18+ only. This guide is for adults. Every technique here is meant for fictional, AI-generated adult characters or for material you own and are fully authorized to use. Do not base a character on a real person's face or photos.
An AI girlfriend image generator is only as good as its ability to show the same woman twice. Generating one attractive portrait is easy; generating the same fictional character across dozens or hundreds of images, outfits, settings and short videos is the real skill, and it is what separates a believable companion persona from a random gallery. Companion apps hand you a pre-made character, but many creators want control over appearance, personality and style. This guide covers the whole process, from writing a character specification to locking the face, training a custom model, building an image library and moving into video, with both a hosted route on ZenCreator and a do-it-yourself route in ComfyUI.
In brief: consistency comes from three ingredients. You need a fixed definition of the face (a reference set or a trained LoRA), a fixed definition of the style, and a workflow that applies both every time you generate. A hosted platform with built-in character tools is the quickest way in; training your own LoRA gives the most control if you have a capable GPU.
Why an AI girlfriend image generator struggles with consistency
Diffusion models do not remember faces. Every generation starts from fresh noise, so even a small change in wording can shift the jawline, eye shape or nose. Text alone is a blunt instrument: "green eyes, auburn hair, oval face" describes thousands of possible women, and the model resolves that description differently on every run.
There are three established ways to fix this, each with a trade-off:
- Image-prompt adapters such as IP-Adapter steer generation using a reference picture of the face. They are quick to set up but need the reference on hand every time.
- Fine-tuning methods such as LoRA or DreamBooth teach the model your specific character. They take time up front but give the most reliable long-term results.
- Hosted character tools package one of these approaches behind a simple interface, so you save a face once and reuse it.
The research behind them is well documented. IP-Adapter uses a "decoupled cross-attention" mechanism that handles text and image features separately, and its authors report that an adapter with only 22M parameters can match or beat a fully fine-tuned image-prompt model (IP-Adapter, arXiv). DreamBooth showed that from just a few images of a subject, a text-to-image model can be fine-tuned to bind a unique identifier to that subject and render it in new scenes (DreamBooth, arXiv).
The workflow at a glance
| Stage | What you do | What you end up with |
|---|---|---|
| 1. Character specification | Decide face shape, eyes, hair, signature details, personality | A written spec and a mood board |
| 2. Face lock | Generate with a reference image or saved face | A consistent starter set of 15–30 images |
| 3. Custom model (optional) | Train a LoRA on that set | A reusable character model with a trigger word |
| 4. Library | Generate portraits, outfits, settings in batches | A curated image library |
| 5. Video | Animate the best stills | Short clips of the same character |
Most people skip stage 1, and that is why their character drifts. A model can only hold on to features you have actually decided.
Stage 1: Write the character specification
Before you open any tool, describe the character as precisely as a casting brief. Start with a clear statement that she is an adult, then define the details.
Face shape and structure. Oval, heart, square or round? High cheekbones or a softer profile? A defined jaw or a narrow chin? Structure is what makes a face recognizable from different angles.
Eyes and brows. Go beyond colour: almond, round or hooded; spacing; brow thickness and arch. Eyes carry most of a character's expression.
Hair. Colour, length, texture, parting and how it frames the face. Hair is often the most recognizable trait at a glance because it survives changes of angle better than fine facial detail.
Signature details. Freckles across the nose, a beauty mark, a particular smile. One or two distinctive features make a character memorable; too many become noise.
Personality through appearance. Decide her default expression, because you will generate it more than any other. A relaxed half-smile works well for a companion character. A playful character might favour bright lighting and casual outfits; a more mysterious one, low-key lighting and darker palettes.
Write the spec as a fixed block of text you will paste into prompts, and keep a simple mood board for lighting, wardrobe and setting. Do not use photos of real people as face references; build the face from the generator itself.
Stage 2: Lock the face
The hosted route: ZenCreator
The fastest way to a consistent character is a platform that stores the face for you. ZenCreator includes a face generator for designing an original face, image generation with face-consistency features, reference-to-image tools, and a Characters feature that describes characters "with a face, a body and a personality" you can also chat with. Photo and video credits are charged when you unlock content.
A practical sequence:
- Use the face generator to create a handful of candidate faces from your written spec, and pick the one that matches best.
- Save it as your character's face, then generate 15–30 images with varied angles, expressions and lighting.
- Keep only the images where the face clearly matches; discard the rest.
- Use that curated set as the reference for everything that follows.
Credits are bought in packs starting at $19.99 for 200, there is no subscription, and credits never expire, so you can build a character over weeks without a monthly bill (ZenCreator Help: Pricing & Credits). Most image models cost 1 or 2 credits per image, and LoRAs do not change the per-image cost in prompt and reference tools.
The do-it-yourself route: IP-Adapter in ComfyUI
If you run models locally, install IP-Adapter through ComfyUI Manager along with a face-focused adapter (such as an IP-Adapter FaceID variant) and the face-analysis models it depends on. In the graph, feed the reference picture into the adapter and connect its output upstream of the sampling step.
Starting values that work for many people:
- Adapter weight around 0.7–0.85 for faces
- Start and end around 0.1 and 0.9, so the face holds while the prompt still controls pose and scene
- A sensible face-detection threshold, so partial faces in the reference are ignored
Generate 10–20 images from your spec with ordinary prompting first, choose the three to five that best match your vision, and use those as references. Good references share consistent lighting, show the face clearly with no hands or hair across it, include slight angle variation, and have enough resolution in the face area (at least 512×512 pixels). Avoid heavy makeup or dramatic lighting you do not want copied into every image.
When generating, describe pose, clothing, setting and mood in the prompt but leave the face out; let the adapter handle it.
Stage 3: Train a character LoRA (optional, advanced)
For the most dependable results, train a LoRA on your curated set. LoRA works by freezing the original model weights and injecting small trainable low-rank matrices into each layer; in the original paper on language models it cut trainable parameters by 10,000 times and GPU memory needs by three times compared with full fine-tuning (LoRA, arXiv). The same idea is now widely used to teach image models a specific character.
Preparing the dataset
- 15–30 strong images. Quality beats quantity: fifteen sharp, well-lit images outperform a hundred mediocre ones.
- Varied angles. Include front, three-quarter and profile views and different head tilts so the model learns the face in three dimensions.
- Identical identity. Every image must show the same character. Any drift in eye colour or bone structure will confuse training.
- Captions that skip the face. Describe clothing, background, pose and expression, but not facial features, so the model ties the face to your trigger word.
Typical starting settings (in tools such as Kohya_ss or AI-Toolkit):
| Setting | Starting point for a character |
|---|---|
| Network rank (dimension) | 32–64 |
| Alpha value | Match the rank, or use 50% of it |
| Learning rate | Between 5e-5 and 1e-4; err toward the low end |
| Training steps | 1,500–3,000 |
| Images per batch | One or two, limited by GPU memory |
Training usually takes from half an hour to two hours depending on hardware. Generate test images every few hundred steps to catch overtraining early.
Using the LoRA: load your base checkpoint, add the character LoRA at a weight of roughly 0.7–1.0, include the trigger word, and let the rest of the prompt control pose, outfit and setting. On ZenCreator, LoRA support in its editor and video tools means you can use a trained character model without hosting it yourself.
Stage 4: Build the character library
With the face locked, generate in planned categories rather than at random:
- Portraits: a range of expressions (happy, thoughtful, playful, relaxed) for profile images and conversation.
- Everyday scenes: reading, cooking, a morning coffee, a walk in the city. These give the character a life beyond posed shots.
- Wardrobe: casual, formal, sporty and evening outfits, all with the same face.
- Locations: home, café, beach, city at night, countryside.
- Seasonal sets: holidays and weather-appropriate looks planned in advance.
Work in batches of 10–20 images per category, using templates that only vary one element at a time (for example, outfit colour while keeping pose and expression fixed). Then curate ruthlessly. A small library of excellent images is worth more than a large one full of near misses.
Stage 5: Move into video
Once you have strong stills, animate them. Image-to-video keeps the face closest to your approved look because the still acts as the opening frame. Research systems such as Stable Video Diffusion were built for exactly this, extending an image model with temporal layers for text-to-video and image-to-video generation (Stable Video Diffusion, arXiv).
For best results, choose centred, well-lit stills with the face unobstructed, and ask for one simple action per clip: a smile, a head turn, a slow camera push. ZenCreator's video generator works from the same account and credits, and our guides to image-to-video and the Wan 3.0 features explain the motion side in more depth.
Common problems and how to fix them
The face drifts from image to image. Raise the adapter weight or LoRA strength, remove facial descriptions from the prompt, and fix the seed for key shots.
The face is stable but the style jumps around. Stick to one checkpoint or model, add a style reference or style LoRA, and keep a fixed style phrase in every prompt.
The LoRA looks plastic or only works in certain poses. It is overtrained. Use fewer steps, a lower learning rate and a more varied dataset.
In group scenes, the wrong person gets her face. Use regional prompting or pose guidance to tell the model which figure is your character.
She looks identical in every image. Vary pose, expression, outfit, setting and seed deliberately while leaving the face lock on.
Rules to follow
A companion character must be clearly fictional and clearly adult.
- No real people. Platform policies are explicit. Civitai states that "content that reproduces the likeness of real people, in any context, is prohibited" (Civitai Safety Center), and OpenAI prohibits using someone's photorealistic likeness without consent in ways that could confuse authenticity (OpenAI Usage Policies).
- Adults only, in appearance as well as description. ZenCreator's content policy requires that anyone depicted in a sexual context is 18 or older and does not appear to be a minor, and bans sexualized fictional or AI-generated characters who look underage (ZenCreator Content Policy).
- The law on intimate imagery. In the US, the TAKE IT DOWN Act covers AI-generated intimate depictions of identifiable people and requires covered platforms to remove them within 48 hours of a valid request (Public Law 119-12, GovInfo).
Keep your spec, prompts and dataset together so you can always show that the character is an original, adult creation.
FAQ
How long does it take to build a consistent character?
With a hosted face tool or IP-Adapter, you can have a usable, consistent set in a few hours. Training a LoRA adds a few more hours. A full library builds up over several weeks of regular sessions.
Can I base my AI girlfriend on a real person?
No. Using a real person's face without consent breaks most platform rules and can break the law, especially for intimate content. Design an original face instead.
Which GPU is enough for the local route?
For IP-Adapter workflows, plan on at least 8GB of VRAM (12GB or more is more comfortable). LoRA training is happier with 12GB or more. A hosted platform like ZenCreator needs no GPU at all.
How many reference images should I use?
Three to five strong references are usually enough for an adapter, and 15–30 for training a LoRA. Consistency and clarity matter more than numbers.
IP-Adapter or LoRA: which is better?
IP-Adapter is faster to set up and needs a reference each time. LoRA takes longer to train but gives steadier results and needs no reference once trained.
Can I animate my character?
Yes. Start from your best stills and use image-to-video for short clips. ZenCreator offers this in the same account as image generation.
Sources
- IP-Adapter, arXiv 2308.06721: https://arxiv.org/abs/2308.06721
- DreamBooth, arXiv 2208.12242: https://arxiv.org/abs/2208.12242
- ZenCreator Help Center, Pricing & Credits: https://help.zencreator.pro/articles/12068902-pricing-and-credits-plans-payment-methods-and-tool-costs?ref=baiwei
- LoRA, arXiv 2106.09685: https://arxiv.org/abs/2106.09685
- Stable Video Diffusion, arXiv 2311.15127: https://arxiv.org/abs/2311.15127
- Civitai Safety Center: https://civitai.com/safety
- OpenAI Usage Policies: https://openai.com/policies/usage-policies
- ZenCreator Content Policy / Acceptable Use Policy: https://help.zencreator.pro/articles/1373215-content-policy-acceptable-use-policy?ref=baiwei
- Public Law 119-12 (TAKE IT DOWN Act), GovInfo: https://www.govinfo.gov/content/pkg/PLAW-119publ12/html/PLAW-119publ12.htm