

• 00:08:39 so concentrate here on where the work runs  and where its files are stored. Watch the  

• 00:08:45 live progress until the image appears. Download  and model loading contribute to the first run,  

• 00:08:51 while later runs can reuse loaded resources.  We are checking a usable route, not claiming  

• 00:08:56 a hardware speed ranking. Open the completed  image large enough to inspect the person,  

• 00:09:09 the workbench and the background. A successful  execution only confirms that the graph ran;  

• 00:09:15 the image still needs to satisfy the prompt you  gave it. Use the image's download control and  

• 00:09:21 complete the browser download. Open the file  locally to verify it arrived. This is now a  

• 00:09:26 copy on our computer, independent of whether  the remote session survives. Save the workflow  

• 00:09:45 in the Workflows panel, then export its JSON  as well. The saved graph records the recipe,  

• 00:09:50 but it does not package every model or  input file that the recipe depends on.  

• 00:10:15 An image upload makes a local file available  to the remote graph. Add a Load Image node,  

• 00:10:21 use its upload button and choose our workshop  

• 00:10:23 reference. Its preview confirms  which image the node will provide.  

• 00:10:46 That image becomes a reference, an initial image  or a control source only through its connections.  

• 00:10:52 A text to image starter does not gain an editing  feature merely because an image sits beside it.  

• 00:10:58 For several files or a video, open Colab's Files  panel and enter ComfyUI's input folder. Upload  

• 00:11:04 the file there, then select its filename  in a compatible loader after refreshing  

• 00:11:09 the available choices. This video loader  reads the uploaded clip from the runtime.  

• 00:11:20 Inspect its preview before connecting it  to anything else. Uploading a video does  

• 00:11:25 not make an image model understand motion  or temporal references. Open an exported  

• 00:11:40 workflow through ComfyUI's normal Open command.  Reconnect the intended input filename and check  

• 00:11:46 the preview. A workflow from another machine can  refer to a file that has not been uploaded here.  

• 00:11:57 Individual downloads are convenient for 1  result. To preserve a session's work together,  

• 00:12:03 save the graph first and return to the notebook.  Stop the server cell and wait for its cleanup  

• 00:12:09 message. The final export cell creates a ZIP from  outputs and saved workflows. Leave the subfolder  

• 00:12:18 blank for everything, or enter a relative output  subfolder when you only want 1 experiment's  

• 00:12:24 results. The input checkbox includes your uploaded  source files when needed for reproduction. Model  

• 00:12:34 weights are excluded because they are large  and should retain their original download and  

• 00:12:40 license information alongside the workflow. Run  the export and read the exact archive path it  

• 00:12:49 prints. In the Files panel, expand exports, open  that ZIP's menu and choose Download. Complete  

• 00:12:55 Save As if your browser presents it. Open the  downloaded ZIP locally. Check the output files,  

• 00:13:21 saved workflow and manifest before ending the  session. A finished archive on the remote machine  

• 00:13:26 is not yet a backup on your own computer. The  output folder contains the rendered images,  

• 00:13:34 while the workflow files preserve  the editable graphs. Check both:  

• 00:13:38 an image is useful evidence of a result, but it  cannot replace the graph and its model selections.  

• 00:14:02 The manifest lists the files actually included in  

• 00:14:05 this archive. It records their relative  paths and sizes, including the uploaded  

• 00:14:10 image and video. Use this list to check  the backup before releasing the runtime.  

• 00:14:21 Restart the server cell on this same  runtime. Use the newly printed address,  

• 00:14:25 then reopen the saved workflow  from Workflows. Stopping the  

• 00:14:29 server process normally leaves  this machine's files in place.  

• 00:15:35 That is different from replacing the  runtime. Changing the GPU, resetting the  

• 00:15:39 machine or terminating it can remove the content  folder. Export useful work before those actions,  

• 00:15:46 even when the notebook itself is saved. Closing a  tab is also different from terminating a runtime.  

• 00:15:52 It may remain allocated, or the service may later  reclaim it. Neither behavior is a reliable storage  

• 00:15:58 plan for your only copy of an output. Choose the  A100 in Change runtime type and connect. Inspect  

• 00:16:07 the machine that is actually allocated.  A paid account offers different choices,  

• 00:16:12 but the requested hardware still  depends on current availability.  

• 00:16:47 Inspect the automatic encoder choice  again. With different available memory,  

• 00:16:51 auto may choose different weights. That means  comparing these 2 runs directly would mix a  

• 00:16:57 hardware change with a model configuration  change. When finished, open Runtime and Manage  

• 00:17:04 sessions. Terminate this session and verify it  disappears. This ends the allocated computer;  

• 00:17:10 our exported files and saved notebook remain  available separately. You now have the cloud  

• 00:17:24 setup and file handling route. The rest of this  lesson uses local ComfyUI, where we can build and  

• 00:17:30 inspect structural controls without depending  on a temporary remote session. Before changing  

• 00:17:40 the local workflow, apply the update routine  from Lecture 2. Open the existing ComfyUI  

• 00:17:46 folder and its command prompt, then activate  the environment used by that installation.  

• 00:18:05 Update the repository and install its requirements  in that environment. Keep the output visible until  

• 00:18:11 it completes. If an update reports a conflict,  resolve that actual conflict before assuming  

• 00:18:17 the application is current. Restart ComfyUI  and inspect the loaded interface. Updating  

• 00:18:25 files while the old process continues running  would leave us testing the previous process,  

• 00:18:30 so the restart is part of verifying  the change. We will keep our inputs,  

• 00:18:35 outputs and workflows organized for this lesson.  The original source image remains unchanged,  

• 00:18:41 and each comparison will retain the settings  that explain how its result was made.  

• 00:18:50 In Lecture 4, a reference helped describe what  we wanted. In Lecture 5, a mask specified where  

• 00:18:56 a change could happen. Structural control adds  a constraint on shape, pose, outlines or spatial  

• 00:19:03 arrangement. The word control does not identify  1 universal connection. Read the working graph  

• 00:19:12 from its loader to the point where the extra  information enters. That path tells you which  

• 00:19:18 mechanism the recipe uses. Here the ControlNet  file is loaded from the controlnet folder. Its  

• 00:19:28 apply node receives an image and modifies the  positive and negative conditioning passed to  

• 00:19:34 sampling. That is the route we will build. A  model patch uses a different loader and changes  

• 00:19:43 the model path. Its filename may still contain  control, but attaching it to a ControlNet loader  

• 00:19:49 would not turn it into this same topology. A LoRA  also changes a model through its own loader. A  

• 00:20:00 control LoRA must follow the recipe it was trained  for. Ordinary style LoRAs are not automatically  

• 00:20:06 structural controllers just because both use LoRA  files. Compare those routes with the reference  

• 00:20:15 graph from Lecture 4. The reference enters the  model's supported conditioning path. A reference  

• 00:20:21 can influence appearance or content without acting  like a pixel aligned depth map. Choose a control  

• 00:20:32 by its job and supported model family, then follow  the exact loader and apply nodes. Similar names  

• 00:20:38 and matching file extensions are not sufficient  evidence that 2 components belong together. For  

• 00:20:48 this example, the InstantX union file matches  original Qwen-Image. The notebook's Qwen-Image  

• 00:20:54 2.1 is a different model. We will load the matched  original base for these structural comparisons.  

• 00:21:05 Open the official workflow's dependency links.  Download the original Qwen-Image diffusion weights  

• 00:21:10 and the InstantX union ControlNet. Their  model pages identify the supported family,  

• 00:21:16 which is more useful than a generic control model  label. Put the diffusion file in diffusion models  

• 00:21:29 and the union file in controlnet. We already have  the matched Qwen text encoder and autoencoder from  

• 00:21:36 the earlier installation; verify their exact  selections rather than downloading duplicates  

• 00:21:42 unnecessarily. The depth preprocessor  needs its own weights. Download Depth  

• 00:21:48 Anything 3 mono large into geometry estimation.  It estimates scene structure from our photograph;  

• 00:21:54 it is separate from the model that  will generate the finished picture.  

• 00:22:02 Refresh the model lists after placing the files.  Each loader should now show the intended filename.  

• 00:22:09 If it does not, check that installation's  configured folders before making another  

• 00:22:14 copy somewhere unrelated. The official  example establishes a compatible graph,  

• 00:22:22 but accelerated recipes can add their own adapters  and sampling choices. We will use the original  

• 00:22:28 base without a speed LoRA so our control schedule  comparison stays straightforward. Our comparison  

• 00:22:35 will use 30 sampling steps and the same seed,  prompt, dimensions and guidance. Changing only the  

• 00:22:41 declared control variable makes the differences  easier to interpret within this specific recipe.  

• 00:22:49 These settings are an experimental starting  point, not a rule for every model. When you  

• 00:22:54 move to another family, use its documented  sampling recipe and verify the graph before  

• 00:22:59 copying our numerical choices. Keep the download  links and model names beside the saved workflows.  

• 00:23:07 A small workflow JSON can reproduce the wiring  and settings, but its recipient still needs the  

• 00:23:13 corresponding model files and node implementation.  Our source is 1 male explorer in a workshop. The  

• 00:23:20 foreground workbench, the standing figure and  the distant doorway provide different depth  

• 00:23:25 layers. The raised arm also gives us a useful  pose to compare later. Load the original image  

• 00:23:37 and inspect it at a readable size. We want a  clear source before extracting any map. Cropping  

• 00:23:43 away the feet or obscuring the raised hand would  remove information the preprocessor needs. Set 1  

• 00:23:50 common working size for the photograph and its  maps. Keep the aspect ratio and crop aligned. A  

• 00:23:56 shifted or stretched map changes the geometry even  when the control strength stays the same. Add Load  

• 00:24:06 Depth Anything 3 and select the downloaded model.  Connect it to the inference node, then connect our  

• 00:24:12 image. The preprocessor receives pixels; it does  not need the generation prompt to estimate depth.  

• 00:24:46 The inference resolution controls how  finely the depth model examines the  

• 00:24:51 source. We will use a practical detailed  setting, then inspect the map. Raising it  

• 00:24:56 costs memory and does not guarantee  that every boundary becomes correct.  

• 00:25:12 Connect the geometry output to the depth  renderer and choose the grayscale depth  

• 00:25:17 view. Use the v2-style normalization  for this control recipe, rather than  

• 00:25:22 feeding a colorful visualization  or unscaled distances into it.  

• 00:25:36 Send that result to Preview Image before  connecting a generator. This separates 2  

• 00:25:41 questions: whether the preprocessor produced  a useful map, and whether the generative model  

• 00:25:47 responds well to that map. Run the map branch and  enlarge its preview. Compare the foreground bench,  

• 00:25:59 the person and the background opening. We  are checking relative layers and boundaries,  

• 00:26:04 rather than assuming this single image provides a  measured 3D survey. Depth estimation infers hidden  

• 00:26:12 geometry from appearance. Reflective surfaces,  thin structures and ambiguous overlaps can be  

• 00:26:17 wrong. A believable grayscale picture is not  enough; inspect the regions that matter to your  

• 00:26:23 intended result. Save the map with the source  image. That preserves the actual constraint  

• 00:26:30 used in our comparisons and lets someone inspect  a result without rerunning the preprocessor under  

• 00:26:36 a potentially changed software version. Load  the original Qwen diffusion model, the matched  

• 00:26:42 text encoder and the Qwen autoencoder. Keep their  filenames visible while connecting them. The model  

• 00:26:48 family links these components, not the fact that  their sockets fit. Write a prompt for a futuristic  

• 00:26:58 research workshop with 1 male explorer in a blue  jacket. We are deliberately changing the scene's  

• 00:27:04 appearance while asking the structural control  to retain useful arrangement from the source.  

• 00:27:11 The text encoder provides positive conditioning  for that description. Keep the negative branch  

• 00:27:16 consistent across the experiment. Guidance,  prompt wording and negative conditioning are  

• 00:27:22 separate variables that we will not change  during the control sweep. Connect the model  

• 00:27:28 through the sampling configuration and into the  sampler. Set the latent size to the same working  

• 00:27:34 aspect ratio. A different framing would make the  correspondence with our source harder to judge.  

• 00:27:47 Use a fixed seed and 30 steps. Keep the seed's  after generation behavior fixed as well;  

• 00:27:52 otherwise a new seed could explain a difference  that we mistakenly attribute to control strength.  

• 00:28:02 Add the ControlNet loader and select the InstantX  union file. Union means this file supports several  

• 00:28:08 kinds of structural hints. It does not mean it  accepts every model family or every possible map  

• 00:28:14 representation. Connect its output to Apply  ControlNet Advanced, then connect the depth  

• 00:28:43 image. The positive and negative conditioning  enter this apply node, and its 2 conditioning  

• 00:28:48 outputs continue toward the sampler. Trace both  branches into the sampler before running. Loading  

• 00:29:10 a control file without placing its apply node on  the active conditioning path leaves the sampled  

• 00:29:16 graph without that intended structural influence.  Connect the matching autoencoder to the apply  

• 00:29:22 node's VAE input. This control route needs the  image represented in the appropriate latent space;  

• 00:29:29 the same compatible autoencoder also decodes  the sampled latent afterward. Decode the sampler  

• 00:29:35 output and connect Save Image. Keep the depth  preview visible beside the final image path.  

• 00:29:40 We can now inspect the source constraint and the  generated result without confusing which branch  

• 00:29:47 produced which picture. Start with a control  strength of 0.8 across the full schedule. This  

• 00:29:54 is a useful baseline within the model author's  suggested range, and the later sweep will test  

• 00:29:59 whether it suits our particular scene. Before  comparing anything, check the complete route once:  

• 00:30:08 source image, aligned map, compatible control,  conditioning path, sampler and decoder. A  

• 00:30:14 correctly connected graph gives the experiment  a meaningful starting point. Generate the prompt  

• 00:30:20 without structural control first, keeping the  seed and sampling settings fixed. This gives  

• 00:30:26 us a reference for what the text and model  choose when the depth constraint is absent.  

• 00:30:52 Inspect that result for the prompt's subject,  setting and style. Its arrangement may differ  

• 00:30:58 from the source because the prompt  describes content without specifying  

• 00:31:03 every spatial relationship in the workshop  photograph. Restore the depth control at 0.8  

• 00:31:09 and run the same settings. We are adding a spatial  constraint while preserving the other inputs,  

• 00:31:15 so compare corresponding regions rather than  unrelated decorative details. Place the source,  

• 00:31:34 depth map and 2 generated images together.  First inspect the workbench's position and the  

• 00:31:39 background opening, then the person's location  and silhouette. Those are concrete tests of  

• 00:31:45 layout adherence. Next inspect the requested  blue jacket and futuristic setting. Stronger  

• 00:31:52 structural adherence is not automatically a better  overall result if the constraint also prevents an  

• 00:31:58 important part of the new prompt from appearing.  The depth map describes distance relationships,  

• 00:32:04 not the person's identity. A similar position or  silhouette does not prove that the generated face  

• 00:32:10 matches the source. Keep identity and structure as  separate evaluation criteria. Record which result  

• 00:32:17 serves the actual brief. For a room redesign,  arrangement may matter most. For a new character  

• 00:32:24 illustration, you may accept looser background  geometry to preserve the requested appearance and  

• 00:32:29 expression. Return to the connected VAE socket on  the apply node. This is easy to overlook because  

• 00:32:36 some other control recipes do not use the same  input. Follow this model's requirements instead  

• 00:32:42 of a remembered generic graph. The autoencoder  translates between image pixels and the model's  

• 00:32:48 latent representation. Here it helps put the  control image into the representation expected  

• 00:32:54 by the control network; decoding the final sample  is another use of it. If you encounter the missing  

• 00:33:00 VAE message, inspect this connection and its  model family first. Increasing the control  

• 00:33:06 strength cannot repair an absent input or make an  incompatible autoencoder's representation match.  

• 00:33:13 For a shape mismatch, check the base model,  control model, autoencoder and map dimensions  

• 00:33:18 together. Read the failing node's actual message.  Changing random values can hide the original cause  

• 00:33:25 without producing a reproducible repair. Our  working graph keeps the connection explicit.  

• 00:33:30 Save it in the workflow rather than relying on  a verbal reminder, so the next person can trace  

• 00:33:36 the same input from its loader to both places it  is used. Strength scales the influence of this  

• 00:33:47 structural constraint. We will compare 0.5, 0.8  and 1 using the same map and seed. These values  

• 00:33:55 belong to this recipe, rather than a universal  strength scale. Set 0.5 and generate the first  

• 00:34:02 comparison. Watch for room to reinterpret the  source: the figure and major surfaces may still  

• 00:34:08 follow its layout while smaller boundaries become  less tightly constrained. Return to 0.8 for the  

• 00:34:23 middle condition. Reuse the verified result if  its complete inputs are unchanged. Cached work  

• 00:34:29 is useful, but only when the settings actually  match the condition named in the comparison. Set  

• 00:34:36 strength to 1 and generate the third condition.  Higher influence can preserve more source  

• 00:34:41 structure, but it can also make the image stiff or  interfere with the changes requested by the text.  

• 00:35:05 Inspect the 3 images at the same scale. Follow  the raised arm, the workbench edge and the distant  

• 00:35:11 opening across them. Compare those features  before allowing a brighter color palette to  

• 00:35:16 decide which image looks best. Then check the  prompt requirements separately. Does the blue  

• 00:35:22 jacket remain clear, and does the room read as  the new setting? This separates adherence to  

• 00:35:29 the constraint from adherence to the requested  transformation. Choose the least restrictive  

• 00:35:35 strength that meets the important geometry  requirements. If every tested value damages the  

• 00:35:40 result, reconsider the map or model pairing before  assuming an even stronger setting will solve it.  

• 00:35:47 1 seed is a controlled example, not a  population study. For a production decision,  

• 00:35:53 repeat the promising settings on additional  seeds and difficult inputs. Keep the chosen  

• 00:35:58 criterion consistent rather than selecting  only flattering examples. Strength answers  

• 00:36:04 how much influence the control has while active.  Start and end percent answer when it is active  

• 00:36:10 along the denoising schedule. They do not directly  represent a percentage of finished visual detail.  

• 00:36:17 Keep the start at 0 and strength at 0.8. We  will compare stopping at 0.3, 0.65 and 1,  

• 00:36:24 with the prompt, map, seed and sampling recipe  unchanged. Before we run the first case, pause and  

• 00:36:31 predict. If control stops after the first 30% of  the schedule, will the broad room layout survive,  

• 00:36:38 and which smaller boundaries might become freer  to change? Set the end to 0.3 and generate. Early  

• 00:36:50 guidance can establish broad arrangement, while  later sampling continues without this control.  

• 00:36:56 The final image will tell us how much of  the imposed structure survives. Now extend  

• 00:37:17 the end to 0.65 and run the middle condition.  More of the schedule receives the constraint,  

• 00:37:23 giving us a way to compare structure  retention with freedom in texture and detail.  

• 00:37:51 Finally use the full schedule condition. Place  the 3 results together and inspect large geometry  

• 00:37:57 first, then the arm outline, material details  and lighting. Separate a preserved layout from  

• 00:38:03 an identical surface appearance. Return to your  prediction. The comparison shows what happened  

• 00:38:09 for this model, sampler and seed; it does not  establish that every model assigns the same kind  

• 00:38:15 of visual work to the same schedule fraction.  An early release can be useful when the general  

• 00:38:21 arrangement matters but a traced appearance does  not. If it loosens a critical boundary too much,  

• 00:38:27 lengthen the control interval or revise the  map deliberately. Save both strength and  

• 00:38:33 schedule with the selected result. 2 images  labeled with the same strength can still  

• 00:38:38 use different constraints if one applies that  strength for only part of the sampling process.  

• 00:38:49 We will extract 4 kinds of guidance from the  same photograph. Keeping the source and crop  

• 00:38:55 identical lets us compare what information  each map retains, instead of confusing a new  

• 00:39:00 photograph with a new control mechanism. Canny  marks strong image edges. Add the Canny node  

• 00:39:07 and preview its output. It can preserve  silhouettes and architectural boundaries,  

• 00:39:11 but it can also capture textures that are  irrelevant to the structure you actually want.  

• 00:39:50 For the soft edge example, install the  auxiliary preprocessor package through  

• 00:39:54 Manager. Restart after installation and  verify its node appears. A custom node is  

• 00:40:00 executable software, so use the documented  package from its established source.  

• 00:40:22 Choose AnyLine for this union model's  demonstrated soft edge route and connect  

• 00:40:27 the same aligned image. Its learned and  line based processing produces a different  

• 00:40:32 constraint from Canny; preview the actual  result rather than relying on the label.  

• 00:41:08 The depth map retains relative spatial  layers while discarding much of the  

• 00:41:13 source texture. It is useful when the room  arrangement matters more than reproducing  

• 00:41:18 every line on a jacket or object surface.  For pose, download the documented SDPose  

• 00:41:24 checkpoint into checkpoints and select it  in its loader. This model extracts joints;  

• 00:41:30 it is separate from the Qwen model  that creates our finished image.  

• 00:42:26 Connect the checkpoint's model and  autoencoder to the keypoint extractor,  

• 00:42:30 then connect the same source image.  The drawing node converts detected  

• 00:42:35 joints into the skeleton image that  the control network will receive.  

• 00:43:10 Keep body, hand and face drawing  choices explicit. A map with only  

• 00:43:14 body joints carries different  information from one that also  

• 00:43:18 constrains fingers and facial landmarks.  Include only the detail your brief needs.  

• 00:43:36 Run the preprocessing branches and  inspect all 4 maps together. Canny  

• 00:43:41 and soft edges emphasize contours,  depth describes spatial layers,  

• 00:43:46 and pose concentrates on the person's joint  arrangement instead of the complete environment.  

• 00:43:57 Connect the Canny map to our existing  union apply node and generate with  

• 00:44:02 the comparison settings. The base model  and prompt remain fixed. We are changing  

• 00:44:07 the kind of structural information supplied to it.  

• 00:44:41 Repeat with the soft edge map,  then depth and pose. Save each  

• 00:44:45 result under its map type. The  union model supports these routes,  

• 00:44:49 but the useful strength can still differ  between different kinds of guidance.  

• 00:46:12 Read the result grid against the maps. Inspect  the workbench and doorway for scene geometry,  

• 00:46:17 then the raised arm and leg positions for  pose. A single overall beauty score would  

• 00:46:23 hide these different capabilities. Pose can  retain a gesture while allowing substantial  

• 00:46:29 freedom in clothing and background. Edges  can preserve more visible boundaries.  

• 00:46:34 Depth can retain the room's arrangement  without demanding every original contour;  

• 00:46:39 judge the actual outputs beside those  expectations. Choose the map that carries  

• 00:46:44 the information your task requires. If the goal is  a new room with the same perspective, background  

• 00:46:50 depth matters. If the goal is a new character  making this gesture, pose may be more relevant.  

• 00:47:13 Enlarge the Canny preview before changing  its thresholds. We need to see both the  

• 00:47:18 active values and the resulting lines. A tiny map  thumbnail makes it difficult to tell a meaningful  

• 00:47:24 change from display noise. Lower thresholds admit  more edges, including weak changes in texture and  

• 00:47:31 lighting. Run the map after the adjustment and  inspect the jacket, bench surface and window.  

• 00:47:37 More visible lines are not automatically more  useful guidance. Raise the thresholds for a  

• 00:47:56 cleaner alternative, then inspect which important  boundaries survive. If an arm or bench edge  

• 00:48:02 disappears, the simplified map may have removed  information that the control needed to preserve.  

• 00:48:24 Place the 2 maps beside the source  at the same scale. We want the major  

• 00:48:29 silhouette and useful room boundaries,  with less distracting surface texture.  

• 00:48:33 Choose the balance for this brief  before generating another image.  

• 00:48:43 Check that the threshold pair remains meaningful  for this implementation. The low and high values  

• 00:48:50 work together in edge selection. Keep  their relationship understandable,  

• 00:48:54 rather than presenting 2 unexplained numbers  as a magic preset. Now enlarge the pose map and  

• 00:49:01 compare joints with the photograph.  Follow shoulder to elbow to wrist,  

• 00:49:05 then hip to knee to ankle. A plausible  skeleton can still put a joint on the wrong  

• 00:49:10 visible location. Confidence filtering  can suppress uncertain detections,  

• 00:49:20 but it cannot invent a missing correct  joint. If an essential limb is wrong,  

• 00:49:25 improve the source or correct the pose with a  suitable editor before trusting a stronger control  

• 00:49:31 setting. All these maps must share the source's  framing. A map shifted by a crop or stretched  

• 00:49:37 to another aspect ratio can impose the wrong  geometry even when the detector itself worked  

• 00:49:43 correctly. This is why map inspection comes  before repeated sampling. If the constraint  

• 00:49:48 encodes the wrong arm or noisy surface texture,  increasing its strength asks the generator to  

• 00:49:55 follow that error more insistently. Depth and pose  describe different parts of our brief: the room's  

• 00:50:02 spatial arrangement and the explorer's gesture.  Combining them can help, provided their source,  

• 00:50:07 crop and intended geometry agree with each  other. Add a second Apply ControlNet Advanced  

• 00:50:14 node after the first. Feed the first node's  positive and negative outputs into the second,  

• 00:50:19 then send the second pair to the sampler.  Both constraints now lie on the active path.  

• 00:50:39 Use the aligned depth map in 1 node and the  pose map in the other. Connect the compatible  

• 00:50:44 autoencoder to both. Reusing the union loader  is valid here because the 2 nodes use that same  

• 00:50:50 supported control file. Set both strengths  to 1 for our first combined condition. Keep  

• 00:51:00 their full schedules and the same seed. This  deliberately tests a demanding constraint pair,  

• 00:51:05 so inspect the finished image  for rigidity and prompt loss.  

• 00:51:40 Reduce depth to 0.5 and pose to 0.6  for the second condition. These are  

• 00:51:46 separate influences, not percentages that must sum  

• 00:51:49 to 1. Their effects interact inside  this particular model and schedule.  

• 00:52:22 Compare the 2 combined results with  the single control examples. Check  

• 00:52:27 whether the background arrangement  and raised arm are both retained,  

• 00:52:31 then inspect clothing, expression and texture  for unwanted restrictions or artifacts.  

• 00:52:42 If 1 control dominates, adjust the one responsible  for that feature. Keep the other unchanged while  

• 00:52:49 testing. Changing both strengths, both schedules  and the prompt together would obscure which  

• 00:52:55 adjustment helped. Compatible maps can still  compete when their geometry disagrees. A pose  

• 00:53:01 moved sideways while the depth figure remains  centered asks for inconsistent arrangements.  

• 00:53:07 Fix the spatial disagreement before treating it as  a strength tuning problem. For overlapping people,  

• 00:53:14 a single crowded skeleton map can make limb  ownership ambiguous. Separate the subjects  

• 00:53:19 spatially or use separate generation and masks  when the application requires reliable control  

• 00:53:24 over who owns each arm and attribute. Save the  selected map pair and complete chained graph.  

• 00:53:31 A screenshot of the final image alone would  not preserve which map fed which node, their  

• 00:53:37 ordering or the strength and schedule used for  each one. Our final brief adds a second subject:  

• 00:53:48 the male explorer on the left and a compact orange  inspection robot on the right. We will compare a  

• 00:53:54 single spatial prompt with separately generated,  editable elements. Use the current Qwen-Image 2.1  

• 00:54:01 graph for this composition example. We are leaving  the matched InstantX experiment, so remove its  

• 00:54:07 control path rather than silently attaching that  file to a different base model. Describe the 2  

• 00:54:13 subjects with distinct positions and attributes.  The explorer has a blue jacket; the robot has an  

• 00:54:19 orange shell. State the shared workshop setting  and keep their roles clear within the prompt.  

• 00:54:40 Generate the complete composition and  inspect left versus right placement.  

• 00:54:45 Then check whether colors or objects have  crossed between subjects. A model can  

• 00:54:50 produce an attractive image while missing  the intended assignment of an attribute.  

• 00:55:06 Spatial language is useful guidance, but  it is not a set of independent editable  

• 00:55:12 layers. If the robot is almost correct but too  close to the explorer, changing the prompt can  

• 00:55:17 also change other parts of the image. For  a loose concept, that single generation  

• 00:55:23 may already be enough. For an illustration  with exact placement or repeated revisions,  

• 00:55:28 we need to consider how easily the result can  be changed after generation. Keep this result as  

• 00:55:34 the first solution to the brief. We will use the  same subject descriptions in the separate route,  

• 00:55:40 so the comparison remains about placement  and editability rather than switching to an  

• 00:55:45 unrelated scene. Generate the explorer and  robot separately, keeping their intended  

• 00:55:55 viewpoint and lighting compatible with the  workshop. Independent generation gives each  

• 00:56:00 subject more room, but it also makes us  responsible for their final relationship.  

• 00:56:57 Inspect each source before extracting it. A broken  

• 00:57:00 hand or an unclear robot shape will  remain a defect after compositing.  

• 00:57:05 Choose usable subjects first instead of hiding  weak generation inside a more elaborate graph.  

• 00:57:16 Use the selection and mask workflow from Lecture  5 to isolate each subject. Inspect the mask edge  

• 00:57:22 around the explorer's hands and the robot's small  parts, where rough selection is easiest to notice.  

• 00:58:14 Place the explorer on the left and the robot on  the right over the prepared workshop background.  

• 00:58:20 Move the actual image and its mask together so the  visible subject stays aligned with its selection.  

• 00:59:13 Adjust scale against the floor  and nearby furniture. Correct  

• 00:59:16 left and right positions are not  sufficient if the robot floats  

• 00:59:20 or the explorer's size conflicts  with the perspective of the room.  

• 00:59:37 Inspect the feet, contact points and edge  transitions at full size. Matching light  

• 00:59:59 direction and adding a believable contact  treatment can matter more than a perfectly  

• 01:00:04 sharp outline. Preserve the separate  elements while making those corrections.  

• 01:01:14 The visible rectangle is a compositing  defect, so we need to repair the contact  

• 01:01:19 layer. Give the soft mask room to fade into 0  before it reaches the edge of its own canvas.  

• 01:01:26 Create a solid mask with value 0. Its width  is 640 and its height is 128. This becomes  

• 01:01:36 the empty area around the existing contact  mask, rather than another visible patch.  

• 01:01:53 Combine the masks by placing the  existing contact shape inside that  

• 01:01:57 larger empty mask. The horizontal offset is  160 pixels and the vertical offset is 32,  

• 01:02:04 leaving padding around the soft shape.  

• 01:02:13 Use the add operation with the empty mask as  the destination and the contact shape as the  

• 01:02:19 source. Feed the combined mask into the  existing processing route so the padded  

• 01:02:24 boundary follows the shadow. The black image  and its mask must use matching dimensions.  

• 01:02:39 Keep each pair aligned when resizing  it; otherwise the image boundary can  

• 01:02:43 appear separately from the soft mask  and reveal the rectangular patch again.  

• 01:03:12 Reposition each contact patch beneath its  subject after changing its dimensions. The  

• 01:03:18 explorer and robot have different footprints,  so give them separate widths and positions,  

• 01:03:23 then run the composite and  inspect the floor again.  

• 01:03:56 Now move the robot slightly while  leaving the explorer fixed. This  

• 01:04:01 is the practical advantage of the separate route:  

• 01:04:04 the composition can change without asking the  image model to redraw every subject again.  

• 01:04:41 The rectangular boundary is gone, but the explorer  still needs stronger contact under the left boot.  

• 01:04:47 Increase both the contact-mask height and  its matching black-image height together,  

• 01:04:52 keeping their dimensions  synchronized at 120 pixels.  

• 01:05:02 Move that contact layer upward to support the  boot, then inspect the completed composition.  

• 01:05:08 This improves grounding while preserving separate  editable subjects. Matching their lighting to the  

• 01:05:14 background would still require more work  for a seamless composite. Place the single  

• 01:05:26 prompt result beside the composite. Compare the  stated positions, attribute separation, natural  

• 01:05:32 integration and the work required to revise each.  Neither route wins every brief; their trade offs  

• 01:05:38 are different. Use the simpler route when it meets  the task. Choose separate elements when placement,  

• 01:05:50 attribution or later editing requires that  control, and budget time for masks, perspective,  

• 01:05:56 light and contact so the assembled image feels  coherent. Export the finished composition,  

• 01:06:10 both subject sources and their masks. Keep the  editable workflow as well. A flattened final image  

• 01:06:16 is useful for delivery, while separate components  preserve the choices needed for the next revision.  

• 01:07:11 Reopen the saved structural workflow. Check that  the diffusion model and control weights belong to  

• 01:07:17 the same family, then confirm the depth and pose  images and their 2 strengths before inspecting  

• 01:07:23 the saved output. Open the saved image beside that  workflow. This is the result produced with depth  

• 01:07:32 at 0.5 and pose at 0.6. Keep it with the graph  to identify the demonstrated settings. Reopen the  

• 01:07:51 editable composition as well. The robot still has  its own horizontal position and aligned mask. Open  

• 01:07:57 the exported composition to confirm that the saved  files match the final placement of both subjects.  

• 01:08:19 The learner materials include the source image,  saved maps, comparison settings and editable  

• 01:08:25 graphs. Use their dependency links to match the  models, and keep your own results labeled when  

• 01:08:31 you change 1 variable. Choose the constraint  by the job: edges for useful boundaries, depth  

• 01:08:38 for spatial layers, pose for joints, and separate  masks when composition needs independent elements.  

• 01:08:45 Inspect the map before asking the generator to  obey it more strongly. You also know how to move  

• 01:08:51 a compatible workflow into Colab and preserve its  files. Next we will focus on detail, restoration  

• 01:08:58 and delivery, building on the structure and  composition decisions we can now control.