Three interventions that turn an AI from slot-machine mode into a real creative tool.
Concepts 10 min Intermediate May 25, 2026
Your base model can generate stunning images — but it gives you limited control. The style isn't quite right? LoRA fixes that. The hands look wrong? Inpainting fixes that. The framing is too tight? Outpainting fixes that.
These three techniques are the difference between "let's see what comes out" and directed creative work. By the end of this article, you will understand how each technique works, when to use it, and where its limits lie.
LoRA — Adding Styles and Concepts
LoRA — Low-Rank Adaptation
AnalogyDefinition
Think of a base model as a versatile painter who can paint in many styles. A LoRA is like a style guide sheet placed on the painter's desk — it nudges the painter toward a specific style (say, Studio Ghibli anime) without retraining the painter from scratch. You can place multiple style guides (combine LoRAs) and control how closely the painter follows each guide (strength slider). But a style guide written for an oil painter (SD 1.5) makes no sense to a watercolor artist (SDXL) — the guide is specific to the painter type.
Analogy:
Think of a base model as a versatile painter who can paint in many styles. A LoRA is like a style guide sheet placed on the painter's desk — it nudges the painter toward a specific style (say, Studio Ghibli anime) without retraining the painter from scratch. You can place multiple style guides (combine LoRAs) and control how closely the painter follows each guide (strength slider). But a style guide written for an oil painter (SD 1.5) makes no sense to a watercolor artist (SDXL) — the guide is specific to the painter type.
Definition:
LoRA (Low-Rank Adaptation) is a method by Hu et al. (2021) that specializes a pre-trained model toward a particular style, face, or concept without retraining the entire model. A LoRA file is tiny (10-200 MB) compared to a full checkpoint (2-7 GB) because it only adjusts a limited subset of neural network weights via low-rank matrix decomposition. Multiple LoRAs can be combined, and a strength slider (0.0-1.0) controls the dominance of the effect.
The style guide analogy has limits: a real painter interprets style guides subjectively. LoRA applies mathematical weight modifications mechanically — there is no interpretation, only computation.
10-200 MB
LoRA file size For comparison: A full checkpoint weighs 2-7 GB.
LoRA
Small add-on file (10-200 MB). Modular system — multiple LoRAs can be combined. Specialized for a specific style or concept. Strength slider from 0.0 to 1.0.
Checkpoint
Full model (2-7 GB). Monolithic — one checkpoint per generation. Broad foundation with many styles. No continuous strength slider.
The LoRA strength determines how dominant the effect is in the final image. Here is an example with a Studio Ghibli LoRA and the prompt "a forest with a river":
1
Strength 0.0 — No effect, pure base model
2
Strength 0.3 — Subtle Ghibli touch in colors and linework
3
Strength 0.7 — Strong Ghibli style, some realism remains
4
Strength 1.0 — Full Ghibli anime rendering
Same prompt, same model — dramatically different results, controlled entirely by the LoRA strength slider.
Common Misconception: LoRA Compatibility
"LoRAs work with any model" — this is false. LoRAs are trained for a specific model architecture. An SD 1.5 LoRA does not work with an SDXL model. Always check which model architecture the LoRA was trained for before downloading.
LoRA represents a broader principle of efficient finetuning: instead of retraining billions of parameters, only small weight subsets are modified. This principle extends beyond image generation — LoRA is also one of the most important finetuning methods for large language models (LLMs).
Deep Dive: LoRA Training
Training your own LoRAs is more accessible than you might think: with just 10-50 well-chosen example images, you can teach a model a specific art style or a particular object. The rank parameter (typically 4-128) determines the capacity: low values for simple styles, high values for complex concepts.
The most important platforms for pre-made LoRAs are Civitai (the largest community) and Hugging Face. Both offer thousands of free LoRAs for download — from art styles to characters to specific objects.
Interactive: Why LoRA Is So Efficient
LoRA trains only a fraction of the parameters. But how dramatic is the difference? Move the slider and observe: a frozen model has zero trainable parameters (green). LoRA grows linearly with model dimension (blue). Full fine-tuning grows quadratically (red). Beyond a model dimension of d=1000, the difference becomes enormous — this is precisely why LoRA matters so much for modern models with millions of parameters.
1010000
Frozen (0 params)1
LoRA (2·d·r)100
Full Fine-Tuning (d²)10.000
Moderate Input
At n=100, the difference becomes visible: O(n²) requires 10.000 operations, while O(n) needs only 100. O(log n) needs just 6.6 — that's 15x less than O(n).
Ratio to O(n)
Complexity
Operations
Factor vs. O(n)
Frozen (0 params)
1
100x faster
LoRA (2·d·r)
100
1x (Reference)
Full Fine-Tuning (d²)
10.000
100x slower
Inpainting — Selectively Regenerating Parts of an Image
Inpainting — Mask-Based Selective Regeneration
AnalogyDefinition
Inpainting is like a restorer working on a damaged painting. The restorer covers the damaged area with masking tape (the mask), studies the surrounding undamaged paint for color, texture, and light direction, then paints new content into only the masked area — blending it so seamlessly that nobody can tell where the original ends and the restoration begins. A light touch (low denoising) makes subtle corrections; a heavy hand (high denoising) paints something entirely new.
Analogy:
Inpainting is like a restorer working on a damaged painting. The restorer covers the damaged area with masking tape (the mask), studies the surrounding undamaged paint for color, texture, and light direction, then paints new content into only the masked area — blending it so seamlessly that nobody can tell where the original ends and the restoration begins. A light touch (low denoising) makes subtle corrections; a heavy hand (high denoising) paints something entirely new.
Definition:
Inpainting is the process of selectively regenerating a specific masked region of an existing image while keeping the rest untouched. You paint a mask over the area you want to change, write a new prompt targeting only that area, and the model generates new content that blends seamlessly with the surrounding pixels — matching lighting, perspective, and style. Denoising strength (0.0-1.0) controls how much the model deviates from the original.
The restorer analogy has limits: a real restorer aims to recreate the original. Inpainting generates something new guided by a prompt — it does not restore what was there before, it creates what you ask for.
1
Generate an image
2
Spot the defect
3
Paint a mask over the area
4
Enter a new prompt for the area
5
Regenerate only the masked area
Typical workflow: You generate a portrait — the face is perfect, but one hand has six fingers. You paint a mask over the deformed hand, set the prompt to "realistic human hand, five fingers, detailed" and the denoising strength to 0.6. The model produces anatomically correct fingers that match the skin tone, lighting, and pose of the rest of the image.
Swap objects Replace a dog with a cat or swap out an element
Change background Replace the entire background with a new scene
Add details Add a painting to an empty wall or insert new objects
Common Misconception: Inpainting Can Fix Everything
Inpainting fixes details, not fundamental composition problems. If the overall composition is fundamentally flawed, the perspective is completely wrong, or the pose does not work, inpainting will not help — in those cases, it is better to regenerate the entire image from scratch.
Specialized inpainting models (such as SD 1.5 Inpainting and SDXL Inpainting) are specifically trained to understand mask boundaries and produce seamless transitions. Another example of how targeted training creates specialized tools.
Deep Dive: Denoising Strength
Denoising strength is the most important parameter for inpainting. It controls on a scale from 0.0 to 1.0 how much the model changes the masked area. At 0.0, everything stays as the original — at 1.0, the model regenerates the area completely without regard for the original image.
For subtle corrections (such as color adjustments), low values around 0.3 work well. For fixing hands or faces, the optimal range is 0.5-0.7 — enough freedom for new structures, but still enough context from the original. At 0.8+, the model largely ignores the original and generates almost freely.
Outpainting — Expanding Beyond Image Borders
Outpainting — Extending the Canvas
AnalogyDefinition
Outpainting is like unrolling more canvas around a finished painting. The painter studies the existing composition — where the horizon line goes, how the light falls, what colors dominate — and then continues painting outward so that the extension looks like it was always part of the original.
Analogy:
Outpainting is like unrolling more canvas around a finished painting. The painter studies the existing composition — where the horizon line goes, how the light falls, what colors dominate — and then continues painting outward so that the extension looks like it was always part of the original.
Definition:
Outpainting extends an image beyond its original borders by generating new content in the empty canvas area. Technically, it is inpainting applied to blank space: the model treats the empty area as the "masked region" and uses the existing image as context. You choose a direction (left, right, up, down) and extension size. The model must maintain perspective, lighting consistency, and style coherence across the boundary.
The canvas analogy has limits: a human painter has perfect understanding of the 3D scene. The AI model only has the pixel context and may produce increasingly inconsistent results the further it extends from the original — style and lighting can "drift."
Inpainting
Masks existing content to replace or correct it. Works within the image boundaries. Ideal for repairs and modifications to the existing image.
Outpainting
Masks empty space to extend the image. Works beyond the image boundaries. Ideal for format changes and image extensions.
Concrete example: You have generated a 512x512 portrait showing head and shoulders. By outpainting downward by 256 pixels, the model generates a torso and legs in matching style. Another 256 pixels down: a full-body image (512x1024) emerges. Alternatively: extend a 512x512 landscape left and right for a 1024x512 panorama as a website header.
Caution: Style Drift on Large Extensions
With very large extensions, the original art style and lighting can noticeably drift over distance. The new edges may no longer visually match the core of the image. Recommendation: prefer multiple small extension steps over one large one — this gives the model better context.
Outpainting impressively demonstrates how the same underlying technology (mask-based generation) serves completely different creative purposes when applied differently. Inpainting and outpainting are technically the same process with different use cases.
Key Takeaways
A LoRA is a small add-on file (10-200 MB) that steers a base model toward a specific style or concept — but it only works with the model architecture it was trained for (SD 1.5 LoRA does not work with SDXL).
Inpainting regenerates a masked area of an existing image while keeping everything else untouched — ideal for fixing hands, swapping objects, or changing backgrounds. Denoising strength controls the balance between preservation and change.
Outpainting extends an image beyond its borders by treating the empty canvas as a masked area — technically identical to inpainting but applied to new space. Large extensions may cause style drift.
Quiz: Control & Finetuning
Question 1 / 4
Not completed
You download a LoRA trained on Stable Diffusion 1.5 and try to use it with an SDXL model. What happens?
1. You download a LoRA trained on Stable Diffusion 1.5 and try to use it with an SDXL model. What happens?
☐ A) It works but at reduced quality.
☐ B) It works perfectly — LoRAs are universal.
☐ C) It does not work — LoRAs are model-architecture-specific.
☐ D) It only works if you set the LoRA strength to 1.0.
2. A generated portrait has a perfect face but deformed hands. You want to fix ONLY the hands. Which technique and settings should you use?
☐ A) Outpainting with the hands at the image border.
☐ B) Inpainting: mask the hands, prompt "realistic hands", denoising ~0.6.
☐ C) Generate the entire image again with a better prompt.
☐ D) Apply a hand-fixing LoRA at strength 1.0.
3. You want to turn a 512x512 portrait into a 1024x512 panorama by adding scenery to the left and right. Which technique is correct?
☐ A) Inpainting the left and right edges.
☐ B) Outpainting: extend the canvas left and right, the model fills in matching scenery.
☐ C) Generate a new 1024x512 image with the same prompt.
☐ D) Use a LoRA trained on panoramic images.
4. Outpainting is technically described as a variant of inpainting. Why is this characterization accurate, and what is the key difference in application?
☐ A) Both use masks — inpainting masks existing content to replace it, outpainting masks empty space to extend it.
☐ B) Both use LoRAs — inpainting uses small LoRAs, outpainting uses large ones.
☐ C) Both change the prompt — inpainting changes part of the prompt, outpainting changes all of it.
☐ D) They are completely different technologies that share a name.
Answer Key: 1) C · 2) B · 3) B · 4) A
Comprehension Check
Explain why a LoRA is much more storage-efficient than a full checkpoint — and what you need to watch out for when combining it with different models.
Describe the difference between inpainting and outpainting: when would you use each technique?
Why should you avoid inpainting on an image with fundamentally bad composition?