The Three Starting Points: How I Decide Where to Begin
Three Doors, Same Building
In the last tutorial, I showed you the big picture — three starting points, a flexible middle, one finish. Now let's get into the starting points themselves.
Every image I make enters the pipeline through one of three doors. The door I choose depends on what I'm working with and what I want out of it.
This isn't a step-by-step guide. These are three separate approaches. You pick one. Which one depends on your source material and your goal.
Starting from a Reference Image
This is my most common starting point. I find an image I like — could be a photo, a screenshot, fan art, a painting, anything — and I use it as the visual foundation.
But I don't just throw the raw image into img2img. There's a step before that.
The Crop
First, I crop the reference image to the exact aspect ratio I want the final output to be. I use a custom cropping tool that locks the aspect ratio so every image comes out at a consistent size. I pick the part of the image I want to base the generation on — maybe it's the face and upper body, maybe it's the full pose, maybe it's just a background composition I liked.
This matters more than people think. The crop determines what the AI focuses on. A tight crop of a face produces a completely different result than a full-body crop of the same image.
Here I'm pulling one character out of a larger scene. The cyan box is locked to my target aspect ratio. That crop is what goes into the next step — everything outside the box is gone.
The Tags
After cropping, I run the image through a tagger — an AI model that looks at the image and extracts tags describing everything it sees. Pose, clothing, lighting, colors, expression, background, style.
Here's where it gets interesting: I have multiple taggers and I pick different ones depending on what I want.
- If I want a faithful interpretation of the reference, I use a tagger that pulls out a lot of detail — every element of the clothing, the exact lighting setup, the specific pose description.
- If I want loose variations, I use a tagger that gives me broader, less specific tags. This gives the AI more room to interpret and creates more variety.
- If I'm working with a realistic photo, I go heavy on the details. Real photos have a lot of nuance — skin texture, fabric type, specific lighting angles — and I want to capture all of that.
- If it's anime or a painting, I go lighter. These images have less detail to extract, and over-tagging them just adds noise.
The tagger choice is one of the biggest decisions in the whole pipeline. Same image, different tagger, completely different output.
Here's what it looks like when I run my tagger on that Nikke crop from earlier. Five different classification models analyze the image, then Florence-2 generates detailed captions in multiple passes. The result: 69 tags plus a full description. All of that becomes the raw material for the prompt.
The Enhancement Pass
Raw tagger output is a starting point, not a finished prompt. The tags are accurate but flat — they describe what's in the image without any atmosphere, style, or mood.
Before I generate anything, I run the tags through an enhancement pass. I have a library of prompt upgrade techniques — each one rewrites the prompt to improve a specific quality. Cinematic lighting. Dynamic composition. Material textures. Color palette. Atmosphere.
I pick whichever ones fit the image I'm going for and run them in sequence. The order matters — I usually go lighting first, then composition, then trim any duplicate or redundant tags that crept in.
Here's a real example. This is the raw tagger output for that Nikke crop:
"...a young woman with short white hair and purple eyes... standing in the middle of the frame with her arms outstretched... the lighting is bright and dramatic... multiple_girls, weapon, pantyhose, gloves, short_hair, cape..."
And here's what it looks like after the enhancement pass:
"...(young woman) with (platinum silver hair) and (deep violet eyes)... (volumetric amber god rays) streaming through (ochre dust clouds), (dramatic copper rim lighting) on silhouettes, (molten tangerine backlight) illuminating (airborne particles)... ((charcoal_pantyhose)), ((ivory_gloves)), ((platinum_hair)), ((burgundy_cape))..."
Same image. Same tags. But the enhanced version has specific color names instead of generic ones, atmospheric lighting instead of "bright and dramatic," and parentheses weighting on the elements that matter most.
Same concept. Completely different image.
I break down the full enhancement toolkit in a paid tutorial for members — but the takeaway here is that there's a step between tagging and generating that most people skip entirely. The tagger gives you the skeleton. The enhancement pass gives it a soul.
What Happens Next
The enhanced tags go into my txt2img pipeline to generate variations. Or the cropped reference goes directly into img2img. Or both — I might do txt2img variations first, then pick the best ones and run those through img2img.
This is what I meant by "the middle is a toolbox." After the start, there are multiple paths forward.
Starting from Extracted Tags
Sometimes I don't want the image itself — I want its DNA.
This approach is similar to the reference image approach, but I skip the img2img entirely. I extract the tags and use them purely as a txt2img prompt.
Why Not Just Use the Image Directly?
Because img2img gives you something that looks like the original. Tag extraction gives you something that feels like the original but looks completely different.
The AI isn't seeing the image. It's seeing words like "standing pose, white dress, dramatic lighting, wind-blown hair, sunset background." Those same words can produce a thousand different images. The DNA is the same. The expression is unique every time.
My Process
Same as before — crop, then tag. But instead of feeding the image into img2img, I take the tags and drop them into txt2img. From there I usually generate a batch of variations and cherry-pick the best compositions.
I use this approach most often when I like the concept of a reference image but not the execution. Maybe the composition is great but the style is wrong. Or the pose is perfect but I want a completely different character. Tags let me keep what works and rebuild the rest.
Starting from a Prompt
This is what most people think I do all day. And sometimes I do — but it's actually the least common starting point for my best work.
Pure Imagination
Sometimes an idea just hits and I want to build it from scratch. No reference. Just a concept in my head, translated into tags and descriptions.
Borrowed Prompts
Other times I'll grab the positive prompt and tags from someone else's image — on Civitai, or anywhere the generation data is available. But I always modify them. I'm not trying to recreate their image. I'm using their tags as a launching pad.
Maybe they nailed a lighting description I hadn't thought of. Maybe they used a tag combination that's interesting. I'll take that, strip out what I don't want, add my own direction, and run it through my pipeline.
The key difference between how I use someone else's prompt and how most people use it: they paste and generate. I disassemble, modify, and rebuild.
When I Choose This Path
Honestly? When I'm feeling creative and want to experiment. Or when I see someone else's work and think "I wonder what would happen if I took that concept in a different direction."
This is the most unpredictable starting point. Sometimes it produces my best work. Sometimes it goes nowhere. That's the trade-off of starting without a visual anchor.
How I Decide Which Door to Walk Through
There's no formula. But here's the general logic:
Realistic photo reference? Start from the reference image. Go heavy on the tags. I want to capture all the detail — the fabric textures, the specific lighting, the subtle color grading. Real photos have information that tags alone can't fully represent, so img2img helps carry that visual information through.
Anime or illustration reference? Could go either way. If I like the composition and pose, I'll use the image directly. If I just like the concept or vibe, I'll extract tags and go txt2img. Anime images have less fine detail to extract, so the tags are usually enough.
Painting or artistic reference? Almost always tag extraction → txt2img. Paintings have a style that img2img tends to flatten. Better to extract the concept and let the AI rebuild it in a different style.
Just feeling creative? Prompt from scratch. Or grab something interesting from Civitai and remix it.
Want maximum variety? Tag extraction with a loose tagger → batch txt2img. This produces the widest range of results from a single starting point.
Want a specific result? Reference image → detailed tagger → img2img. This is the most controlled path.
What Comes After the Start
No matter which door you walk through, you end up in the same place: the middle of the pipeline. That's where batch generation, recipe systems, cherry-picking, and refinement happen.
The starting point determines the raw material. The middle determines what you build with it. And the finish — the img2img passes — is what polishes it into a final piece.
The middle and the finish are where the real depth is. I'll be breaking those down in upcoming tutorials for Prompt Insider and Full Workshop members.
Try It Yourself
Pick one approach and try it this week:
If you've never used img2img: Find an image you love, crop it to your target aspect ratio, and run it through img2img with a simple prompt. See how different the result is from txt2img alone.
If you've never used tag extraction: Find a tagger tool (there are free ones built into A1111 extensions), run it on an image you like, and use the output as a txt2img prompt. You'll be surprised.
If you always write prompts from scratch: Go grab a prompt from an image you admire on Civitai. Don't use it as-is. Take it apart, understand what each piece does, modify it, and make it yours.