How much memory does it need?

On Windows, check the GPU name and dedicated GPU memory in Task Manager. On an Apple Silicon Mac, check the chip and unified memory in About This Mac. Then choose an installation path below. Model files and intermediate generation data both use memory, so start with one image.

Image and video generation

Local image generation: software, hardware and your first image

Create poster-background candidates with local image generation

Local AI image generation turns a written scene into an image on your computer. Connect an image model to ComfyUI to create thumbnails and poster backgrounds. Start with GPU¹ requirements, then work through setup and your first image.

To make a thumbnail or poster background, start with the computer you already own. Choose the app and model files for that hardware, save one image successfully, then increase the resolution or image count.

Contents

Requirements and key details

The app, model, workflow, and weight files are distinct components.
Qwen-Image 2.1 does not use the same file configuration as Qwen-Image-2512; it needs a dedicated VAE and Qwen3-VL encoder.
Image count and resolution affect peak memory and total waiting time in different ways.

Choose a starting path for your PC

On Windows, check the GPU name and dedicated GPU memory in Task Manager. On an Apple Silicon Mac, check the chip and unified memory² in About This Mac. Then choose an installation path below. Model files and intermediate generation data both use memory, so start with one image.

A first execution path for the hardware you own
Your hardware	First run	Next check
NVIDIA · VRAM 8GB	Small checkpoint · 512×512 · one image	Save a basic image, then check model files and memory
NVIDIA · VRAM 12~16GB	ComfyUI · model-specific quantization and offload	Qwen-Image GGUF model, encoder and VAE files
NVIDIA · VRAM 24~32GB	ComfyUI · the chosen model's basic workflow	Generation time at a fixed resolution and step count
Apple Silicon Mac	ComfyUI's Metal path or a Mac GGUF runtime	Free unified memory and the model's Mac instructions
AMD Radeon · Intel Arc	An installation matching ComfyUI's support table	OS, GPU and PyTorch combination; custom-node support

With 8 GB VRAM³, start with a small checkpoint⁴ and a basic workflow that generates one 512×512 image. Once it saves successfully, check the file set and memory requirements of the model you want to try next. The standard FLUX.2 [klein] 4B configuration lists about 13 GB VRAM; an 8 GB card needs a separate path supporting quantization⁵ or CPU⁶ offload⁷.

ComfyUI hardware support · FLUX.2 klein 4B

Distinguish the app, model, and workflow

ComfyUI loads image models and connects generation steps. Model files contain the weights used to compute an image; a workflow connects loading, the prompt⁸, generation and saving. Start by completing one basic workflow supported by your installation.

Open the workflow in the first-image guide and select the checkpoint. Describe the scene in the prompt and run one image. When the result appears and a file is saved in the output folder, the first run is complete. Change the model or output size one at a time before adding custom nodes.

If a node turns red, check whether its named custom node is installed. If the model selector is empty, check the file folder against the workflow's loader. For an out-of-memory error, return to one image, lower the resolution and rerun.

An illustration of a local workstation producing thumbnails, product images, posters, and reference-image variations — Local image generation can repeat drafts, product shots, posters, and image variations on one device.

Qwen-Image 2.1 has a dedicated file configuration

Qwen-Image 2.1 was released in September 2026 and does not share the execution-file configuration of Qwen-Image-2512. Generation and reference-image editing use a dedicated VAE⁹ and Qwen3-VL encoder, so follow the file and support requirements in the installation guide for this variant. This version is under a research license. Keep the unpublished festival draft as practice; check terms before publishing for an actual event.

Set the composition on a small draft, then check final size

Once the app, model, and workflow are ready and saving works, generate a small background draft. Open the saved image and check the cinema entrance, warm windows, and empty title area on the right. If decoration fills the space, revise one condition—for example, “entrance on the left, leave the right side empty.” Do not assume a composition that works at low resolution will remain at larger output; check the publication size too.

Inpainting can change a selected area, but it is not available in the same way for every model and app combination. To revise a sign or window, check whether the workflow accepts an image input or mask. If not, adding the title or removing a small flaw in an editor may be simpler than regenerating the whole image. Keep the model name, image dimensions, and changed settings with each candidate file.

An illustration comparing one large image with several small drafts laid out like a contact sheet, showing the effects of resolution and batch size — Higher resolution and batch size change time and memory use; separate drafts from final output.

Test batch size and resolution in the actual workflow

Generating several images at once can raise peak memory because the system keeps more intermediate data. If a four-image batch runs out of memory, check the memory needed for one image and the supported resolution. When producing several candidates, compare the total completion time.

Occasionally generating one image and producing many candidates every week have different memory and waiting costs. Include whether you must close your browser or editor during generation as part of the real workflow. Record peak memory, failures, and total wait as you raise resolution or batch size¹⁰ to decide whether hardware changes are needed.

An illustration of a reference image passing through a local computer and being saved in an archive in a workspace disconnected from the network — For sensitive product images or unreleased designs, check external calls and sync paths too.

Review saved files and publication conditions

When the run completes, open the saved file at full size and inspect edge artifacts, patterns in the title area, and unintended text. Enter the event name and time in an editor rather than relying on model output, then check legibility on mobile. Before publishing, review model license, reference-photo rights, faces and trademarks, and platform policy; also check whether extensions send images externally and whether the work folder syncs automatically.

The generation demo estimates wait time from registered models and task conditions. It does not generate a new image on your device. To measure your actual workflow, record model loading, image generation, and time spent retrying failed candidates separately.

If one image saves reliably and the wait for candidates is acceptable, you can start with your current setup. If repeated generation stalls, check memory; if memory is available but generation is slow, consider an upgrade based on the supported execution path and measured bottleneck.

Terminology notes

GPU — A processor designed to handle many calculations in parallel. It performs model computations during AI inference.
Back to the text
Unified memory — An architecture where the CPU and GPU share one physical memory pool.
Back to the text
VRAM — Memory used by a graphics card’s GPU to store model weights and intermediate values.
Back to the text
Checkpoint — A file containing saved model weights and related state. Versions or tasks in one model family may use different checkpoints.
Back to the text
Quantization — A way to represent model weights or values with fewer bits, reducing storage and memory use. The coarser representation can also affect accuracy.
Back to the text
CPU — The central processor that runs the operating system and general-purpose instructions; in AI workloads it also handles tasks such as preprocessing and data movement.
Back to the text
Offloading — Moving some model data from GPU memory to system RAM or storage when capacity is limited. This adds data transfer.
Back to the text
Prompt — The input text sent to a model. For generative models it describes the desired result and strongly shapes the output.
Back to the text
VAE — Short for Variational Autoencoder. It can encode input into a compact latent representation or decode that representation into an output.
Back to the text
Batch size — The number of inputs or requests processed together in one batch.
Back to the text

Frequently asked questions

What should you check first in this article?: Local AI image generation turns a written scene into an image on your computer. Connect an image model to ComfyUI to create thumbnails and poster backgrounds. Start with GPU requirements, then work through setup and your first image.
How much memory does it need?: On Windows, check the GPU name and dedicated GPU memory in Task Manager. On an Apple Silicon Mac, check the chip and unified memory in About This Mac. Then choose an installation path below. Model files and intermediate generation data both use memory, so start with one image.
Where can you check the conditions and sources?: Qwen-Image 2.1 was released in September 2026 and does not share the execution-file configuration of Qwen-Image-2512. Generation and reference-image editing use a dedicated VAE and Qwen3-VL encoder, so follow the file and support requirements in the installation guide for this variant. This version is under a research license. Keep the unpublished festival draft as practice; check terms before publishing for an actual event.