Once you can generate articles with a local LLM, the next thing you need is images.
A blog post needs more than just body text — it also needs featured images and images for sharing on social media.
You could easily generate these with an external image generation API, but if you create multiple candidates for each article, the fees add up.
So I use an M1 Max MacBook Pro and an M4 Pro Mac mini to generate images locally as well.
Is Image Generation on a Mac Possible?
Apple Silicon Macs can run image generation models such as Stable Diffusion and FLUX.
Compared to dedicated machines with NVIDIA GPUs, they can be at a disadvantage in terms of compatibility and generation speed.
On the other hand, Apple Silicon uses a unified memory architecture in which the CPU, GPU, and memory are integrated.
With 64GB of memory, it becomes easier to try out multiple image generation models.
My main use cases this time are the following:
- Featured images for blog articles
- Post images for X and Bluesky
- Illustrative images for introducing apps
- Background images for websites
- Creating variations of existing images
- Fixing parts of an image
- Upscaling images to higher resolutions
Draw Things and ComfyUI
As ways to get started with image generation on a Mac, I shortlist Draw Things and ComfyUI.
Draw Things
Draw Things is an app that makes it relatively easy to try image generation on a Mac.
Because you configure models and generation settings through a GUI, it's easy to pick up even if you're not used to working in the terminal.
It is well suited for the following use cases:
- You want to try image generation for the first time
- You want to generate images manually while tuning prompts
- You don't need complex automation yet
- You want to try multiple models with minimal setup
It works well as a first environment for checking the quality and speed of local image generation.
ComfyUI
ComfyUI is a tool for combining image generation steps as nodes.
You can visually build workflows covering model loading, prompt input, image generation, upscaling, saving, and more.
It is well suited for the following use cases:
- You want to automate image generation
- You want to run the same process repeatedly
- You want to invoke workflows from an API
- You want to combine multiple steps
- You want to automatically generate featured images for each article
Since the ultimate goal this time is to receive image generation jobs from a VPS, I plan to use ComfyUI in production.
That said, the initial generation tests can also be done in Draw Things.
Which Model to Use
Each image generation model has its own characteristics.
Stable Diffusion 1.5 family
It is relatively lightweight, with an abundance of information and add-on models.
It's a good fit when starting at lower resolutions or when trying things across a wide range of environments, including older Macs.
SDXL family
Larger than the Stable Diffusion 1.5 family, this model makes it easier to generate high-resolution images.
For blog featured images and website graphics, it's a practical candidate.
FLUX family
You can expect high-quality images and relatively strong adherence to prompts.
On the other hand, the models are large, and the required memory and generation time can increase.
Where lightweight or quantized versions are available, I'll compare how practical they are on a Mac.
When choosing a model, always check the commercial use terms, not just quality and speed.
Workflow for Generating Article Images
When generating a featured image from an article, the following flow comes to mind:
- Create the article body with a local LLM
- Summarize the article content briefly
- Create an image generation prompt
- Pass the prompt to ComfyUI
- Generate multiple candidate images
- Upscale if needed
- Return the images to the VPS
- Register them as media in Payload CMS
The important point is not to pass the article body directly to the image model.
Distill the article's subject, mood, composition, colors, intended use, and so on, and convert that into a prompt for the image.
Image Sizes and Generation Load
In image generation, memory usage and generation time increase as the resolution goes up.
In initial testing, don't jump straight to large images — work up through the following steps:
- Check the composition at low resolution
- If it looks good, generate at standard resolution
- Upscale only the final image
- Compress and convert for web publishing
For blog images, there's no need to make the source image larger than necessary.
Decide the final output size based on the display size and your website's image optimization setup.
Comparing the M1 Max and M4 Pro
As with the local LLM, I run the same image generation workflow on both Macs.
The items to compare are the following:
- Model loading time
- Time to generate one image
- Memory usage
- Swap usage
- Chassis temperature
- Fan behavior
- Stability during high-resolution generation
- Impact when other apps are in use
For image generation, the M1 Max's GPU configuration and memory bandwidth may work in its favor.
On the other hand, depending on the app and model implementations, the M4 Pro may be faster.
Based on the measured results, I'll also consider splitting article generation and image generation between the two machines.
Should the LLM and Image Model Run at the Same Time?
Even on a 64GB Mac, loading a large LLM and an image generation model into memory at the same time can leave little headroom.
At first, process them sequentially as follows:
- Generate the article and image prompt with the LLM
- Release the LLM from memory
- Load the image generation model
- Generate the images
- Release the image generation model
This approach saves memory compared to keeping both resident at the same time.
If the number of jobs and the waiting time grow, another option is to split article generation and image generation across the two Macs.
Precautions for Automatically Generated Images
When using generated images on a blog or social media, check the following:
- The license of the model used
- The terms for LoRAs and add-on models
- How trademarks and logos are handled
- Images resembling real people
- Excessive reliance on copyrighted works or the style of specific artists
- Unnatural text or details
- Consistency with the article content
- Whether anything could come across as misleading
In the early stages, even automatically generated images will be reviewed by a human before use.
Summary
Local image generation is possible on Apple Silicon Macs too.
Draw Things is a strong candidate for the manual experimentation stage, while ComfyUI is the better fit when automation is the premise.
I'll try the same models and workflows on both the M1 Max and M4 Pro, comparing generation speed, memory usage, and stability.
Next time, I'll build a mechanism that lets the Mac at home securely pick up generation jobs registered on the ConoHa VPS.