What is AI image generation? Understand the mechanism and main services

What is AI image generation? Understand the mechanism and main services | Sugiyama Nobutsugu

Why AI image generation stops the worksite

I think many people have already encountered AI image generation.、When I try to use it in practice, it stops at the same place.。

  • Even though the instructions are the same, the results are not stable.
  • The output is completely different depending on the service.
  • I don't know which one to choose

This is not a skill issue。

The cause is、Not understanding the mechanism and service structure separatelyです。

AI image generation is not a “tool”、It's a system that goes into the production process.。
I need to sort this out first.、No matter what you use, reproducibility will not improve.。

Why the results vary (understanding the mechanism)

Structure for generating images from noise

Many of the current image generation AIs、It works on a mechanism called a diffusion model.。

this is、

  • Start with random noise
  • gradually converted into images

This is the process。

In other words、We are not creating a complete form from the beginning.、
Stochastically converges to a “likely state”Only。

Therefore, in practice、

  • Same instructions, different results
  • cannot be completely reproduced
  • The task is to “bring it closer”

The premise is that。

Prompts are not “instructions” but “weighted”

Text input (prompt) is not an instruction。

  • Strongly written elements are more likely to be reflected
  • Weak elements may be ignored

In other words、this is
Rather than specifying conditions、trend controlです。

If you misunderstand this、

  • It is better to write long
  • The more detail you write, the more accurate it will be.

I think so.、Actually it's the opposite、
Designing what to prioritizebecomes important。

Image input is a “control device”

Because text alone is unstable、Use images in practice。

If you insert an image、

  • The composition is stable
  • The colors match
  • details are fixed

In other words、

  • Text = Direction
  • Image = Control

The role will be。

I wonder if it is possible to separate these two、In practice it makes a big difference。

Difference between image generation cloud type AI and local AI

This is the first branch in understanding AI image generation.。

However, rather than "which is better"、
Differences in how much control is requiredshould be understood as。

Cloud-based AI：Mechanism to output the completed image

Representative things：

  • Midjourney
  • DALL-E
  • Adobe Firefly
  • Gemini
  • ChatGPT
  • Grok et al.

Features：

  • generated on the server side
  • High initial quality
  • Get results right away

Behavior in practice：

  • Works even with vague instructions
  • the atmosphere is strong
  • However, detailed control is difficult

In terms of shooting、
Shooting in an already completed studioです。

local AI：Mechanism to control the production process

represent：

  • Stable Diffusion

Features：

  • Works on PC
  • Configurable/customizable
  • Can create reproducibility

Behavior in practice：

  • conditions can be fixed
  • Can reproduce the same composition
  • Strong in mass production

In terms of shooting、
Assembling your own lighting and equipmentです。

There are few local types of image generation AI。

Cloud and local have different roles

These two are not in competition。

In practice, it is divided as follows。

  • Cloud → rough/direction/initial generation
  • Local → control/reproduction/mass production

Without this understanding、

  • Failed when trying to mass produce in the cloud
  • It is inefficient to create rough information locally.

A discrepancy occurs.。

Differences in major services (practical perspective)

It's not about "performance" here.、Differences in design philosophyI'll see it at。

Midjourney：create a direction

  • strong atmosphere
  • Art-oriented
  • Strong against rough generation

use：

  • Key visual examination
  • tone design

DALL-E：Verify instructions

  • Easy to understand text
  • Stable composition
  • Fewer bankruptcies

use：

  • Confirm instructions
  • Composition arrangement

Adobe Firefly：Incorporate into production

  • Design tool collaboration
  • Strong partial generation

use：

  • Retouching aid
  • Replacement work

Stable Diffusion：control and mass production

  • Customizable
  • reproducible

use：

  • Product image mass production
  • Fixed composition generation

Overall picture of AI image generation (production flow)

AI image generation is not a standalone、Differently used in the process。

① Rough/directional design

→ Midjourney

② Instructions/composition verification

→ FROM-E

③ Connection to actual production

→ Firefly

④ Mass production/operation

→ Stable Diffusion

like this、
Roles are divided in the production processis the reality。

Common failure patterns

These three are the most common in practice.。

① Try to do everything with one service

→ There will always be a limit

② Try to solve the problem using prompts

→ Control is achieved through structure.

③ Immediately use it for production production

→ The verification process is skipped.

In terms of shooting、

  • Production without testing
  • All supported by fixed equipment

is in the same state as。

Connecting human production and AI

I'll sort it out at the end。

AI is in charge

  • rough generation
  • Composition verification
  • Variation development

Responsible for people

  • concept design
  • brand judgment
  • final quality

Only after this separation is achieved、
AI will be incorporated into production。

summary：Understanding AI image generation in terms of structure

There are three points to understand about AI image generation.。

  • How it works (why it breaks)
  • Service (why different)
  • Process (where to use it)

If you press this、

  • Don't worry about choosing tools
  • Improves reproducibility
  • Can be incorporated into production

It will look like this。

AI image generation is not a technology、
Design elements of the production processです。

If you can understand this far、
It will be ready for practical use for the first time.。

▶︎ [Required environment for AI image generation | Difference between cloud AI and local AI]

▶︎ [AI image generation depends on PC performance | Differences between Mac and Windows environments]

▶︎ [Is GPU necessary for AI image generation? Difference from CPU and role]