{/* banner prompt: A playful, polished 3D clay-render in matte sage green, cream and terracotta: a stylized NVIDIA RTX 3060 graphics card holding a large traditional Chinese ink brush, painting a tiny misty mountain landscape onto a small scroll of cream rice paper, a little red seal-stamp prop and an ink stone beside it, soft studio shadows, simple pale background, 16:9. */}

Four posts into this series, the same RTX 3060 has learned to make
[images](/blog/2026/08/03-two-years-of-open-source-image-ai-on-one-rtx-3060/),
[video](/blog/2026/08/03-now-it-moves-video-generation-on-the-same-rtx-3060/),
[video with sound](/blog/2026/08/04-now-it-talks-open-source-video-with-sound-on-the-same-rtx-3060/),
and even [reason like a 27B language model](/blog/2026/08/18-now-it-thinks-a-27b-language-model-on-the-same-rtx-3060/).
This time I wanted to point it at something harder to fake: a *specific,
demanding art tradition*. Not "an image in the style of," but **traditional
Chinese ink-wash painting, 水墨 (shui-mo), on rice paper.**

It's a good test because the style is unforgiving. Real ink painting lives in
what *isn't* painted: the empty rice paper, the 留白 (liú bái) negative space,
the way a single dry-brush stroke suggests a whole pine tree. Get it slightly
wrong and it looks like a greeting card. So: can a five-year-old graphics card,
sitting in a home office, paint a convincing misty-mountain landscape? And how
close does it get to the cloud giants?

## The machine (still the same one)

Same RTX 3060, 12GB of VRAM, 64GB of system RAM, that this whole series is built
around, a card that shipped back in **2021**. Nothing upgraded.

## The test

One prompt, given to every model: a classic 山水 (shan-shui, "mountain-water")
scene:

> *Traditional Chinese ink wash painting on xuan rice paper: misty mountains, a
> lone fisherman's boat, gnarled pines on the cliffs, drifting fog, generous
> negative space, monochrome ink with grey gradients, a small red seal.*

Three models ran **locally on the 3060**. Three ran in the **cloud**, for
comparison.

## The local heroes: painted on the 3060

Here's the headline: **it works, and it works beautifully.** The trick isn't a
giant model; it's a tiny add-on. You take a general open image model already
running on the card and bolt on a **LoRA**: a small (~40–55MB) "style adapter"
trained on Chinese ink paintings. Drop the file in, add a trigger word, and the
model paints in that style.

**SDXL + a Chinese-ink LoRA** gave the most *authentic* result: loose,
expressive, with the faint colour washes and worn look of a scroll that has hung
on a wall for a century:

![A Chinese ink-wash landscape generated locally on an RTX 3060 with SDXL and a Chinese-ink LoRA: misty mountains, a small boat with two figures, pine trees, on aged rice paper](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2026/09/now-it-paints-chinese-ink-landscapes-on-the-same-rtx-3060/sdxl-ink.png)

**FLUX.1-dev + an ink-style LoRA** was the more *refined* look: crisper pines,
deeper atmospheric mist, a flawless composition. It reads as premium modern
"digital ink" rather than a genuine antique, but as an image it's stunning:

![A Chinese ink-wash landscape generated locally on an RTX 3060 with FLUX and the zyd232 ink-style LoRA: sharp misty peaks, pine trees on cliffs, a lone boatman on a still river](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2026/09/now-it-paints-chinese-ink-landscapes-on-the-same-rtx-3060/flux-zyd232.png)

Swap in a *different* ink LoRA and Flux shifts mood: softer, mistier, the peaks
dissolving further into fog, proof that the "look" is a small file you can
change, not a whole new model to download:

![A softer, mistier Chinese ink landscape from FLUX with a second ink-style LoRA (liujiyou), peaks dissolving into fog](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2026/09/now-it-paints-chinese-ink-landscapes-on-the-same-rtx-3060/flux-liujiyou.png)

Each took **about a minute** on the 3060. Both came out of a card that cost less
than a phone, in a home office, for the price of the electricity.

## The honest numbers

<Table data={
    {
        headers: ['', 'Local (RTX 3060)', 'Cloud'],
        rows: [
            ['Setup', 'Forge + a ~50MB LoRA file', 'An API key'],
            ['Time per image', 'About 1 minute', 'Seconds to a minute'],
            ['Cost', 'Electricity', 'A fee per image'],
            ['Privacy', 'Nothing leaves the room', 'Sent to a server'],
            ['Chinese calligraphy', 'Decorative gibberish', 'Legible (gpt-image)'],
            ['Best for', 'Authentic painterly landscapes', 'When you need real text and seals'],
        ]
    }
} />

## The cloud ceiling

To see how far ahead the big players are, I gave the same prompt to three cloud
models. This is where you see both the ceiling, and the one thing local can't
yet touch.

**OpenAI's gpt-image** was the standout of the entire test. Not just
museum-grade brushwork, but **real, legible calligraphy**. It wrote an actual
classical couplet by the poet Du Fu, with a proper signature line and seal
stamps. Look at the characters in the top-left; they're *real writing*, not
decoration:

![A highly authentic Chinese ink landscape by OpenAI gpt-image, with real legible calligraphy, a signature line and red seals](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2026/09/now-it-paints-chinese-ink-landscapes-on-the-same-rtx-3060/openai-gptimage.png)

**Zhipu's GLM (CogView-4)**, a Chinese model, produced a clean, textbook
shan-shui, confident and correct, if a little safe:

![A clean monochrome Chinese ink landscape by Zhipu GLM CogView-4](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2026/09/now-it-paints-chinese-ink-landscapes-on-the-same-rtx-3060/glm-cogview.png)

**Google's Gemini** did something unexpected. It rendered the painting as a
fully *mounted hanging scroll on a wall*, silk border and all, plus two zoomed
detail crops. Impressive, but more a product photo than a painting:

![Google Gemini's take, showing the ink painting as a mounted hanging scroll on a wall with detail crops](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2026/09/now-it-paints-chinese-ink-landscapes-on-the-same-rtx-3060/gemini.png)

## What I learned: the writing is the tell

Here's the interesting part. On the *painting*, the mountains, the mist, the
pines, the mood, the local 3060 results genuinely hold their own against the
cloud. Side by side, you'd struggle to pick the home-GPU images out of a lineup.

The gap is the **calligraphy.** Every local model painted beautiful *decorative*
brushstrokes where the inscription should be: squiggles that look like writing
but say nothing. Only gpt-image produced characters you can actually read. If
your painting needs a real poem or a legible seal, that's still a cloud job.
For everything else, the art itself, the card on your desk is enough.

## Where this leaves a small business

If you sell tea, run a restaurant, teach calligraphy, or just want a beautiful,
culturally-grounded header image for a newsletter, you no longer need a stock-art
subscription or a design agency. A modest computer you may already own can paint
you an original ink landscape in about a minute, privately, for the cost of the
power it draws, and you can generate a hundred variations until one is right.

The cloud still wins when the words on the page have to be real. But the *art*?
That quietly moved into the home office, onto a graphics card from 2021.

*Same card. Five years old. Now it paints.*


c07ca0c52d37ba12acd90bc4f968ebb08cc50c81babd70ce1b1c5438d1938837

Generating traditional Chinese ink-wash painting (shui-mo / guohua) locally on a consumer RTX 3060 with SDXL and FLUX plus ink-style LoRAs, compared against cloud models (OpenAI gpt-image, Google Gemini, Zhipu GLM/CogView): real results, honest limits, and the one thing local models still can't do.

The fifth in our RTX 3060 series. We test whether a five-year-old 12GB graphics card can paint traditional Chinese ink-wash landscapes (shui-mo) on rice paper, locally and with open models, and pit the results against the cloud heavyweights.

Now It Paints: Chinese Ink Landscapes on the Same RTX 3060

{/* banner prompt: A playful but polished 3D render of a stylized RTX 3060 as a small factory on a simple background, using matte clay-like materials in a color story of sage green, cream, and terracotta; a conveyor belt on the left carries misshapen, glitchy figures, while a belt on the right carries perfectly formed photorealistic objects. */}

Eighteen months ago I wrote a tutorial on
[mastering human portraits with Stable Diffusion](/blog/2025/02-06-mastering-human-portraits-with-stable-diffusion-ai/).
Reading it back is a little painful. It opens with "trust me, this is hard," and
its list of search keywords reads like a bug report: *two heads, extra heads,
deformed face, extra eyes, disfigured, duplicate.* Those weren't edge cases,
they were what you fought on every render.

This week I ran open-source image generation on the **exact same computer**. Same
RTX 3060, same 12GB of VRAM. The only thing that changed is the software, and the
gap is hard to overstate.

## The machine (deliberately unchanged)

It's the same box from our [local-LLM
benchmark](/blog/2026/05/11-64gb-ram-12gb-vram-the-honest-local-llm-benchmark/):

- **CPU:** Intel Core i7-12700F
- **RAM:** 64GB DDR4
- **GPU:** NVIDIA GeForce RTX 3060, 12GB VRAM
- **OS:** Debian 13

Holding the hardware constant is the whole point. Any improvement you see is the
open-source ecosystem getting better, not a new graphics card.

## Early 2025: Stable Diffusion 1.5, 512 pixels, and a war on anatomy

Back then, the state of the art you could actually run at home was Stable
Diffusion 1.5. My old parameters tell the story: 512×512, Euler a, CFG 7. To get
anything lifelike you leaned on community finetunes (RealisticVision, ChilloutMix,
F222), stacked LoRAs, and wrote paragraph-long negative prompts to *suppress* the
model's habit of adding a second head or a sixth finger. Then you generated a
dozen and kept the one that wasn't broken. It worked, but it was a craft, and a
grind.

## Mid 2026: the same card, a different league

Here's a portrait from the same 3060 today, straight out of the model,
no inpainting, no ControlNet, no clean-up:

![A photoreal portrait generated on the RTX 3060 in 2026](https://www.tyolab.com/api/tynion/images/f9f8fb60-d635-1048-cb10-c3492d7a4c4a/file)

Three things changed the game. **Flux** and modern **SDXL** finetunes replaced
SD 1.5 as the open baseline. They render natively at 1024px and up, and they get
hands, faces and proportions right the first time far more often than not.

They also learned to spell, the single clearest "we crossed a line" moment.
Readable text in an image was essentially impossible in 2025:

![A shopfront with legible hand-painted signage, generated locally](https://www.tyolab.com/api/tynion/images/3fa3d0ac-f3af-f6e7-952e-27b04f9b545e/file)

And the range on one card is wide: photoreal wildlife and clean anime styles
from the same machine, same afternoon:

![A photorealistic wildlife shot from the same GPU](https://www.tyolab.com/api/tynion/images/adbceadb-bc09-46f0-c8c7-bd76f312361f/file)

![A detailed anime illustration from the same GPU](https://www.tyolab.com/api/tynion/images/3d737b80-9cbb-186f-ce2c-c496fb918589/file)

## Then and now, side by side

<Table data={
    {
        headers: ['', 'Early 2025', 'Mid 2026'],
        rows: [
            ['Best model you can run', 'Stable Diffusion 1.5 + finetunes', 'Flux, SDXL finetunes'],
            ['Native resolution', '512 x 512', '1024px and up'],
            ['Anatomy', 'Fighting extra heads, hands, eyes', 'Right on the first try, mostly'],
            ['Text in the image', 'Gibberish', 'Often readable'],
            ['Keepers per ten tries', 'One or two', 'Most of them'],
            ['Effort per shot', 'Long negative prompts, LoRAs, inpainting', 'A plain-English sentence'],
            ['Video', 'Not on this card', 'A few seconds, locally'],
        ]
    }
} />

## The part that was pure science fiction in 2025: video

The same RTX 3060 now generates **video**. Wan 2.2 and LTX-Video produce short
text-to-video and image-to-video clips locally: a five-second scene in a few
minutes, on a card that two years ago struggled with a single 512px still. The
next open model on our list, MiniMax H3, even adds a synchronised soundtrack.

## What actually changed, and why it matters

It wasn't the GPU. Three shifts in the open-source world did the work:

- **Better base models.** Flux and SDXL are simply generations ahead of SD 1.5.
- **Quantisation.** Techniques like fp8 and 4-bit weights shrink 12-billion-parameter
  models enough to fit inside 12GB of VRAM with quality mostly intact.
- **Smarter memory management.** The tools now offload the heavy text encoders to
  system RAM, so the GPU only holds what it must.

Put together, a consumer card that launched years ago runs models that didn't
exist when you bought it, because the community learned to make them fit.

For a small business or a hobbyist, that's the real headline: the *same* modest
hardware budget now buys studio-grade stills and short video, from software
that's free and runs on your own desk. Every image in this post was generated on
that one 3060; you can browse and remix the prompts in our
[gallery](https://www.tyolab.com/prompts).

*Next up: putting MiniMax H3, open-weights video with native audio, on the very
same card, and seeing how far 12GB can really go.*


7fccf546dccfab1e31d5dea9d7e220f6bd2721130b85ec2280897e22f3bb17af

A hands-on before-and-after: how open-source text-to-image models improved on the exact same RTX 3060 12GB between early 2025 and 2026, comparing Stable Diffusion 1.5 vs Flux and SDXL finetunes, resolution, anatomy, text rendering, hit-rate, and video.

We ran open-source image generation on the same RTX 3060 we used 18 months ago. The hardware is identical; the results are unrecognisable: from 512px portraits fighting extra heads to 1024px Flux, near-perfect anatomy, readable text, and even video.

Two Years of Open-Source Image AI on One RTX 3060


# Creating Celebrity Look-Alike Images with Stable Diffusion  

## Introduction  

We all have idols or people we admire, figures from movies, music, sports, or history who have left a lasting impression on us. Naturally, we enjoy seeing familiar faces in the work we create, whether for artistic inspiration, storytelling, or just for fun. With AI tools like Stable Diffusion, it’s now possible to generate images that resemble well-known individuals.  

Stable Diffusion has become one of the most powerful AI models for generating photorealistic images, allowing users to create portraits that bear a striking resemblance to real celebrities. However, while this technology is impressive, it also raises ethical and legal concerns. In this guide, we’ll explore how Stable Diffusion can be used to generate celebrity look-alike images, the best techniques to achieve high-quality results, and the responsible use of this technology.  
 

## Understanding Stable Diffusion’s Capabilities  

Stable Diffusion is a text-to-image AI model that can generate detailed images based on prompts. While it doesn’t directly recreate real celebrities, users can fine-tune prompts and apply specific techniques to achieve results that resemble well-known individuals.  

If you are not familiar with how to use Stable Diffusion, we recommend checking out a few guides listed below:
- [Using Stable Diffusion WebUI Docker for Image Generation – A Newbie's Guide](/blog/2025/02-05-using-stable-diffusion-webui-docker-for-image-generation-a-newbies-guide/)
- [Stable Diffusion Parameter Guide](/blog/2025/02-05-stable-diffusion-parameter-guide/)
- [Mastering Human Portraits](/blog/2025/02-06-mastering-human-portraits-with-stable-diffusion-ai/)

### Key Features of Stable Diffusion for This Task  
- **Prompt Control** – Through careful prompt control (prompt scheduling and prompt weighting), you can achieve the desired effect and likeness.  
- **Fine-Tuning** – Using pre-trained models that enhance facial similarities.  

## Prompt Control

### Some Warm Ups

Through some testing, we can see how well Stable Diffusion works in generating images that resemble well-known individuals.

#### Attempt #1

>Abraham Lincoln

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/creating-celebrity-look-alike-images-with-stable-diffusion/00100-206793973.png)

#### Attempt #2

**Parameters:**

- **Steps:** 20  
- **Sampler:** DPM++ 2M  
- **Schedule type:** Karras  
- **CFG scale:** 12  
- **Seed:** 4221999450  
- **Size:** 512x512  
- **Model hash:** c6bbc15e32  
- **Model:** sd-v1-5-inpainting  
- **Conditional mask weight:** 1.0  
- **Version:** v1.9.4  

Now adding more details:

>A highly detailed, colorful, photorealistic portrait of Abraham Lincoln, 4K resolution, dramatic lighting, sharp facial details, deep wrinkles, strong jawline, piercing eyes, dignified expression, black suit with bow tie, cinematic composition, DSLR shot, volumetric lighting, ultra-HD, realistic skin texture, ultra quality, dof, film grain, Fujifilm XT3, crystal clear

**Negative prompt:**
>deformed face, disfigured, asymmetric, duplicate, tattoos, marks, Extra heads, extra faces, multiple faces, extra eyes, immature, cartoon, anime, 3d, painting, b&w, multiple figures, multiple heads, extra hands

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/creating-celebrity-look-alike-images-with-stable-diffusion/00101-4221999450.png)

Stable Diffusion does a great job of recreating the likeness with a more detailed prompts.

### Prompt Scheduling and Weighting

Now we are gonna play with prompt scheduling and weighting to achieve the desired effect.

#### **Prompt Scheduling:**  

Prompt scheduling in Stable Diffusion refers to the technique of modifying the text prompt dynamically over the course of the image generation process. Instead of using a single static prompt, prompt scheduling allows for gradual changes in the description, influencing how the model interprets and renders the image over multiple denoising steps. With this technique, we can precisely control the generated person's face and seamlessly blend the features of two celebrities.

Prompt scheduling is typically available in automatic1111’s WebUI, for example:
>A photorealistic portrait of a woman | A Van Gogh-style painting of the same woman :0.5

The :0.5 ratio means the prompt starts shifting at 50% of the total steps.

Now back to the task at hand, we change the prompt to:

>A highly detailed, colorful, photorealistic portrait of Abraham Lincoln | Henry Cavill :0.5, 4K resolution, dramatic lighting, sharp facial details, deep wrinkles, strong jawline, piercing eyes, dignified expression, black suit with bow tie, cinematic composition, DSLR shot, volumetric lighting, ultra-HD, realistic skin texture, ultra quality, dof, film grain, Fujifilm XT3, crystal clear

Here are our **Cavilcoln**:

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/creating-celebrity-look-alike-images-with-stable-diffusion/00103-312295037.png)

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/creating-celebrity-look-alike-images-with-stable-diffusion/00104-1101313381.png)

#### **Prompt Weighting:**  

Prompt weighting in Stable Diffusion allows users to assign different levels of importance to specific words or phrases in a prompt. This helps the AI model prioritize certain details over others when generating an image.

By adjusting the weight of words, you can fine-tune the composition, style, or specific elements of the image without needing extensive prompt rewriting.

**Usage:**

In prompt, parentheses () and brackets [] are used for weighting:
- Parentheses () increase emphasis on a word or phrase.
- Brackets [] decrease emphasis, making an element less prominent.

Now we change the prompt to:

>A highly detailed, photorealistic image of Abraham Lincoln | Henry Cavill :0.6 sitting on a (chair:2.3), black suit with bow tie, [history accuracy] , 4K resolution, cinematic lighting, DSLR shot, ultra-HD, realistic skin texture,  cinematic composition, DSLR shot, volumetric lighting, ultra-HD, realistic skin texture, ultra quality, dof, film grain, Canon 70D

**Parameters:**
Steps: 20, Sampler: DPM2, Schedule type: Karras, CFG scale: 10, Seed: 1101313381, Size: 512x512, Model hash: c6bbc15e32, Model: sd-v1-5-inpainting, Conditional mask weight: 1.0, Version: v1.9.4

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/creating-celebrity-look-alike-images-with-stable-diffusion/00106-1101313381.png)


## Using Pre-trained Models for Better Results

There are models that have been tuned specifically for creating human portrait images. A list of those models are:

- **F222**: Known for better face generation.
- **RealisticVision**: A popular choice for photorealistic AI art.
- **ChilloutMix**: Good for anime-style and realistic portraits, especially for Asian facial features.

For the details of how to use these models, please check out [this article](/blog/2025/02-06-mastering-human-portraits-with-stable-diffusion-ai/).

Now we test the prompts that have been used previously. Below you can see the difference between Stable Diffusion base model and the models mentioned above.

With model **F222**:

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/creating-celebrity-look-alike-images-with-stable-diffusion/00113-4237562600.png)

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/creating-celebrity-look-alike-images-with-stable-diffusion/00114-1779733346.png)


## Final Thoughts  

While AI-generated images can be used for artistic and creative purposes, **it is crucial to avoid generating images that impersonate real people for deceptive, defamatory, or commercial exploitation.** And remember, these techniques have the potential to be used for unauthorized purposes, so consider the legal and ethical implications of your work.  

Stable Diffusion offers an exciting way to explore AI-generated art, but ethical responsibility is key. By focusing on creativity while respecting privacy and legal boundaries, users can harness this technology without engaging in harmful practices.  

## References

- [How to generate realistic people in Stable Diffusion](https://stable-diffusion-art.com/realistic-people/)


<ContactFooter />

88d180d442f28dae6e692fd5d176ee4a806cbe0009d0b6fdd1e6e1cf01aef7cd

In this article, we will cover the steps required to create celebrity look-alike images with Stable Diffusion, including prompt control, fine-tuning, and advanced techniques for enhancing results.

In this tutorial, we will explore the steps required to create celebrity look-alike images with Stable Diffusion.

Creating Celebrity Look-Alike Images with Stable Diffusion


# Mastering Human Portraits with Stable Diffusion AI

Trust me, using Stable Diffusion to create lifelike, realistic human images is hard. While AI-generated portraits have improved significantly, achieving high-quality, photorealistic results requires understanding prompts, fine-tuning models, and leveraging the right tools. In this article, we will explore techniques and strategies to master human portrait generation with Stable Diffusion AI.

## Understanding Stable Diffusion for Portraits

Stable Diffusion is a powerful AI model for image generation, but its default outputs often lack the fine details necessary for realistic human portraits. Achieving high fidelity requires:

- **Prompt Engineering**: Crafting detailed and structured prompts.
- **CFG Scale & Sampling Methods**: Adjusting parameters for optimal results.
- **Fine-tuned Models**: Using specialized models trained on high-quality human images.
- **ControlNet & Inpainting**: Refining features with additional tools.

In this article, we will focus on prompt engineering and fine-tuning models to create lifelike human portraits, highlighting the importance of these techniques because they can significantly impact the quality of the final result. Choosing a right CFG Scale or sampling method, it may help in some scenarios, but it may not always work. Also using ControlNet and Inpainting requires advanced knowledge and skills, and they will be not covered in this article.

## Let's Dive in

### Parameters 

- **Steps**: 20  
- **Sampler**: Euler a  
- **Schedule type**: Karras  
- **CFG scale**: 7  
- **Seed**: -13456  
- **Size**: 512x512  
- **Model hash**: c6bbc15e32  
- **Model**: sd-v1-5-inpainting  
- **Conditional mask weight**: 1.0  
- **Version**: v1.9.4  

### Let's Prompt

Here are some basic prompts that can help you get started:

- **A highly detailed portrait of a woman**.
- **A highly detailed portrait of a man**.
- **A highly detailed portrait of a child**.
- **A highly detailed portrait of a person**.

Type in prompt:
>A highly detailed portrait of a woman

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00064--13456.png)

Not quite what we expected, let's add more details to the prompt:

>A highly detailed portrait of a woman standing next to a window of a bridal dress shop.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00069--13456.png)

It is getting better, let's add more details:

>A highly detailed portrait of a young woman standing next to a window of a bridal dress shop, facing the camera

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00074--13456.png)

So the more details we add to the prompt, the better the result.

Now let's add add a negative prompt to prevent the model from generating unnecessary elements:

>deformed face, disfigured, asymmetric, duplicate, tattoos, marks

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00079--13456.png)

Wow, now that is a lot better! But it is still not there yet.

Let's keep working on that:
>A highly detailed portrait of a young woman standing next to a window of a bridal dress shop, facing the camera, ultra-realistic, cinematic lighting, DSLR sharp focus, soft shadows, perfect skin texture, 8K resolution, photorealism, HDR

Finger crossed!!!

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00084--13456.png)

How unfortunate that the model still generates disfigured faces.

Don't give up, let's keep working on that.

**Change the parameters:**

- **Steps**: 20  
- **Sampler**: DPM2  
- **Schedule type**: Karras  
- **CFG scale**: 8  
- **Seed**: 13456  
- **Size**: 512x512  
- **Model hash**: c6bbc15e32  
- **Model**: sd-v1-5-inpainting  
- **Conditional mask weight**: 1.0  
- **Version**: v1.9.4  

**Change the prompt:**
>A highly detailed portrait of a young woman with highlight hair, standing next to a window of a bridal dress shop looking at the camera.  The face is illuminated by soft, natural lighting, highlighting lifelike skin texture, expressive eyes, and a natural facial expression. The image has a cinematic look, with balanced contrast, depth of field,  studio lighting, ultra-realistic, cinematic lighting, DSLR sharp focus, soft shadows, perfect skin texture, 8K resolution, photorealism, dslr, ultra quality, tack sharp, dof, film grain, Fujifilm XT3, crystal clear,

We throw lots of fancy words in it

**Change the negative prompt:**
>deformed face, disfigured, asymmetric, duplicate, tattoos, marks, Extra heads, extra faces, multiple faces, extra eyes, ugly, bad, immature, cartoon, anime, 3d, painting, b&w

Tadah!

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00089-13456.png)

Finally that is something we can use.

I suspect words like "film grain, Fujifilm XT3" are making the difference.

So with the same negative prompt, we are gonna change the prompt to:

>A highly detailed portrait of a young woman with highlight hair, standing next to a window of a bridal dress shop. She is looking at the camera, so we can see her eyes.  full torso, shoulders and chest visible. The face is illuminated by soft, natural lighting, highlighting lifelike skin texture, expressive eyes, and a natural facial expression. The image is symmetric and has a cinematic look, with balanced contrast, studio lighting, ultra-realistic,  DSLR sharp focus, soft shadows, perfect skin texture, ultra quality, dof, film grain, Fujifilm XT3, crystal clear,

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00094-456.png)

Not bad.

Now we want to make an image of a different size with aspect ratio of 16:9, so we can use this image for the marketing material of a bridal dress shop.

Oops, new issue!

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00096-13457.png)

The image has two figures.

This is a common issue with Stable Diffusion, we could either 1) stay with size of 512x512, or 2) adjust both the positive and the negative prompts, change the CFG scale, seeding, the sampling method or lots of image generations.

Now the new parameters are:

- **Steps**: 20  
- **Sampler**: DPM++ 2M  
- **Schedule type**: Karras  
- **CFG scale**: 15  
- **Seed**: 457  
- **Size**: 910x512  
- **Model hash**: c6bbc15e32  
- **Model**: sd-v1-5-inpainting  
- **Conditional mask weight**: 1.0  
- **Version**: v1.9.4  

New prompt:
>A highly detailed portrait of a young woman with highlight hair, standing next to a window of a bridal dress shop. She is looking at the camera, so we can see her eyes.  full torso, shoulders and chest visible. The face is illuminated by soft, natural lighting, highlighting lifelike skin texture, expressive eyes, and a natural facial expression. In the background, on the woman's right, there is a table. The image has a cinematic look, with balanced contrast, studio lighting, ultra-realistic,  DSLR sharp focus, soft shadows, perfect skin texture, ultra quality, dof, film grain, Fujifilm XT3, crystal clear,

New negative prompt:
>deformed face, disfigured, asymmetric, duplicate, tattoos, marks, Extra heads, extra faces, multiple faces, extra eyes, ugly, bad, immature, cartoon, anime, 3d, painting, b&w, multiple figures, multiple heads

The idea of adding the table in the description is to fill up the space so Stable Diffusion won't generate two heads.

Finally we have a nice one:

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00097-457.png)


### More Examples

#### Red haired woman

##### Prompt:
>A photograph of a red haired woman, realistic face, symmetrical, highly detailed, ultra HD

##### Negative prompt: 
>Extra heads, extra faces, multiple faces, extra eyes, deformed face, disfigured, asymmetric, duplicate, tattoos, marks

##### Parameters Used:

- **Steps**: 139
- **Sampler**: DPM++ 3M SDE
- **Schedule type**: Karras
- **CFG scale**: 12
- **Seed**: 655
- **Size**: 512x768
- **Model hash**: c6bbc15e32
- **Model**: sd-v1-5-inpainting
- **Conditional mask weight**: 1.0
- **Version**: v1.9.4

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00011-655.png)

#### A Young Woman

##### Prompt:
>a young woman with gorgeous face, delicate features, symmetrical, highly detailed

##### Negative prompt: 
>Extra heads, extra faces, multiple faces, extra eyes, deformed face, disfigured, asymmetric, duplicate, tattoos, marks

##### Parameters Used:
- **Steps**: 20
- **Sampler**: UniPC
- **Schedule type**: Karras
- **CFG scale**: 16
- **Seed**: 655
- **Size**: 512x512
- **Model hash**: c6bbc15e32
- **Model**: sd-v1-5-inpainting
- **Conditional mask weight**: 1.0
- **Version**: v1.9.4

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/00027-655.png)

In order to get one you really like, you may have to repeat this process until you get a good result. If you don't want to go through this tedious process, you can always use a model that is specifically trained to generate realistic human portraits.

## Choosing the Right Model

The base Stable Diffusion model may not always generate photorealistic people without proper prompting. Instead, consider using:

- **F222**: Known for better face generation.
- **RealisticVision**: A popular choice for photorealistic AI art.
- **ChilloutMix**: Good for anime-style and realistic portraits, especially for Asian facial features.
- **Custom LoRAs**: Fine-tuned weights designed for specific enhancements in face rendering.

In this article, we will test generating human portraits with model **F222**.

### Model Installation

### F222

Presumably you are running Stable Diffusion with **Stable Diffusion WebUI Docker** with **Automatic1111 UI**.

In the directory where you have installed Stable Diffusion WebUI docker, find the folder `data/models/Stable-diffusion`, you can see the following files:

```
-rw-r--r-- 1 root root 4265437280 Feb  4 13:39 sd-v1-5-inpainting.ckpt
-rw-r--r-- 1 root root 4265380512 Feb  4 13:39 v1-5-pruned-emaonly.ckpt
```

So go to the folder and run the following command to download the `f222` model.

```
cd data/models/Stable-diffusion
wget https://huggingface.co/acheong08/f222/resolve/main/f222.ckpt
```

After downloading, on the top left corner of the UI, click on the refresh button, then you can see the installed `f222` model shown in the dropdown menu.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/2025-02-07-19-07-21.png)

#### ChilloutMix

```bash
cd data/models/Stable-diffusion
```

Assume you have `curl` installed.

```bash
curl https://civitai.com/api/download/models/11745 -o ChilloutMix.safetensors
```

Or with `wget`

```bash
wget https://civitai.com/api/download/models/11745
mv 11745 ChilloutMix.safetensors
```

If you dont have `curl` installed, you can download the model from [Civitai](https://civitai.com/models/11745) and save it to `data/models/Stable-diffusion`.


### Testing the Model

After selecting the `f222` model, we can generate the image with the same prompts as before.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00098-458.png)

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/00099-457.png)

This model does generate nice images, but it cannot eliminate the extra figure in the image completely. So we need to run it a few times to find some good ones.

## Making it a Billboard Ad for a Bridal Dress Shop

Since we have the ideal image, we can create a billboard ad for a bridal dress shop.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/mastering-human-portraits-with-stable-diffusion-ai/billboard.png)

## Final Thoughts

Mastering human portraits with Stable Diffusion AI requires patience and practice. By selecting the right model, refining your prompts, and leveraging additional tools like ControlNet and inpainting, you can achieve stunning, lifelike AI-generated portraits. Experiment with different settings and techniques to improve your results over time.

Are you ready to create your own masterpiece? Start experimenting with Stable Diffusion today and push the boundaries of AI-generated human portraits! For details about how to run Stable Diffusion with user-friendly UI, check out [Using Stable Diffusion WebUI Docker for Image Generation – A Newbie's Guide](/blog/2025/02-05-using-stable-diffusion-webui-docker-for-image-generation-a-newbies-guide/).

If you need to create a new prompt from scratch and choose an appropriate negative prompt, consider using a text-to-image generation prompt helper like the one we provide at [Tynion](https://www.tyolab.com/tynion/).

## References

- [How to generate realistic people in Stable Diffusion](https://stable-diffusion-art.com/realistic-people/)


<ContactFooter />

4e2e09a48db7ddd4e32283931aafd8e337bf534a2cf4aabb5d7603997a84e74f

In this article, we will cover the steps required to master human portraits with Stable Diffusion AI, including model selection, prompt engineering, and advanced techniques for enhancing results.

In this tutorial, we will explore the steps required to master human portraits with Stable Diffusion AI.

Mastering Human Portraits with Stable Diffusion AI


# Stable Diffusion Parameter Guide

**Stable Diffusion** is a powerful AI image generation model that transforms text prompts into high-quality visuals. To get the best results, it's essential to understand the various parameters that influence image generation. This guide explains key parameters, their impact, and how to fine-tune them for optimal results.

## Classifier Free Guidance (CFG)

CFG controls the balance between creativity and prompt adherence. The default value is 7. Lower values increase randomness, while higher values ensure stronger adherence to the prompt. This parameter does not affect VRAM usage or generation time.

CFG 2-6: High creativity, but less aligned with the prompt.

CFG 7-10: Recommended for most cases, balancing creativity and control.

CFG 10-15: Useful when the prompt is highly detailed and specific.

CFG 16-20: Not recommended unless the prompt is exceptionally detailed, as it may reduce coherence and image quality.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/merged_cfg_comparison.png)

## Steps (Sampling Steps / Precision)

Step (Sampling Steps) has a default value of 50. Stable Diffusion begins with a noise-filled canvas and gradually removes noise to produce the final image. The Step parameter controls the number of denoising iterations. Generally, the higher the step count, the better the image quality. Steps do not affect VRAM usage, but generation time increases proportionally.

For beginners, using the default value of 50 is recommended.

Higher steps = better image quality but longer generation time.

Lower steps = faster generation but may result in less detailed images.

Higher step counts (75-100) may improve fine details but with diminishing returns beyond a certain point.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/merged_image.png)

## Seed (Random Seed Control)

The Seed parameter determines the initial noise used for image generation. The default value is -1, meaning a random seed is used for each generation.

Using the same seed with identical prompts and settings will reproduce the same image.

Changing the seed results in a different image, even with the same prompt.

If a particular seed produces high-quality results, it can be saved and reused.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/merged_seed_image.png)

## Sampler (Diffusion Sampler Method)

The Sampler determines how Stable Diffusion removes noise during image generation. Different samplers use different mathematical methods to estimate the next denoising step, impacting the final image’s smoothness, detail, and coherence. Some samplers require more steps for high-quality output, while others can achieve good results with fewer iterations.

### Common Samplers and Their Effects:

- **DDIM (Denoising Diffusion Implicit Models):** Fast and efficient; requires fewer steps (10-25) but may sacrifice fine details.

- **PLMS (Predictive Linear Multistep):** Faster than DDIM but sometimes struggles with detailed features.

- **Euler / Euler A (Ancestral):** Sharp and well-defined images; Euler A adds randomness, making results more varied.

- **Heun:** Balances detail and smoothness.

- **DPM++ (Denoising Probabilistic Models):** Advanced sampler offering high-quality outputs with fewer steps.

- **LMS (Linear Multistep):** Provides smooth gradients and balanced details.

### Choosing a Sampler:

If speed is a priority, use DDIM (fewer steps, faster results).

If detail is important, use Euler or DPM++.

If experimenting with creative variance, use Euler A or PLMS.

Experimenting with different samplers may yield varying results in style and detail.

Here are the recommended samplers for different styles.

{/*
Style	Recommended Sampler
Photorealistic animals (dogs, cats, wolves, birds, etc.)	DPM2 Karras
Highly detailed textures (fur, feathers, scales)	DPM2
Sharp and well-defined details	Euler
Smooth shading and softer fur rendering	Heun
Strong contrast and bold textures	LMS
Quick experiments and artistic stylization	DDIM
*/}

<Table data={
    {
        headers: ['Style', 'Recommended Sampler'],
        rows: [
            ['Photorealistic animals (dogs, cats, wolves, birds, etc.)', 'DPM2 Karras'],
            ['Highly detailed textures (fur, feathers, scales)', 'DPM2'],
            ['Sharp and well-defined details', 'Euler'],
            ['Smooth shading and softer fur rendering', 'Heun'],
            ['Strong contrast and bold textures', 'LMS'],
            ['Quick experiments and artistic stylization', 'DDIM'],
        ]
    }
} />

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/merged_dpm_image.png)

## Text Prompt for Image Generation

Stable Diffusion generates images based on textual descriptions, determining what appears in the final image. A well-structured prompt significantly impacts output quality. 

### General Structure of a Prompt

Below are key elements to consider when crafting prompts:

- Subject: Clearly specify the main object (e.g., a dog, a cat, a bear, a warrior with a sword).

- Image Type: Define the medium: painting, photograph, watercolor, oil painting, comic, digital art, illustration, 3D render, sculpture, etc.

- Artist/Style: Combine multiple artists or describe the style explicitly (e.g., "Studio Ghibli + Van Gogh" or "very coherent symmetrical artwork").

- Color Scheme: Define the dominant colors (e.g., "blue color scheme").

- Reference Platform: Mention where similar styles appear (e.g., "using styles from the trending on CIVITAI").

### Example Prompt #1:

>A photograph of a red haired woman, realistic face, symmetrical, highly detailed, ultra HD using the style trending on CIVITAI.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/00022-655.png)

### Example Prompt #2:

>A photograph of a dark haired woman, realistic face, symmetrical, highly detailed, ultra HD using the style trending on Pixiv.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/00024-655.png)

## Advanced Prompting Techniques

To improve generated images, add detailed descriptions:

- Secondary Elements: Add 2-3 additional objects (e.g., "beautiful forest background, desert!!, (((sunset)))").

- Character Features: Add details like "detailed gorgeous face, silver hair, olive skin, delicate features".

- Special Effects: Use terms like "insanely detailed, intricate, surrealism, smooth, sharp focus, 8K resolution".

- Lighting Effects: Specify lighting conditions (e.g., "Natural Lighting, Studio Lighting, Cinematic Lighting, Backlight, Rim Light").

- Camera Angles: Use photographic terms (e.g., "Cinematic, Golden Hour, Depth of Field, Side-View, Bokeh").

- Camera Position: Specify the location (e.g., "Front View, Back View, Left View, Right View").

- Camera Type: Mention the camera model (e.g., "Canon 70D, Nikon D850, Fujifilm X100, Fujifilm XT3").

- Image Quality: Ensure high fidelity by adding "professional, award-winning, groundbreaking, ultra HD, film grain".

### Example Prompt #3:
>a young woman with gorgeous face, delicate features, symmetrical, highly detailed 


![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/00027-655.png)

## Prompt Weighting

Stable Diffusion allows weight adjustments to control element emphasis:

Increase prominence: (element:10), e.g., A planet in space:10 | bursting with color:4 | aliens:-10.

Decrease prominence: (element:-10), e.g., tree:-10 ensures no trees appear.

Word order matters: Earlier words have higher weight. If an element is missing, move it to the beginning.

## Negative Prompts

Negative prompts exclude unwanted elements from the image. This is useful for avoiding distortions, unwanted themes, or specific artifacts.

### Common Negative Prompts

- **Fix anatomy issues**: disfigured, deformed hands, blurry, grainy, broken, extra digits, bad anatomy.

- **Avoid nudity**: nudity, bare breasts.

- **Avoid black-and-white images**: black and white, monochrome.

For example, same Prompt as Example **#1** with Negative Prompting:

>Extra heads, extra faces, multiple faces, extra eyes, deformed face, disfigured, asymmetric, duplicate, tattoos, marks

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/00023-655.png)

## Prompt Matrix (Variations in Image Generation)

The Prompt Matrix function generates multiple variations of an image by combining different elements. This is particularly useful for creating assets for animations or video sequences.

In a prompt:

- The elements that remain constant across all images in the matrix are enclosed in &lt;angle brackets&gt;.

- The elements that vary between images are preceded by a pipe character (|). This character is later replaced with either a comma or a space, depending on the script's configuration.

For example, generating an image sequence for a white dog like husky with different lighting conditions:

>&lt;a white dog like husky, 4k, &gt; lights | Night:1 | Sunrise:2 | Rain:3 | Snow:4

This will generate different versions of the same image under various lighting conditions.

![image](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/00028-2107875282.jpg)

If you are on Automatic1111, you can select the Prompt Matrix function by clicking on the "Script" button under the "Seed" section.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/stable-diffusion-parameter-guide/2025-02-06-15-16-43.png)

If you're interested in converting images into animations, consider using [Deforum Stable Diffusion](https://deforum.art/).

## Conclusion

Understanding Stable Diffusion’s parameters enables precise control over image generation, improving quality and efficiency. By experimenting with different CFG values, steps, samplers, and prompt structures, users can fine-tune their creations to achieve stunning results.

For further enhancements, tools like [PromptHero](https://prompthero.com/), [OpenArt](https://openart.ai/), and [PromptMatrix](https://promptmatrix.com/) can help you find existing prompts and discover new creative possibilities. If you need to construct a new prompt from scratch, consider using an Text-to-Image generation prompt helper like we provide at [Tynion](https://www.tyolab.com/tynion/).

<ContactFooter />

ded30c4945a653407de7142be12c2f7641eb1c7e576fe894c4ebb1452fb0c681

In this guide, we explore the key parameters of Stable Diffusion, an AI image generation model, and how to fine-tune them for optimal results.

Stable Diffusion Parameter Guide


# How to Generate Images with AI Models on Your Own Computer

## **Introduction**  

AI-generated images have revolutionized the creative industry, enabling anyone to create stunning visuals without requiring advanced artistic skills. While many online services offer AI image generation, running models on your own computer gives you greater control, privacy, and flexibility, and it is totally free.  

This guide will walk you through the process of setting up and using AI models for image generation locally. Whether you're an artist exploring AI-assisted creativity or a developer integrating AI into your projects, you'll learn how to install the necessary software, choose the right models, and generate images efficiently on your machine.  

## **Setting Up Your Environment**  

Before running AI models, you need to set up your system properly.  

### **Hardware Requirements**  
- **GPU** – A dedicated GPU (NVIDIA recommended) with at least 8GB VRAM for efficient performance.  
- **RAM** – At least 16GB RAM is recommended, though some models may work with less.  
- **Storage** – AI models can be large (2GB–10GB per model). Ensure you have sufficient SSD storage. 

If you want to run a particular model you desired, which might require a lot of VRAM, make sure that the GPU has enough VRAM.

### **Software Requirements**  

#### **Operating System**

While you can use run AI models on all main operating systems including Windows, MacOS and Linux. This guide will cover running AI models on only Ubuntu or other Linux distributions.

#### **Installing Python and Dependencies**  
1. **Install Python (3.8 or later)**

```bash
sudo apt install python3
sudo apt install python3-venv
sudo apt install python-is-python3
```

2. **Initialize Virtual Environment**:  

You may need `gnome-terminal` to run the commands.

```bash
python -m venv ai-env
source ai-env/bin/activate
```

3. **Install Dependencies**

```bash
pip install torch torchvision torchaudio
pip install diffusers
pip install transformers
pip install accelerate
pip install sentencepiece
```

4. **Install Hugging Face Hub**

```bash
pip install huggingface-hub
```

For logging in to Hugging Face, you will need to have a Hugging Face account. If you don't have one, you can create one here: [https://huggingface.co/join](https://huggingface.co/join).

5. **Installing git**

```bash
apt install git git-lfs
```

This is required for downloading the model.

## **Choosing AI Models**  

There are many AI models available for image generation, but some of the most popular ones are listed below.

### Popular AI Models for Image Generation

1. **Stable Diffusion** – High-quality text-to-image model.
2. **DeepFloyd IF** – Advanced diffusion model for detailed images.
3. **SDXL** – Improved version of Stable Diffusion for higher resolution.

### Text-to-Image vs Image-to-Image Models

- **Text-to-Image** – Create images from text prompts.
- **Image-to-Image** – Modify or enhance existing images using AI.

### Where to Download Pre-Trained Models

- **Hugging Face** (huggingface.co) – Hosts various AI models.
- **CivitAI** (civitai.com) – Community-driven model sharing.
- **GitHub** – Open-source repositories for AI tools.

### Pre-trained Models

In this guide, we will demonstrate how to use text-to-image models and download pre-trained models from Hugging Face (huggingface.co). Please note that compatibility depends on your hardware specifications, as some models may require more VRAM and RAM than your machine can support.

The pre-trained models we will be testing are:

- [small-stable-diffusion-v0](https://huggingface.co/OFA-Sys/small-stable-diffusion-v0)
- [AWPortrait-FL](https://huggingface.co/Shakker-Labs/AWPortrait-FL)

## **Running AI Models Locally**

### Download the Pre-trained Models

We might need to create a directory where we will download the models.

```bash
mkdir -p /data/ai/huggingface
cd /data/ai/huggingface
```

#### Download small-stable-diffusion-v0

```bash
git-lfs clone https://huggingface.co/OFA-Sys/small-stable-diffusion-v0 OFA-Sys/small-stable-diffusion-v0
```

#### Download AWPortrait-FL

If you have a NVIDIA 3090/4090/5090 GPU, you can try this model.

```bash
git-lfs clone https://huggingface.co/Shakker-Labs/AWPortrait-FL Shakker-Labs/AWPortrait-FL
```

#### Download FLUX.1-dev

To save memory, when using AWPortrait-FL, we will need FLUX.1-dev too. In order to download this model, you will need to login to Hugging Face and accept the license: [https://huggingface.co/black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev). Please refer to the [instruction](https://huggingface.co/docs/hub/en/security-tokens) for access token generation. After you have the access token, then you can login to Hugging Face with the following command:

```bash
# Input your access token when you see the prompt
# you can get it from https://huggingface.co/settings/tokens
huggingface-cli login
# you may see a warning message from git about credentials store
# if you see this warning, you will need to run the following command
# git config --global credential.helper store
# re-login
# huggingface-cli login
```

```bash
git-lfs clone https://huggingface.co/black-forest-labs/FLUX.1-dev black-forest-labs/FLUX.1-dev 
```

Since these models are huge, downloading them might take a while. If you're in Australia, well… grab a coffee, binge a whole TV series, maybe even take a nap, or three. Just start the download, go about your life, and check back tomorrow.

>Why did the Aussie try to download an AI model at home?
>Because they wanted to experience what life was like before the fast internet existed! 🚶‍♂️🐢

### Running AI Models Locally

#### Running small-stable-diffusion-v0

Go the directory where you downloaded the model.
```bash
cd /data/ai/huggingface
```

Create a python script file called `test-small-stable-diffusion-v0.py`:

```python
import torch
from diffusers import StableDiffusionPipeline

model_id = "OFA-Sys/small-stable-diffusion-v0"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")

prompt = "a pig, 4k"
image = pipe(prompt).images[0]  
    
image.save("pig.png")
```

You can use any text (code) editor to edit the script. And if you would like to change the prompt, you can change the text between the double quotes.

Run the script.

```bash
python test-small-stable-diffusion-v0.py
```

Now we have a pig image in `pig.png`.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/how-to-generate-images-with-ai-models-on-your-own-computer/pig.png)


Not very impressive, is it? Because we are using a very small model. But we can try to use a bigger model to generate a better image.

#### Running AWPortrait-FL

If your GPU has enough VRAM, you can try this model.

```bash
cd /data/ai/huggingface
```

Create a python script file called `test-awportrait-fl.py`:

```python
import os 

import torch
from diffusers import FluxPipeline

os.environ['CUDA_VISIBLE_DEVICES'] = '0' 

pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16)
pipe.load_lora_weights('Shakker-Labs/AWPortrait-FL', weight_name='AWPortrait-FL-lora.safetensors')
pipe.fuse_lora(lora_scale=0.9)
pipe.to("cuda")

prompt = "close up portrait, Amidst the interplay of light and shadows in a photography studio,a soft spotlight traces the contours of a face,highlighting a figure clad in a sleek black turtleneck. The garment,hugging the skin with subtle luxury,complements the Caucasian model's understated makeup,embodying minimalist elegance. Behind,a pale gray backdrop extends,its fine texture shimmering subtly in the dim light,artfully balancing the composition and focusing attention on the subject. In a palette of black,gray,and skin tones,simplicity intertwines with profundity,as every detail whispers untold stories."

image = pipe(prompt, 
             num_inference_steps=24, 
             guidance_scale=3.5,
             width=768, height=1024,
            ).images[0]
image.save(f"example.png")
```

```bash
python test-awportrait-fl.py
```

Unfortunately, this model cannot work on my computer because I do not have enough VRAM on my GPU. But you can see the result generated by the model author.

![](https://storage.googleapis.com/www-tyolab-com-blog-assets/images/blog/2025/how-to-generate-images-with-ai-models-on-your-own-computer/2025-01-31-14-39-00.png)

#### Common Issues

A most common issue will be "CUDA out of memory". This is most likely because the VRAM is not enough. You can try to reduce the resolution of the image, or reduce the guidance scale. And if you are trying models that is different from shown above, you may want to try another model to see if it works. If you are testing the "small-stable-diffusion-v0" model and you are encounter this issue, you may want to try a different GPU.

<ContactFooter />








89dc1dcb9a0e519c488347e148ade14258af5dd6059c07c486bbbe227ee251c3

This article will show you how to generate images with AI models on your own computer.

Tag: Image Generation