One anchor, one face: training Odile on Krea 2

A downloaded character LoRA is a trigger and the weights. Nobody gets my anchor photo or my edit workflow, so the test I care about is whether the face survives on text alone.

On Krea 2 a character LoRA pays twice. The base alone won't hold a person from a reference image, so the LoRA ends up being a tool I use myself, on top of being the thing I publish.

Where she came from

Odile started as one image: a close-up at a bar, Krea 2 Turbo, seed 22. Everything else she is came out of that same session. I pushed the anchor into new scenes, poses and outfits with conradlocke's Identity Edit LoRA for Krea 2, version 1.2 (the v2 is non-commercial without the author's license), always starting from the same face. Every image in her dataset descends from that one face.

The contrast: Arak on Klein 4B

Before Odile I spent three rounds on Arak, a four-eyed creature character, on Flux.2 Klein 4B, 500 Buzz a round.

• v1: the captions described her skin and eyes every time, and the trigger learned "pale woman".

• v2: 30 images from her sheet and older work, captions describing only what varies. The four eyes and the skin came from the trigger alone in 9 of 10 tests. The face changed from image to image.

• v3: 21 new images generated from her references, plus 15 older ones with high face scores. The trainer samples looked closer to her. On the distilled Klein 4B, same seed, v3 came out nearly identical to v2.

Changing the dataset didn't move the face. With text only, the 4B LoRA learned the creature markers and left the face to the base model, whatever I fed it. Arak holds when the model gets her reference and the LoRA together, and a download doesn't give anyone that. She's paused.

Her v2 also mixed images from different engines and phases, and its face changed from image to image. Odile's dataset came from one source.

One anchor, 31 images

What I cut by hand from her session:

• the mirror shot, because the reflection flips the beauty mark

• a beach shot with her face covered

• a set from an older phase where her hair went all blue

• the bookstore and piano shots, for broken hands

• anything with a second person in it

The 33 candidates left went through a Gemini vision prompt that scores each feature against the anchor from 0 to 10, where 7 means present and correct. Hair, streak and facial features came back 7 on every one. The low scores were the beauty mark alone, which disappears in wide shots anyway. 31 made the zip.

Captions are full sentences, since Krea 2 reads sentences, and they only describe what changes: scene, outfit, pose, light. Her fixed traits (the bob, the bangs, the streak, olive skin, the beauty mark) never appear in a caption, because what the caption leaves out sticks to the trigger. Arak's v1 proved the reverse.

Age words stay out too (young, youthful, petite, girl). Krea already pulls young on its own.

Training

Civitai's own trainer on .red, Krea 2 base, 2,000 steps, defaults otherwise: 1024, flip off, dim and alpha 32. It cost 1,400 Buzz, which is US$1.40. Krea 2 trains at 512 or 1024, never 768. It trains on Raw and runs on Turbo.

Two trainer details that already cost me:

• If your sentence captions open with the trigger, leave the trainer's trigger field empty, or Civitai pastes it in front a second time. The trigger goes in at publish.

• Trainer samples render on the base model, and people run the distilled one. A sample is a first look. The verdict comes from a battery on the distilled model, same seed.

Ten scenes, no reference

The battery ran on my local 16 GB card at zero cost: Krea 2 Turbo, LoRA at 1.0, seed 101, 1024x1280, ten scenes across epochs 10, 8 and 6, about 26 seconds an image. Trigger only, no reference image anywhere. One control on top: the close-up prompt without the LoRA.

By my eye, the face held in all ten scenes and all three epochs, on seed 101. The control gave me a different woman.

One character isn't a study, and Arak differed from Odile in three ways at once: creature against human, Klein 4B against Krea 2, mixed sources against one. Still, it's why my next human characters start on Krea 2 from a single anchor.

I shipped epoch 10.

What it gets wrong

These went on the model page, because you'll run into them:

• Leave the outfit out of the prompt and you get her black slip from the dataset. Name the clothes and she wears them, every time.

• Signage in the scene can print the trigger. My neon sign read ODILESINT.

• In every epoch the face reads a few years older than the anchor.

One upload tip. The gallery PNGs carry the prompt in A1111 format, including the Steps: line, with the epoch's AutoV2 hash under Lora hashes. Drop the Steps: line and Civitai links the resource by hash but shows an empty prompt. It happened to me on another upload.

Using her

• Trigger: odilesint

• Weight 0.8 to 1.2 (my battery ran at 1.0)

• Images you generate are yours to sell. The model isn't for resale. The base sits under Krea 2's community license (commercial use under US$1M a year), so read it before you build a product on top.

Try it: https://civitai.red/models/3024218/odile

If you've trained a human character from mixed sources, did the face hold without a reference, or did you get the average?