Character Consistency
Character consistency is the ability to generate images of the same fictional person across different poses, outfits, and settings — so that Instagram followers (and AI-detection systems) cannot tell the character has drifted between posts.
What is character consistency?
Character consistency is the ability to produce images of the same fictional character that look like the same person across different contexts — different outfits, locations, poses, and lighting conditions — without visible face drift.
Why it is the hardest problem in virtual influencer building: standard Stable Diffusion, when given the same text prompt, produces a different face every generation. The faces share broad characteristics (same hair colour, similar age) but the specific identity drifts. After 20 posts, the “character” no longer looks like the same person.
Character-consistent generation requires a custom LoRA — a fine-tuned model trained on 20-50 images of the specific character you want to maintain.
Why consistency matters commercially
For a virtual influencer account, visible face drift is the primary trust-breaker with followers and the primary signal AI-detector tools (used by Instagram and brand safety platforms) look for. When the face changes between posts:
- Long-time followers notice and call it out in comments
- Brand safety platforms flag the account as synthetic
- Brand-deal opportunities evaporate
Aitana López’s team maintains consistency by running the same LoRA checkpoint for every new image generation — the LoRA is the “identity document” of the character.
How to achieve character consistency
The standard workflow for a solo operator:
- Design the base character in Midjourney (£24/mo) — generate 50+ variations and select the 20-50 best as your LoRA training set
- Train a custom LoRA on RunPod or a local GPU — 4-hour training run, ~£5 in cloud GPU costs
- Deploy the LoRA in ComfyUI — load it in every future generation workflow
- Use IPAdapter for face-conditioning — an additional extension that anchors face structure from a reference image
- Maintain a “canon seed” — store the exact generation seed for your character’s face so you can reproduce it
Consistency loss: when it happens
Character drift typically appears when:
- You switch to a newer SD checkpoint (different base model)
- You generate at very different aspect ratios without adjusting
- The prompt description changes significantly between generations
- The LoRA training set was too small (under 15 images)
Related
- LoRA — the fine-tuning method that enables consistency
- Stable Diffusion — the base model
- IPAdapter — face-conditioning extension
- Character consistency — the real problem