How to get exactly the image you have in mind — instead of something pretty but different.
Concepts 10 min Intermediate May 25, 2026
In previous articles, you learned to generate images, use Stable Diffusion, and fine-tune with LoRA. But you have probably noticed: sometimes the pose is wrong, the resolution too low, or you cannot get that specific composition.
This article gives you the professional toolkit: ControlNet for precise spatial control, AI upscaling for print-ready resolution, and the 7-step workflow that turns random results into deliberate creations.
ControlNet — Steering Composition
ControlNet
AnalogyDefinition
A text prompt is like telling an architect: "Build me a modern house." ControlNet adds the floor plan — the architect now knows not just WHAT to build but WHERE each room goes.
Analogy:
A text prompt is like telling an architect: "Build me a modern house." ControlNet adds the floor plan — the architect now knows not just WHAT to build but WHERE each room goes.
Definition:
ControlNet adds a second input channel to image generation: alongside the text prompt, a control image provides structural constraints (pose, edges, depth, or sketch). The AI generates an image matching both the text description and the spatial structure.
The core insight: The text prompt controls WHAT appears in the image. ControlNet controls WHERE it goes.
OpenPose Detects body poses and extracts a skeleton from photos.
Canny Edge Detects object edges and outlines for structural control.
Depth Captures spatial depth for three-dimensional composition.
Scribble Uses hand-drawn sketches as composition templates.
Example: OpenPose workflow. (1) Take a photo of yourself standing with arms crossed. (2) Run it through the OpenPose preprocessor to extract a skeleton.
(3) Combine the skeleton with the prompt "a medieval knight in full armor, castle background." (4) Result: a knight in YOUR exact pose. The pose is preserved, everything else is new.
Preprocessors extract control images automatically from photos. You need no technical knowledge — upload a reference photo, choose the mode, and the result appears immediately.
Misconception: ControlNet is too complicated for beginners
The basic modes (OpenPose, Canny) require no technical knowledge. You upload a reference photo, the preprocessor does the extraction automatically, and the result is immediate. Advanced configuration is optional.
Deep Dive: How ControlNet works under the hood
ControlNet uses a dual-encoder architecture: the text encoder processes the prompt, a separate control encoder processes the control image. Both signals feed into the diffusion process. The control strength can be adjusted via a weight — at low weight, the AI interprets the structure more freely; at high weight, it follows more precisely.
Upscaling — From Small to Print-Ready
Why not simply generate at high resolution directly? Because it causes severe artifacts: double heads, merged limbs, distorted bodies. The solution: generate small, scale big.
Simple Resize
Pixels are stretched | Blurry result | No new details | Instant
AI Upscaling
New details are generated | Sharp result | Plausible textures | Seconds
Simple resize is like zooming into a newspaper photo — you see bigger dots but no new information. AI upscaling is like a forensic artist who takes a blurry security camera image and paints in realistic skin texture, fabric patterns, and hair strands.
E
ESRGAN
Fast and deterministic. Ideal for photos and anime. Adds no creative details but processes images in seconds. 4x enlargement in ~2 seconds.
S
SD Upscale
Lightly re-noises the image and de-noises it again. Adds genuine new details (skin texture, fabric patterns). Slower but more creative. Ideal for print products.
T
Tiled Upscaling
Processes the image in tiles. Works with as little as 4 GB VRAM. Ideal for large images on limited hardware.
The core principle: Generate at your model's native base resolution (512×512 for SD 1.5, 1024×1024 for SDXL), then upscale. Never the other way around.
Concrete example: (1) Generate portrait at 512×512. (2) Apply ESRGAN 4× → 2048×2048 in 2 seconds, instant print quality. Alternative: (1) Generate at 512×512. (2) SD Upscale 2× with denoising 0.3 → 1024×1024 with new details. (3) ESRGAN 2× → 2048×2048 final.
Misconception: Upscaling is just as good as native high-resolution
Native high-resolution has more genuine detail. But direct high-res generation causes severe composition artifacts. Upscaling is the practical compromise: good composition at low resolution, then good detail through scaling.
Deep Dive: ESRGAN vs. SD Upscale — When to use which?
ESRGAN is deterministic: same input always produces the same output. It sharpens and smooths intelligently without adding creative details. Ideal for fast results and consistent quality. SD Upscale lightly re-noises the image and lets the AI de-noise it again — creating genuine new details like skin texture or fabric folds. Ideal for print products where maximum detail matters. Downside: slower and may slightly alter the image.
The Professional Workflow — Putting It All Together
Professional image generation follows a 7-step iterative pipeline. It is not about typing one perfect prompt, but about iterative refinement:
1
Formulate prompt
2
Batch generate (4–8 variants at 512×512)
3
Select best composition
4
ControlNet for structural refinement
5
Inpainting for detail fixes (as learned in the previous article)
6
Upscale to target resolution
7
Post-processing (Photoshop/GIMP)
Why batch generation? Because you create 8 variants at 512×512 in 30 seconds. The chance of finding a good composition increases dramatically. Then you only refine the best variant further.
Tools like ComfyUI enable this workflow as a visual, node-based pipeline — each step is a building block you can connect, save, and reuse.
Example: Book cover workflow. (1) Prompt: "fantasy landscape, ancient tower, dramatic sky." (2) Generate 8 variants at 512×512 in 30 seconds. (3) Select best composition, note the seed.
(4) Apply ControlNet Depth for spatial arrangement. (5) Inpaint the title area to keep it clear. (6) ESRGAN 4× to 2048×2048. (7) Export as 300 DPI CMYK for print.
Misconception: AI replaces graphic designers
AI is a tool FOR designers. The creative vision, composition decisions, refinement choices, and final polish remain human tasks. The 7-step workflow shows: AI generates raw material, the human shapes it into a finished product.
Interactive: The 7-Step Pipeline as a Node Diagram
articleWorkflow.enrichment.intro.title
articleWorkflow.enrichment.intro.description
Step 0 of 7
Start animation
Click "Play" to see the RAG pipeline step by step.
articleWorkflow.enrichment.edu.title
articleWorkflow.enrichment.edu.description
Key Takeaways
ControlNet solves the "what goes where" problem: text prompts control content, ControlNet controls composition (pose, edges, depth, sketch).
Always generate at your model's native base resolution (e.g. 512×512 for SD 1.5, 1024×1024 for SDXL) and upscale afterward — direct high-resolution generation causes severe artifacts.
Professional image generation is an iterative 7-step pipeline, not a single prompt. Generate many, select the best, refine, upscale.
With this workflow, you now have the complete toolkit for generative image art. You know the path from first prompt to print-ready result — the same principles of iterative refinement apply to video generation and other generative AI workflows.
Quiz: Workflow Optimization
Question 1 / 5
Not completed
You want to generate an image that preserves the exact spatial layout of an architecture photo but changes the style to "futuristic sci-fi city." Which ControlNet mode do you use?
1. You want to generate an image that preserves the exact spatial layout of an architecture photo but changes the style to "futuristic sci-fi city." Which ControlNet mode do you use?
☐ A) OpenPose — detects body poses
☐ B) Depth — captures the spatial depth and arrangement of the scene
☐ C) Scribble — uses hand-drawn sketches as templates
☐ D) Canny Edge — only detects object edges
2. Why do professionals generate images at 512×512 and upscale afterward, instead of generating directly at 2048×2048?
☐ A) Because 512×512 images load faster on websites
☐ B) Because 2048×2048 images cannot be saved as PNG files
☐ C) Because direct high-resolution generation causes composition artifacts (double heads, merged limbs), while low-resolution + upscaling avoids this
☐ D) Because upscaling is mandatory for all AI-generated images
3. You have a photo of a person sitting cross-legged and want to generate a fantasy character in the exact same body position. Which ControlNet mode and workflow do you use?
☐ A) Canny Edge — extract outlines, prompt "fantasy character"
☐ B) OpenPose — extract skeleton from photo, combine with prompt "fantasy character, sitting cross-legged"
☐ C) Depth — extract depth map, prompt "fantasy character"
☐ D) Scribble — hand-draw the pose, prompt "fantasy character"
4. You generated a portrait at 512×512 and want to print it as a photorealistic book cover at 2048×2048. You have two options: ESRGAN 4× (fast, deterministic) or SD Upscale 2× + ESRGAN 2× (slower, adds details). Which do you choose for the book cover?
☐ A) ESRGAN 4× — speed is always more important than quality
☐ B) ESRGAN 4× — because SD Upscale always produces worse results
☐ C) Neither — just generate directly at 2048×2048
☐ D) SD Upscale + ESRGAN — the SD Upscale step adds realistic skin texture and fabric details that pure ESRGAN cannot create
5. A friend says: "I just type one prompt, take the first result, and resize it in Paint. That's my workflow." Evaluate this approach against the 7-step professional pipeline. What are the two biggest problems?
☐ A) No batch generation means missing potentially better compositions, and simple resize (instead of AI upscaling) produces blurry enlargements without new details
☐ B) Paint is slower than Photoshop, and the prompt should be in English
☐ C) The prompt is probably too short, and the friend should use more negative prompts
☐ D) The only problem is that Paint cannot save in CMYK format
Answer Key: 1) B · 2) C · 3) B · 4) D · 5) A
Checkpoint
Explain the difference between a text prompt and ControlNet: what does each control? Name a specific ControlNet mode and its use case.
Why does generating directly at 2048×2048 often produce errors — and how does the detour via 512×512 + upscaling solve this problem?
In the 7-step workflow, batch generation comes before ControlNet. Why does this order make sense?