✦ Aerin's Tensor.Art Guide ✦

A beginner-friendly guide to generating images on Tensor, including image generation styles you can use :3

Welcome!

Hi! I'm Aerin or Aerin___ on J.AI. I've whipped up a small guide on how to use Tensor, so hopefully this can help you out in your bot making!

If you want to skip to image generation styles you can use, click here

But if you've opened Tensor.Art for and thought "what the hell do any of these settings mean?" and you want to learn more then you're in the right place. There are a lot of buttons, models, numbers, sliders, and weird words thrown at you. The good news is that you do not need to understand everything to start making good images.

This guide will go through the important parts of Tensor.Art in simple, beginner-friendly terms, including:

• Models
• Base Models
• LoRAs
• Embeddings
• ControlNet
• VAE
• Aspect Ratio
• Samplers
• Schedulers
• Sampling Steps
• CFG Scale
• Seed
• Advanced Settings
• Clip Skip
• ENSD
• Upscaling
• ADetailer
• Layer Diffusion

....as well as how to prompt!

You don't need to use every feature here. Some of them are completely optional.

✦ 01 — THE MODEL SECTION ✦

The Model is one of the most important parts of your generation. Think of the model as the thing that actually knows how to draw.
Different models have different strengths, styles, and things they understand.

For example, one model might be designed for:

• Anime
• Semi-realistic characters
• Realistic people
• Fantasy art
• Illustration
• Photorealism
• Certain art styles

So if you use the exact same prompt with two completely different models, you can get very different images.

Your prompt tells the AI what you want > The model determines how it interprets that request.

What is a Base Model?

A Base Model is the foundation that your generation is built on.

Think of it like the main recipe.

The base model has learned things like:

• Faces
• Bodies
• Clothing
• Objects
• Environments
• Colors
• Poses
• Art styles
• How words relate to images

There are different model families, such as:

• SD 1.5
• SDXL
• Pony
• Illustrious
• Flux
• And others

These are not interchangeable.

A LoRA made for one model family may not work properly with another. This is why you should always check what Base Model a LoRA or other add-on was made for before using it.

LoRAs

LoRA stands for Low-Rank Adaptation.

A LoRA is basically a small add-on that teaches the model something specific. Think of your Base Model as the main artist. The LoRA is giving that artist an extra skill or instruction.

A LoRA can be made for things such as:

• A specific character
• A specific clothing style
• A pose
• An art style
• A particular aesthetic
• Certain facial features
• Objects
• Hairstyles
• Concepts
• Expressions

LoRA Strength

LoRAs usually have sliders for strength/weight.

This controls how strongly the LoRA affects your image.

For example:

1

0.5 | 0.7 | 1.0 | 1.2

A lower value means the LoRA has less influence. A higher value means the LoRA has more influence. However, higher does not automatically mean better.

Too much LoRA strength can cause:

• Weird anatomy
• Overpowering colors
• Excessive clothing details
• Distorted faces
• The LoRA taking over the entire image

Embeddings

Embeddings are another type of add-on. They are generally much smaller than full models and LoRAs and can help the AI recognize a particular concept or avoid certain unwanted results.

You can think of an embedding as a shortcut. Instead of writing a long collection of words every time, an embedding can represent a specific concept or group of information.

They can be used for things like:

• Improving certain details
• Avoiding common problems
• Improving anatomy
• Negative prompts
• Particular visual concepts

Do I use Embeddings?

Personally, I don't.

You absolutely can use them, but I don't consider them necessary for beginners.

If you're just starting out, I'd recommend learning how the Base Model + Prompt + LoRA + Settings work first. You can always experiment with embeddings later.

ControlNet

ControlNet is used when you want more control over the composition of your image.

Normally, you give the AI a prompt and let it decide how it'll figure out the pose, composition, and structure.

ControlNet lets you give it additional information to follow. For example, you can give ControlNet an image containing a pose and tell it to use that pose.

Some common ControlNet types include:

• OpenPose
  Used primarily to control body pose. Useful when you want your character to stand in a specific way, sit in a particular pose, hold their arms a certain way or copy a reference pose
• Canny
  Uses the outlines and edges of an image. Useful when you want to preserve the basic shape and structure of a reference.
• Depth
  Uses depth information to help preserve the foreground/background structure of an image.
• LineArt
  Uses line art to influence the structure of the generated image.
• Reference / IP-Adapter
  Can help guide the generation using a reference image. This can be useful when you want to influence things like, appearance, composition, style or character features

ControlNet is optional. You do not need it for normal text-to-image generation.

VAE

VAE stands for Variational Autoencoder.

The name sounds much more complicated than it actually needs to be for a beginner.

Very simply: The VAE helps translate the AI's internal image information into the final visible image.

You can think of it as a translator between the model's internal data and the image you actually see.

The wrong VAE can sometimes cause:

• Washed-out colors
• Strange colors
• Muddy details
• Poor contrast
• Other visual problems

What should I use?

If you're unsure, Automatic is usually the easiest option. With Automatic, Tensor.Art can use the appropriate VAE associated with the model/workflow when available. This means you don't have to sit there wondering: Which random VAE am I supposed to pick?

Some models already have their VAE built in, while others may recommend a specific one.

✦ 02 — SETTINGS ✦

Now we're getting into the settings that actually control how the image is generated.

Aspect Ratio

Aspect Ratio is the shape of your image. It describes the relationship between the image's width and height.

For example:

1
2
3
4
5

Square - 1:1

Portrait - 2:3

Landscape - 3:2

Sampling Method

The Sampling Method, also called the Sampler, determines how the AI turns random noise into your final image.

This is one of the most confusing settings when you're new.

Think of it like this:

The AI starts with visual noise > The sampler is the method it uses to gradually turn that noise into an image.

Different samplers can produce slightly different:

• Details
• Textures
• Sharpness
• Composition
• Smoothness
• Overall appearance

Some common samplers you'll see include:

1
2
3
4
5

Euler
Euler a
DPM++ 2M Karras
DPM++ 2M SDE Karras
DPM++ 2S a Karras

Should I turn Sampling Method on?

Yes — if Tensor.Art gives you a Sampling Method option, you need a sampler selected for normal generation.

However, you don't need to constantly change it.

For beginners: Use the sampler recommended by the model first.

If the model doesn't provide one, you can experiment with a common sampler and compare the results.

Scheduler

The Scheduler works alongside the sampler. While the sampler determines how the image is denoised, the scheduler determines how the denoising process is distributed across the steps.

That sounds complicated.

For now, just think:

Sampler = the method
Scheduler = how that method is paced

Common scheduler names include:

• Karras
• Exponential
• Automatic
• Normal
• Simple
• Beta

What should I use?

Again:

Follow the model's recommended setting first.

If you're experimenting, Karras is a very common choice for many Stable Diffusion workflows. But it is not universally correct for every model.

Sampling Steps

Sampling Steps are the number of stages the AI goes through while turning noise into your final image. Think of it like refining a rough sketch.

1
2
3
4
5
6
7

Few steps
↓
Rougher / faster

More steps
↓
More refinement / slower

For example:

1
2
3
4

15 steps
20 steps
30 steps
40 steps

More steps do not mean infinite improvement. Eventually, adding more steps gives you very little benefit while taking longer and using more credits/resources. For many models, something around 20–30 steps is a reasonable starting point, but the ideal number depends on the model and sampler..

If the model says: Recommended: 25 steps

Use 25.

CFG Scale

CFG stands for Classifier-Free Guidance.

In simple terms: CFG controls how strongly the AI tries to follow your prompt.

Think of it like giving the AI instructions.

Lower CFG: The AI has more freedom to interpret your prompt.
Higher CFG: The AI is pushed harder to follow your prompt.

But higher isn't always better.

If CFG is pushed too high, you can start getting:

• Overly harsh details
• Strange colors
• Artifacts
• Unnatural images
• Overcooked-looking results

A common starting range for many SD-style models is around:

1

5–8

but always check the model's recommended CFG first.

Seed

The Seed controls the randomness used to create your image. Think of it like giving the AI a specific starting lottery number.

If you use:

1

Seed: 123456

and keep your other settings the same, you can reproduce a very similar generation.

Random Seed

If you leave the seed empty, it sets at random. You'll get a different starting point each time. This is useful when you're searching for a good composition.

✦ 03 — ADVANCED SETTINGS ✦

Now we're entering the "you probably don't need to touch this" section. If you're a beginner, you can safely leave most advanced settings alone until you understand what they do.

Clip Skip

CLIP is part of the system that helps turn your written prompt into information the image model can understand. Clip Skip tells the system to skip some of CLIP's later processing layers.

In extremely simple terms: Clip Skip changes how your prompt is interpreted before it reaches the image generation process.

A common value is:

1

1

or:

1

2

Some anime-focused models commonly use 2, while other models may use 1 or another value.

The model's recommended Clip Skip matters.

Don't assume that 2 is always better. It isn't. Some models specifically require or recommend a particular value.

My advice is if the model page says:

1

Clip Skip: 2

use 2.

If it says:

1

Clip Skip: 1

use 1.

If you're unsure: check the model's description/recommended settings first.

ENSD

ENSD stands for: Eta Noise Seed Delta

ENSD is an older/technical parameter related to how noise is handled during sampling.

For most modern Tensor.Art workflows: Leave ENSD at its default unless the model or workflow specifically tells you otherwise.

You generally shouldn't change it. If you're following a specific guide that tells you to use a particular ENSD value, then follow that guide. Otherwise, leave it alone.

Upscale

Upscaling means taking your generated image and making it larger and/or more detailed.

For example:

1
2
3

768 × 1152
        ↓
1536 × 2304

An upscaler doesn't just stretch the pixels. It uses another process to reconstruct additional detail while increasing the resolution.

This can help with:

• Higher resolution
• Sharper details
• Better-looking faces
• Cleaner textures
• Preparing images for larger displays

ADetailer

ADetailer is basically a cleanup tool.

It's especially useful for things like:

• Faces
• Eyes
• Hands
• Small details

Sometimes the main generation looks great, but you zoom in and discover weird hands, facial features, etc.

ADetailer can detect certain parts of the image and regenerate/refine them separately.

Face ADetailer

One of the most common uses is a face detector.

It can help clean up:

• Eyes
• Facial features
• Skin details
• Small faces

Hand ADetailer

You can also use an appropriate detector for hands when available.

This can help with:

• Missing fingers
• Weird fingers
• Deformed hands
• Incorrect hand shapes

Layer Diffusion

Layer Diffusion is a more specialized feature.

One of its major uses is generating images with a transparent background / alpha channel.

This can be extremely useful if you're making:

• PNG character cutouts
• Stickers
• Character assets
• Overlays
• Graphics
• Images you want to place over another background

✦ 04 — PROMPTING ✦

A small section for the people who stare at the prompt box for 20 minutes and still don't know what to type. ♡

How to Prompt

A basic guide to building prompts, what to include, how to describe characters/scenes, and how to get more consistent results without writing an entire novel into the prompt box.

Prompt Structure

A simple way to organize your prompt:

subject → appearance → clothing → pose → expression → environment → lighting → composition → style

You don't necessarily need every category every time. Start with the things that matter most to your image and build from there.

What to Describe

Subject: Who or what is in the image?
Appearance: Hair, eyes, skin, body type, distinctive features, etc.
Clothing: Outfit, accessories, materials, colors.
Pose: Standing, sitting, looking away, holding something, etc.
Expression: Smiling, annoyed, tired, neutral, surprised, etc.
Environment: Bedroom, forest, city, fantasy castle, street, etc.
Lighting: Soft lighting, dramatic lighting, backlighting, sunset, neon, etc.
Composition:: Close-up, portrait, full body, from above, side view, etc.

Prompt Weights

Prompt weights let you tell the model which parts of your prompt should receive more or less attention.

A common format is: (keyword:1.2)

The number controls the strength of the word or phrase.

Examples:

(red hair:1.3) → stronger emphasis
(red hair:1.1) → slightly stronger
red hair → normal emphasis
(red hair:0.8) → weaker emphasis

You can also weight multiple words together: (long black hair, crimson eyes:1.2)

General Guide:

0.5–0.8 → decrease emphasis
0.9–1.1 → subtle / normal
1.2–1.4 → noticeable emphasis
1.5+ → strong emphasis and potentially unstable results

Prompting Style Samples

Different image generators interpret prompts differently. There isn't one universal prompt format, so here's a quick comparison of the styles you may commonly see.

Negative Prompting

Negative prompts tell the image generator what you don't want to appear in the image. They're especially common in Stable Diffusion-style workflows, although how useful they are depends heavily on the model and generator. You can find this text box directly below the main prompt text box.

Instead of telling the model what you want:

long black hair, crimson eyes, gothic dress

A negative prompt tells it what you want to avoid:

blurry, low quality, bad anatomy, extra fingers, distorted hands

Negative prompts are most useful for common problems such as anatomy errors, unwanted objects, image artifacts, or visual qualities you don't want.

SAMPLE NEGATIVE PROMPT:

low quality, blurry, pixelated, jpeg artifacts, bad anatomy, bad hands, extra fingers, missing fingers, fused fingers, extra limbs, malformed hands, deformed face, asymmetrical eyes, distorted body, duplicate, multiple people, cropped, out of frame, text, watermark, logo

Tips & Notes

• Don't overstuff your prompt. More words don't automatically mean more detail.
• Be specific about important features. If something really matters to the character, describe it clearly rather than hoping the model understands what you mean.
• Keep your wording consistent when comparing models or LoRAs. Changing the prompt between generations makes it harder to tell whether the difference came from the model/LoRA or the prompt itself.
• Weights aren't magic. If something isn't appearing correctly, increasing the weight over and over may not fix the underlying problem.
• Negative prompts can help, but don't rely on them for everything. Sometimes the better solution is simply describing what you actually want more clearly.
• Different models understand prompts differently. A prompt that works well with one model may produce completely different results with another.
• LoRAs can change the way your prompt behaves. Some are sensitive to their trigger words, recommended weights, or specific prompting styles.

✦ QUICK REFERENCE ✦

🡻 CLICK ME 🡻

✦ IMAGE GENERATION STYLES ✦

🢁 CLICK ME 🢁

Hopefully this helped make Tensor a little less confusing!

If you found this guide useful, consider checking out my bots on JanitorAI. I make a nice mix of characters, so there's probably something lurking around that you'll like.

→ ✦ My JanitorAI Profile

Got questions? You can find me on RevoSpring:

→ ✦ RevoSpring

Thanks for reading, and happy generating! ♡

Made by Aerin

Back to Top