AI Empire
Open menu

Free LoRA dataset planner

A balanced 50-image character dataset is 17 close-up or upper-body shots, 17 cowboy or three-quarter shots and 16 full-body shots, spread over seven angles and five expressions. Set your image count and trigger word, and this planner writes the shot list, a checklist and one caption per image.

Why 20 to 50 images: how many images you need for a LoRA. Why train one at all: how to keep an AI character consistent. The split below is a starting plan, not a rule.

By Ai Emperor. Updated . Runs in your browser; nothing is sent anywhere.

Plan your dataset

20 to 80. We train on 50; 20 to 50 is the usual range.

Lean towards what you post. Every focus still mixes all five framings.

Starts every caption. What’s a trigger word?

50 images: 17 close-up or upper body, 17 cowboy or three-quarter, 16 full body. 20 outfits, 16 places.

Framing split
FramingImages
Close-upHead and shoulders. Teaches the face in detail.9
Upper bodyWaist up. Most selfies look like this.8
Cowboy shotMid-thigh up.9
Three-quarter bodyKnees up.8
Full bodyHead to feet. The gap most datasets have.16
Angle split
AngleImages
Front“facing the camera”15
Three-quarter left“looking slightly to the left”8
Three-quarter right“looking slightly to the right”7
Looking up“looking up”5
Looking down“looking down”5
Left profile“left side profile”5
Right profile“right side profile”5
Expression split
ExpressionImages
Neutral“neutral expression”15
Soft smile“soft smile”13
Smiling“smiling”10
Serious“serious expression”7
Laughing“laughing”5

Outfits, places and light

  • 20 outfits, none in more than 3 images.
  • 16 places: 26 images indoors, 18 outdoors by day, 6 at night.
  • Light: overcast daylight 6, golden hour 6, bright sunlight 6, soft window light 12, warm indoor light 14, night with neon lights 4, phone flash at night 2.

Checklist

  • Make 50 images, then delete any that isn’t clearly her. A clean set of 30 beats a messy set of 80.
  • Same face in every image. Only the angle, expression, outfit, place, light and framing change.
  • At least 16 full-body images. Skipping them is the most common gap.
  • No outfit in more than 3 images, so no outfit becomes part of her.
  • Indoors and out, day and night: 26 indoor, 18 outdoor by day, 6 at night.
  • Every image sharp and at least 1024 pixels on its short side.
  • One .txt caption per image with the same file name: image_01.png and image_01.txt.
  • The trigger word “zvx woman” spelled exactly the same in every caption.
  • No caption mentions her face, hair, eyes or body.

Captions, one per image

Order: trigger word, shot and pose, expression, outfit, place, light, framing last. Close-ups name the shot first instead. Never her face or hair.

  1. image_01.txtzvx woman, close-up portrait facing the camera, serious expression, oversized grey hoodie, in a cozy cafe, soft window light
  2. image_02.txtzvx woman, close-up portrait looking slightly to the left, neutral expression, navy blazer over a white top, in a modern kitchen, warm indoor light
  3. image_03.txtzvx woman, close-up portrait looking slightly to the right, laughing, black puffer jacket, outside a restaurant at night, night with neon lights
  4. image_04.txtzvx woman, close-up portrait facing the camera, soft smile, white linen button-up shirt, on a quiet suburban street, overcast daylight
  5. image_05.txtzvx woman, close-up portrait looking up, smiling, denim jacket over a floral sundress, on a city street at night, night with neon lights
  6. image_06.txtzvx woman, close-up portrait looking down, neutral expression, light blue oxford shirt, in a living room with a beige sofa, warm indoor light
  7. image_07.txtzvx woman, close-up portrait in left side profile, soft smile, black leather jacket over a white t-shirt, in a train station, soft window light
  8. image_08.txtzvx woman, close-up portrait in right side profile, serious expression, black turtleneck, on a beach boardwalk, golden hour
  9. image_09.txtzvx woman, close-up portrait facing the camera, neutral expression, emerald green midi dress, on a rooftop terrace with the skyline behind, bright sunlight
  10. image_10.txtzvx woman, looking slightly to the left, smiling, beige trench coat, in an airport lounge, warm indoor light, upper body
  11. image_11.txtzvx woman, looking slightly to the right, soft smile, white tank top, in a bright bedroom with plants, soft window light, upper body
  12. image_12.txtzvx woman, facing the camera, neutral expression, cropped white t-shirt, in a steel elevator, warm indoor light, upper body
  13. image_13.txtzvx woman, facing the camera, laughing, cream cable-knit sweater, at a farmers market, overcast daylight, upper body
  14. image_14.txtzvx woman, looking up, soft smile, yellow linen summer dress, on a city sidewalk, golden hour, upper body
  15. image_15.txtzvx woman, looking down, smiling, brown suede jacket over a cream blouse, in a park with autumn trees, bright sunlight, upper body
  16. image_16.txtzvx woman, left side profile, serious expression, sage green sports bra and zip jacket, in a modern gym, warm indoor light, upper body
  17. image_17.txtzvx woman, right side profile, neutral expression, striped breton top, in a cozy cafe, warm indoor light, upper body
  18. image_18.txtzvx woman, looking slightly to the left, soft smile, pastel pink knit sweater and white skirt, in a modern kitchen, soft window light, cowboy shot
  19. image_19.txtzvx woman, looking slightly to the right, neutral expression, red satin slip dress, outside a restaurant at night, phone flash at night, cowboy shot
  20. image_20.txtzvx woman, facing the camera, smiling, burgundy cardigan and black jeans, on a quiet suburban street, bright sunlight, cowboy shot
  21. image_21.txtzvx woman, facing the camera, neutral expression, oversized grey hoodie and black bike shorts, on a city street at night, phone flash at night, cowboy shot
  22. image_22.txtzvx woman, looking slightly to the left, soft smile, navy blazer over a white top and black trousers, in a living room with a beige sofa, soft window light, cowboy shot
  23. image_23.txtzvx woman, facing the camera, serious expression, black puffer jacket and black leggings, in a train station, warm indoor light, cowboy shot
  24. image_24.txtzvx woman, looking slightly to the right, laughing, white linen button-up shirt and light blue jeans, on a beach boardwalk, overcast daylight, cowboy shot
  25. image_25.txtzvx woman, looking up, smiling, denim jacket over a floral sundress, on a rooftop terrace with the skyline behind, golden hour, cowboy shot
  26. image_26.txtzvx woman, looking down, neutral expression, light blue oxford shirt and khaki shorts, in an airport lounge, soft window light, cowboy shot
  27. image_27.txtzvx woman, left side profile, soft smile, black leather jacket over a white t-shirt and black jeans, in a bright bedroom with plants, warm indoor light, three-quarter body
  28. image_28.txtzvx woman, right side profile, neutral expression, black turtleneck and grey wool skirt, in a steel elevator, soft window light, three-quarter body
  29. image_29.txtzvx woman, looking slightly to the left, serious expression, emerald green midi dress, at a farmers market, bright sunlight, three-quarter body
  30. image_30.txtzvx woman, facing the camera, smiling, beige trench coat and dark jeans, on a city sidewalk, overcast daylight, three-quarter body
  31. image_31.txtzvx woman, facing the camera, soft smile, white tank top and olive cargo pants, in a park with autumn trees, golden hour, three-quarter body
  32. image_32.txtzvx woman, looking slightly to the right, neutral expression, cropped white t-shirt and grey sweatpants, in a modern gym, soft window light, three-quarter body
  33. image_33.txtzvx woman, looking slightly to the left, laughing, cream cable-knit sweater and brown corduroy skirt, in a cozy cafe, soft window light, three-quarter body
  34. image_34.txtzvx woman, facing the camera, soft smile, yellow linen summer dress, in a modern kitchen, warm indoor light, three-quarter body
  35. image_35.txtzvx woman, looking up, smiling, brown suede jacket over a cream blouse and blue jeans, outside a restaurant at night, night with neon lights, full body
  36. image_36.txtzvx woman, looking down, neutral expression, sage green sports bra and zip jacket with matching leggings, on a quiet suburban street, golden hour, full body
  37. image_37.txtzvx woman, left side profile, serious expression, striped breton top and white jeans, on a city street at night, night with neon lights, full body
  38. image_38.txtzvx woman, right side profile, soft smile, pastel pink knit sweater and white skirt, in a living room with a beige sofa, warm indoor light, full body
  39. image_39.txtzvx woman, facing the camera, neutral expression, red satin slip dress, in a train station, soft window light, full body
  40. image_40.txtzvx woman, looking slightly to the right, smiling, burgundy cardigan and black jeans, on a beach boardwalk, bright sunlight, full body
  41. image_41.txtzvx woman, looking slightly to the left, soft smile, oversized grey hoodie and black bike shorts, on a rooftop terrace with the skyline behind, overcast daylight, full body
  42. image_42.txtzvx woman, facing the camera, neutral expression, navy blazer over a white top and black trousers, in an airport lounge, warm indoor light, full body
  43. image_43.txtzvx woman, facing the camera, laughing, black puffer jacket and black leggings, in a bright bedroom with plants, soft window light, full body
  44. image_44.txtzvx woman, looking up, serious expression, white linen button-up shirt and light blue jeans, in a steel elevator, warm indoor light, full body
  45. image_45.txtzvx woman, looking down, smiling, denim jacket over a floral sundress, at a farmers market, golden hour, full body
  46. image_46.txtzvx woman, left side profile, soft smile, light blue oxford shirt and khaki shorts, on a city sidewalk, bright sunlight, full body
  47. image_47.txtzvx woman, right side profile, neutral expression, black leather jacket over a white t-shirt and black jeans, in a park with autumn trees, overcast daylight, full body
  48. image_48.txtzvx woman, looking slightly to the right, neutral expression, black turtleneck and grey wool skirt, in a modern gym, warm indoor light, full body
  49. image_49.txtzvx woman, looking slightly to the left, soft smile, emerald green midi dress, in a cozy cafe, warm indoor light, full body
  50. image_50.txtzvx woman, facing the camera, smiling, beige trench coat and dark jeans, in a modern kitchen, soft window light, full body

How does the planner split a 50-image dataset?

Into three equal groups by framing, then across angles and expressions. Fifty images become 17 close-up or upper body, 17 cowboy or three-quarter body and 16 full body. Angles lean on the front view, expressions lean on neutral and a soft smile, and 20outfits rotate so none appears more than 3 times.

FramingImages (of 50)What it shows
Close-up9Head and shoulders. Teaches the face in detail.
Upper body8Waist up. Most selfies look like this.
Cowboy shot9Mid-thigh up.
Three-quarter body8Knees up.
Full body16Head to feet. The gap most datasets have.

The 1/3 split is the same mix the course uses for prompts. Full body gets a whole third on its own because it’s the shot most datasets skip, and a LoRA that rarely saw her full body invents one.

Which angles should a character LoRA dataset cover?

Seven: the front view plus the six angles the course prompts an edit model for. Those are looking slightly left and slightly right with the head tilted, looking up, looking down, and the left and right side profiles. The front gets the most images, because that’s how most of her photos will look.

AngleImages (of 50)Caption words
Front15facing the camera
Three-quarter left8looking slightly to the left
Three-quarter right7looking slightly to the right
Looking up5looking up
Looking down5looking down
Left profile5left side profile
Right profile5right side profile

The course’s first angle prompt, for an image-editing model, reads:

subject must be looking slightly to the left with her head slightly tilted to the left

The other five swap in “to the right”, “looking down”, “looking up”, “left side profile” and “right side profile”.

Blonde AI character, front view, neutral expression.Blonde AI character, three-quarter view to the left.Blonde AI character, left side profile.Blonde AI character, facing the camera, eyes looking to the side.Blonde AI character, front view, smiling.Blonde AI character, front view, soft expression.
One face from several angles and expressions: front, three-quarter, profile, eyes aside, smiling. All AI-generated.

How should you caption each image?

With one line in a fixed order: trigger word, shot and pose, expression, outfit, place, light, and the framing last. That’s the order our Dataset Maker writes. Save it as a .txt file with the same name as the image. Leave her face and hair out of every caption, so the trainer ties them to the trigger word instead of treating them as details that can change.

Image 10 of the default plan, saved as image_10.txt next to image_10.png:

zvx woman, looking slightly to the left, smiling, beige trench coat, in an airport lounge, warm indoor light, upper body

Use your own trigger word in place of zvx woman, spelled the same in every caption. The planner only knows where she looks, so add her pose to that first part once you’ve made the image: “sitting on a bench looking up”. Close-ups name the shot first (“close-up portrait looking up”) and need no framing word at the end. Close-ups and upper-body shots name only the top she wears, because the rest of the outfit isn’t in the frame. Copy all captions at once, or download the plan and paste each line into its file.

When should you change the split?

When your feed leans one way. If you mostly post full-body outfit photos, pick “More full body” and 40% of the set becomes full body. If you post selfies, pick “More close-ups”. Every focus keeps at least 30% in each group, because a LoRA that never saw a framing has to guess at it.

The plan doesn’t depend on the base model; training settings and results do. Whatever you train on, test a few checkpoints with the same prompts before you commit to one: how to test a LoRA has the method we use.

How do you make the images in the plan?

Start from one clear, front-facing face of your original AI character, then use an image-editing model to make each planned shot of that same face. Our public AI Empire Dataset Maker does it in one run: one face photo in, 50 images with a caption file each out, zipped.

Its captions use the same format as this planner. If you make the images another way, use the plan as your shot list and the six angle prompts above. Only ever use images of your own AI character, never photos of a real person. A rented GPU is enough for all of it; our RunPod guide covers which one to pick.

What can go wrong?

  • A second face sneaks in. One or two images where she looks slightly different teach the LoRA a blend. Delete them, even if it leaves you short of the plan.
  • Captions mention her hair or face. The LoRA treats them as changeable and they drift. Strip them from every caption.
  • One outfit in half the set. The LoRA learns it as part of her. Keep every outfit to a few images.
  • Too few full-body shots. Close-ups look right and full-body shots show a different body. Add full-body images and retrain.
  • File names don’t match. image_01.png with img_01.txt trains that image with no caption.
  • A typo in the trigger word. One caption with zxv woman teaches a second word. Copy the captions instead of typing them.

Questions people ask

How many images should a character LoRA dataset have?

20 to 50. We train on 50. Below 20, the face tends to drift in angles the LoRA never saw; past 50 you mostly add training time and the risk of an off image. The planner goes up to 80 if you want room to cut. More in how many images you need for a LoRA.

What should a LoRA dataset include?

One face in every image, and variety in everything else: five framings, seven angles, several expressions, many outfits, indoor and outdoor places, and different light. The planner splits all of these for the image count you pick.

Should the captions describe her hair or face?

No. Whatever the captions leave out, the trainer ties to the trigger word. Leave out her face and hair and they become part of her. Describe them and the LoRA treats them as details that can change, so they drift.

Does the plan depend on the base model?

No. The plan is about what the images show, which matters for any base model. Training settings do depend on the model, and so do results, so test a few checkpoints before you commit to one.

Is the planner free, and does it send my data anywhere?

It’s free and needs no account. It runs in your browser and sends nothing anywhere. The download is a .txt file made on your own device.

Read next