Image generation (diffusion) How a text prompt becomes a picture, and what a local box needs to do it.
Short answer. Image generation by diffusion models starts from random noise and removes it step by step until a picture matching a text prompt appears. The method was set out in the 2020 paper on denoising diffusion probabilistic models, and latent diffusion, the basis of Stable Diffusion, made it affordable by working in a compressed image space instead of on pixels. Models run locally in tools such as ComfyUI, which says it can run large models on as little as 4 GB of VRAM with 8 GB of RAM. Limits matter for offices: Stability AI’s own SDXL model card says the model cannot render legible text and may not generate faces and people properly.
01
How diffusion works
A diffusion model is trained to remove noise from images. To make a new image it starts from pure noise and repeats the denoising step many times, steered by a text prompt. Latent diffusion researchers explain that cross-attention layers turn these models into flexible generators for inputs such as text.
Working directly on pixels is costly: the latent diffusion paper says training can take hundreds of GPU days and inference is expensive because the steps are evaluated one after another. Its fix is to work in the compressed latent space of a pretrained autoencoder, which greatly reduces the computing needed. SDXL, a later Stable Diffusion model, uses a UNet three times larger than earlier versions, a second text encoder and a separate refinement model.
02
What it needs to run
On a box with one memory pool the image model and the chat model compete for the same memory, so size the memory for both.
Item
What the sources say
Memory floor
ComfyUI says it can run even the biggest open models on as little as 4 GB of VRAM and 8 GB of RAM, using asynchronous weight streaming
Hardware
ComfyUI supports NVIDIA, AMD, Intel, Apple Silicon and Ascend GPUs
Offline use
ComfyUI says its core downloads nothing unless you request it, and an offline flag disables the optional paid API nodes
Sharing with a language model
AMD ran Stable Diffusion 3.5 Large and a 14-billion-parameter language model together on a 128 GB Ryzen AI Max+ 395 system, which shows the two kinds of model can share one memory pool
Model families
ComfyUI lists Stable Diffusion 1.5, SDXL, SD3.5 and Flux among many supported image models
03
What the open models cannot do
Stability AI’s model card for SDXL 1.0 lists limitations: the model does not achieve perfect photorealism, cannot render legible text, and may not generate faces and people properly. It also says the model was not trained to be a factual or true representation of people or events, so using it to generate such content is out of scope. Newer models differ, so read the card of each model you use.
04
Rules for office use
Licence: SDXL is released under the CreativeML Open RAIL++-M licence. Read the licence and commercial terms of every model before using its images in client work.
Ownership of what a model produces is a legal question that depends on the licence and the law; see our question on who owns what a local model writes.
Do not present generated images of real people or real events as factual.
Label AI-made images, in line with your AI disclosure policy.
Add logos and text in a design tool, because text rendering in images is unreliable.
If images must stay in the building, avoid cloud services: ComfyUI’s partner nodes give access to closed-source cloud models.
05
Where ANVE stands
ANVE generates images on the box’s own GPU through ComfyUI. Its vision models do the opposite job: they read scanned documents, photos, diagrams and video frames. ANVE has published no speed figures for either, and no sealed unit exists yet.
06
Questions people ask
Can AI generate images without the internet?
Yes, with local models. ComfyUI says it runs fully offline: its core downloads nothing unless you ask, and a flag disables the optional paid API nodes that call cloud models.
What hardware do I need to generate images locally?
Memory matters most. ComfyUI quotes as little as 4 GB of VRAM with 8 GB of RAM for large open models, while more memory is faster and lets image and language models run together. Our hardware question page has the detail.
Why does AI get text in images wrong?
The SDXL model card lists legible text as a limitation. Newer models differ, but add logos and wording in a design tool to be safe.
Who owns an AI-generated image?
That depends on the model’s licence and on the law, which differs by country. Our question page covers it. This is general information, not legal advice.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
