These are the Reference photos, examples, and outputs for my AI Starter Guide on the forum.

https://forums.e-hentai.org/index.php?showtopic=294629

5/11/2026
Added simple overviews of the other model types.

I made an error in the original post and called one of the tagging methods story telling, which is not correct that style is actually called Novel Prompting. The animated images after the I2V examples are the output of the prompt for both Novel and Tagging style prompts.

Page 26-29 contain the ESRGAN Model & Simple Upscale Method being used With RIFE Frame Interp. Doubling their native framerate of 16 to 32 and quality by 2X & 4X. Its subtle but the animations have 2x the frames of the originals.

5/12/26
Fixed the webp files not looping infinitely.

removed duplicates and replaced the ESRGAN photos with higher quality variants and some post processing.

I have seen a number of comments on posts asking how to get started with AI. This is a simple starter guide, AI is pretty simple to start if you want safety and access to both local and cloud rendering. ComfyUI or Automatic1111 are your best bet. ComfyUI is significantly more complicated and requires some basic knowledge of node coding and Stable Diffusion processes. To start generating anything you need a few things.

1:Identify whether your hardware is up to the task of generative AI.

To generate images from T2I (Text to Image) or I2I(Image to Image) It requires a minimum spec of a 4GB GPU. GPU power is important for generating high quality images, more VRAM, more quality. I have a 3090 so my card has 24GB VRAM, I can generate about 50 images per 2 minutes: at a Batch size of 5@1024x1024 every 12 seconds for a count of 10, 10 sets of 5, 5 times per minute = 50 images over 2 minutes. With a 12 GB card you could do maybe 35. An 8 maybe 15 and a 4 GB card about 5-6. More VRAM: More Speed & More Quality. It's about finding a balance between processing power for image generation and maximum vram for batches. AI also does support crossfire and other MultiGPU set ups.

2: Choose your preferred software: Automatic1111 & ComfyUI have their benefits and drawbacks. Both programs are free as of 4/10/2026.

Auto1111 is completely open source and runs fairly well. It is a web based GUI. Meaning that the software is controlled through a webpage. The webpage is just a GUI it does not store or send any data to your web browser. It's also compatible with all web browsers. You can turn your rig into an AI website if you desired with auto1111. Some people dedicate a percentage of their GPU and loan it out to AI data servers in exchange for money. Auto1111 is the primary generation program for most local Generative AI & has 1000's of free addons to do incredible things like control nets, upscaling, gif creation, image editing, inpainting, multi prompt manager, etc. That make your generations & Pipelines more varied. However A1111 is unfortunately to my knowledge a Local Generation Only. I am not certain if there's an API for a backend. A1111 is compatible with most general Stable diffusion models and checkpoints. except FLUX. However It is severely lacking in other areas like RIFE, ESRGAN, or WAN support.

ComfyUI is a managed free program that is quite a bit more advanced and complicated than A1111. The biggest benefit of Comfy is compatibility and Fine Tuning. Comfy is a NODE system. So you have to have at least a basic understanding of node coding. It is simple to learn however, just intimidating the first time you open the editor. Comfy Has compatibility with all forms of generation and all diffusion models. It also supports Image 2 3D, I2V, T2V, I2I, T2I, and dozens of other programs and workflows. For beginners you can download free templates of common setups for video generation, 3d model creation, Image Creation, and image Editing. Comfy has both local and cloud generation support. You can make your own templates and run the generation parameters you've set and generate it on a token system provided by a few sources. I'm not 100% certain how the cloud generation works all my generations are done locally, so i don't use it. The cloud system obviously does usually require a subscription though. Comfy does also support many different addons as well, most are free as well.

Both have their benefits and drawbacks. Just depends on you and your situation.

3: Generation Models and basics

Generation models and checkpoint are required for all generation, The most common types are. SD1.5, SDXL, Illustrious, Pony, FLUX, WAN, ESRGAN, and RIFE. Each has their own benefit and drawbacks.

SD1.5 are the most commonly available models. They are very powerful and are fantastic for both realistic and Cartoony. SD1.5 also has thousands of LORA's, and every ControlNet model. SD1.5 is fantastic for beginners, Light weight, and has simple prompting.

SDXL is just SD1.5 with a few more extras. The generation quality can differ in very specific ways. There isn't very many SDXL's in comparison to SD1.5.

Illustrious is fantastic for cartoon and artistic works and is a derivative of the PONY models. Illustrious is compatible with both Illustrious and Pony LORAs

PONY is a relatively new model that is an awkward cousin of the FLUX models. Pony is an Automatic1111 compatible version of the FLUX models. Pony has the potential to create some truly fantastic image quality and dynamic lighting. However PONY models have a tendency to be extremely heavy and slow. PONY and FLUX are not recommended for weaker hardware.

FLUX models are the most commonly generated types of images you see lately. It has a very specific generative style. If you've seen the ones that are cartoony but also highly realistic and shiny. Like their covered in oil. That's a signature of flux generation. This Gallery was likely generated on either a FLUX or PONY Model.

WAN is an image and text to video model. It allows you to input an image and then uses the prompt window to describe the image and the actions being performed. WAN models are very computationally expensive and require both high GPU VRAM but also high CPU processing power. 24GB VRAM and at least 6 cores of 3.4 GHz. WAN is capable of both realistic and artistic generations.

ESRGAN is an image and video upscaling tool, it functions the same as an image to image generator. it uses the source image to try and subdivide the pixels and fill in the gaps with new details based around the existing details. It is functional and is capable of some incredible things however the hallucinations get worse the lower the quality the input is.

RIFE is a relatively new model that is a video editing tool for doubling, quadrupling, and sextupling the frame count of animation and video. It is actually a fairly lightweight program and can run on lower end systems fairly well. The doubling quality is actually exceptional for animated media, the realistic media is shockingly accurate. I've used it to double the frame rates of a few fight scenes from naruto with some pretty crazy results. The Pain fight had somewhat decent frames instead of the stretchy mess it is. I did it with a couple Jackie Chan movie scenes to some hilarious results. sometimes its awesome and sometimes it botches the split frames and creates some pretty wild things. It could be really good for Psychedelic animations.

LORAs are a special type of checkpoint that can be called during prompting. LORA's are really specific details that can be added to steer the generation in particular directions. My most common LORA's I use are for, certain styles, certain locations, certain vagina shapes, dick sizes, breast size, clothing types, clothing brands, etc.

Hypernetworks and Textual Inversions are a topic you will likely encounter, but they are largely deprecated and useless to the average user. They are primarily targeted at people who want to create their own hybrid checkpoints and LORAs. They are processor intensive, time consuming, and very slow to use and train.

LORA's created through KohyaSS are superior in quality.

ControlNets are an advanced tool to assist in image generation broadening your generation capabilities. allowing for custom poses, depth mapping, image manipulation and dozens of other tools that can be programed and applied at any point in the generation process

Multi-Prompt and Wildcards are exceptional addons. They allow you create txt files containing lists of random topics, colors, animal types, clothing types, locations, orientations, sexual positions and kinks. Then you can load multiple of these wildcards to procedurally select random prompts based on the selected parameters. This gallery has likely been generated using a pose wildcard or an OpenPose control net.

4: General prompting and glossary of terms

Batch is how many images are generated at the same time per generation.

Count is how many times to regenerate this batch of generations.

Sampling steps is how many times the model is applied on the image, more sampling steps means more passes and quality depending on the sampling method and schedule type.

Sampling Method is how the image is sampled. The most common types of sampling are:

DPM++2M: General Purpose sampler that is good for all types of generation, is moderately lower, but has a higher range of generation. It can also be used at absurdly high sampling steps with no drawbacks.

EULER Ancestral: A very powerful sampler capable of rapidly creating artistic images and is specifically created for low sample and lower end hardware. It can produce extremely high quality images even at low samples.

Schedule Type is a more advanced sampling modifier that changes how the sampling method is applied. Automatic, Karras, and simple are the only types i use consistently.

HiRezFix is for images at qualities above 1024x1024. The image has to be generated 2x otherwise the high qualities will cause extreme hallucinations and loss of quality. Hires fix also uses double the processing power of a 1024, because it upscales the image to 2048x2048 using ESRGAN

Refiners allow you to hot swap between models mid generation to do some pretty interesting things. I have very limited experience using this tool.

Then Width and Height are self explanatory.

CFG Scale is related to prompt exactness, higher number means more specific prompts but less creativity and higher possibility of artifacting, lower numbers are usually preferable. 4-7 is a good starting range.

**There is a mod called CFG Fix which allows higher numbers without loss of quality.

Positive PROMPT Every checkpoint has specific prompting etiquette related to its type. For example PONY models use a scoring system for image quality. Score_10_up, Score_9_up, Score_8_Up, etc. Masterpiece is also used heavily. Prompting is as much an art as a skill, it requires a solid linguistic eloquence. All words have multiple different meanings, and understanding those differences can wildly change the generation. Your prompt can be done in one of two ways, Tagging or Story Telling.

Tagging is relatively simple: female, Furry:1, Judy Hopps:1, lingerie:1.1, lace, medium breasts, long ears, Rabbit, thigh highs, purple eyes, garter belt, thong, black lingerie, sexy, fluffy tail, short fur, shiny skin, multicolored skin, Camel toe

Story Telling is pretty simple: Disney's Zootopia character Judy Hopps, Wearing sexy black lingerie, she is posing with a white background, the lingerie is transparent and connected to her thigh high socks with garter belts. Her eyes are purple and wearing dark purple eye liner.

Negative prompting is primarily done in tagging format. Text, Watermark, Signature

Weights The:1 is what's called a weight any tag or block followed by a : and a number between 0.1-2 will add a weight to that tag, the generation will contain more of that tag based on its weight, 0.1-1.0 is fairly low weight but will still be added to the image. 1-2 is high weight The generation will definitely have this and ensure it matches as closely as possible.

Upscaling is common there are two types basic and AI generated. A basic upscale uses one of several different methods to essentially subdivide the pixel density to increase the quality at large scale without affecting the current quality. Basic upscales use algorithms like LANCZOS

AI Generated like ESRGAN Adds details but can sometimes produce lower quality results. The most common model used is R-ESRGAN 4X+ Anime6B

That should be a good basic starter guide.

Last edited on 11 May 2026, 10:30.

Base +5, unsn +6, Mr. Xecret +7, sarusa99 +6, Previous -7, glyphtrail +2, Maximus Stallion -6