Your own image workshop: how to generate AI images on a PC

What "generating images locally" means

A local generator is a program with a model on your computer. You describe a scene, press start, and the graphics card creates the image. Most free local models do not have a daily image counter: you spend your PC's time and electricity.

To start, you need three things: a computer, a launch program, and model files. You can think of the program as a photo studio and the model as a tool with its own capabilities and result character. You can use the same program with different models.

After downloading all components, generation can work without the internet. However, a button inside the installed program sometimes calls a paid server. In this guide, such options are marked separately.

Why you need your own image workshop

Local AI is especially interesting when you need many trials, the ability to revert to settings, or working with your own materials. For example, you are making a game and want to compare ten versions of a single location, rather than ten completely different worlds.

The table can be scrolled sideways →

It is also a helpful tool for a 3D artist: you can look for scene imagery before modeling or test render mood options. A generated image itself does not provide a clean 3D mesh, materials, or a consistent character from all sides.

Can your computer handle it

The most important number is the dedicated video card memory, or VRAM. Computer operating memory is called RAM. SSD stores large model files. These three types of memory are not interchangeable: a large disk does not turn an 8 GB card into a 24 GB card.

In Windows, open Task Manager → Performance → GPU. Note the exact card name and the amount of dedicated memory. "Shared GPU memory" is borrowed RAM; it works differently.

The table can be scrolled sideways →

The table is a guide for selection, not our speed measurement. Numbers for SD, SDXL, and the FLUX FP8 path are tied to the documentation Invoke [18]. BFL describes FLUX.2 klein 4B with approximately 13 GB VRAM; ComfyUI indicates that Z-Image-Turbo fits within 16 GB. Different compressed versions and moving part of the work to RAM change the requirements. [5] [6]

For the newer FLUX 3 Klein, the manufacturer has an INT8 package for running on 8-gigabyte GPUs, specifically the RTX 3070. This is a separate optimized path: the generator, text encoder, and auxiliary model files do not need to reside in VRAM simultaneously. [8]

NVIDIA, Radeon, and Mac

NVIDIA: start with a ready-made build for your card and current driver. For a standard run, you don't need to install the entire CUDA developer kit right away.

AMD Radeon: the current ComfyUI has AMD support on Windows and Linux, while Portable has a separate AMD package. Check your exact card, OS, and driver in the official instructions. An NVIDIA instruction from a random video is not a universal instruction for Radeon. [2] [3]

Apple Silicon: you can use ComfyUI or Draw Things. Mac memory is shared between the system and graphics, so leave some headroom for other programs. Support for a specific model and format depends on the application. [2] [20]

Installing and getting your first image without code

In ComfyUI, the workflow is assembled from connected blocks. It might look complicated at first. To get your first result, you don't need to build the scheme yourself: take a ready-made template — a workflow.

1. Download Comfy Desktop from the official website. Select your OS.
2. During the first launch, select Local. This is an installation on your PC. Cloud runs computations on a server. Desktop prepares Python and the necessary environment itself. Check the detected video card and installation folder. [1]
3. Open Templates / Template Library. Find the FLUX.2 Klein 4B template for text-to-image generation. On a suitable PC, you can use Z-Image-Turbo instead. [4] [6]
4. Download missing models from the template prompts. All its components are required. If automatic download is unavailable, use the file table below and the official links.
5. Paste the scene description. For the first attempt, leave the template settings and one result per run — batch 1. Press Run or Queue depending on the interface version.
6. View and save the result. In the standard scheme, SaveImage writes the file to ComfyUI/output/. Then change one detail in the prompt and compare.

A small paper fox standing on a wet tram stop bench at dusk. Warm amber window light, deep navy city background, gentle rain, stylized hand-painted illustration. Medium shot, clear silhouette, one character, no text, no logos.

English example — a template for comparison, not a requirement to communicate only in English. You can try Ukrainian. If the model doesn't understand the description well, translate it and repeat with the same settings. It's better to add precise Ukrainian text for a banner in a graphic editor.

To check autonomy simply: after downloading all files, disconnect the network and run a new generation. If the workflow contains a cloud node, it will not work without a network.

Which local models are worth knowing

Choose based on the task: creating an image from scratch, editing a specific area, preserving a product, or placing layout elements. One model might be good at creating atmosphere but worse at following complex composition.

A number followed by the letter B usually means billions of model parameters. A larger number often means more required memory, but does not guarantee a better image. Besides the main model, the template may also load other components.

The table can be scrolled sideways →

What's special about the new FLUX 3

Distilled is an accelerated version; the official standard mode has 16 steps. Base has a different mode. The word 'steps' refers to the number of image refinement stages, not guaranteed quality. It's better to start with your specific template settings. [8]

The template may use another model that expands a short prompt into a full scene description — a prompt upsampler. It also consumes memory. If you turn it off, prepare a detailed description yourself: who is in the frame, what they are doing, where they are standing, the lighting, and the composition. New templates require the appropriate version of ComfyUI; FLUX 3 Klein — from 0.41 with a fix in 0.41.1, MOSS — from 0.42. [10] [15]

A Model is not a Program: where else is it convenient to work

The table can be scrolled sideways →

You don't need to install everything. If you only need one image for a post, one working template is enough. If you frequently edit specific parts of an image, a drawing interface might be more convenient than a large block diagram.

What is free, what is paid, and where the limits are

Check these three things separately: the program price, the computation location, and the model license. A free program might connect to a paid API, and a freely downloaded model might prohibit commercial use.

The table can be scrolled sideways →

Whether the result can be used for work

For an easier choice of a commercial route, look at the official FLUX.2 klein 4B, Z-Image-Turbo, MOSS-Image, or Qwen-Image-2512 with Apache 2.0. This does not exempt you from checking rights to source photos, logos, or specific LoRAs.

Qwen-Image-2.1 and Turbo have different terms: The Research License allows non-commercial research and evaluation; a separate license is required for commercial use. The word "newer" here does not mean "freer". [13]

FLUX 3 Kommunity allows commercial use of the output images for users who meet the definition of a Qualifying User, specifically a threshold of aggregate annual revenue from affiliates of less than $5 million, and other conditions. For all other scenarios, check separate licensing. This is not an unconditional Apache license. [9]

"Unrestricted" is better read as "no cloud quota for this local run." Physical PC limits, model conditions, and output quality remain. And an "uncensored" tag does not guarantee more accurate anatomy or better editing.

Mini-project: one illustration in four meaningful attempts

Instead of the prompt "make something beautiful," set a specific goal: an illustration for a post about an evening in the city. Let's take a fictional paper fox at a tram stop — that same frame can later be brought to life in a guide on local video.

1. Composition. Make several variations of one description. Select the one where the character and action are clear.
2. Light. Keep the composition as a reference, change only the lighting. Check if the fox itself has changed.
3. One edit. Mask an unsuccessful object and regenerate the area if your template supports it.
4. Final. Check the fine details, prepare the required size, and add precise text in an editor.

Use the attached image as the reference. Keep the same adult character, facial features, hairstyle, body proportions, outfit and composition. Change only the background weather from rain to clear evening. Preserve the existing warm lighting on the character. Do not add people, text or objects.

The "save all" command only sets an intention. If faces, clothing, or product shape change after editing, the version should be rejected or corrected. A reference helps with consistency but does not provide absolute identity.

For pros: how to get more control

Mask, reference, ControlNet, and LoRA — different tools

• Mask shows which area needs to be changed. Inpainting — drawing within the frame, outpainting — expanding beyond its borders.
• Reference — a sample of a character, object, or style. The model must support the corresponding way of working with it.
• ControlNet and similar tools help set contours, depth, or pose. An add-on must match the model family; an SDXL file does not automatically become an add-on for FLUX.
• LoRA — a small additional model setting for a style or object. It does not replace the base model and requires compatibility and its own verification of terms.

For example, for a character, the process "reference → composition control → local edit → manual selection" is useful. For product advertising, it is better to keep a verified product layer and work with the background than to trust the generator with every small label on the packaging every time.

What to save along with the result

Record the exact model file, software version, workflow, prompt, size, number of steps, and seed — the initial number for generation randomness. The same seed helps compare edits in the same environment, but does not guarantee identical pixels after changing the model, version, or hardware.

Quantization — a way to reduce the model weight volume. FP8, INT8, NF4, and GGUF are not interchangeable terms for "small file": specific format and hardware support is required. For example, the NVFP4 version of FLUX 3 requires Blackwell. [8]

Purpose of the image: Where it will be shown and what format: Main object: What it does: Background and lighting: What to keep from the reference: What exactly to change: Which details to check manually: Exact model, file, and license:

What to do if it doesn't work

The table can be scrolled sideways →

Start with one task and one model. Once a stable result appears, it will become clear what is lacking: memory, better composition, or a more precise editing tool. Regarding general local AI work, there is a separate simple guide, and regarding motion and sound — an article based on a local video.

Sources and preparation method

This is an original editorial guide based on official documentation, model cards, and licenses. Selection advice serves as our guidelines, not a ranking based on our own lab tests. We have not compared all models on a single PC. Before downloading, verify the template version, and before paying, check the current tariff.