LoRA vs reference images vs .char: which to use for consistent characters
Use a LoRA when you need the strongest likeness on one model and can wait for training. Use reference images when you want speed and do not mind attaching them every time. A .char picks the right method for each model automatically.
TL;DR
| ROW | LoRA | Reference images | .char |
|---|---|---|---|
| Training needed | Yes, per model, and slow | No | None on FLUX.2 and H3 R2V. An adapter trained once on Krea 2 and other H3 nodes |
| Photos required | 20 to 50 is a good range | 1 or more, attached every time | 1 to 2 |
| Models supported | One model per file | Models that accept references | FLUX.2, Krea 2, FLUX.1, Z-Image Turbo, MiniMax H3, LTX-2.5, plus Nano Banana Pro and H3 R2V via fal |
| Portable file | Yes, one file per model | No, re-attached every time | Yes, one .char file |
| Face / body / wardrobe / voice | Whatever the training set covers, and no voice | Whatever is in the photos, and no voice | Face, body, wardrobe and an optional voice clip |
| Runs locally | Yes | Yes | Yes, in the OmniChar app and ComfyUI, or hosted on OmniChar Cloud |
| Open source | Yes, the training tools are | Yes, with open models | Yes, GPL-3.0 |
| Cost | Free on your own GPU | Free on your own GPU | Free to run locally. OmniChar Cloud uses credits |
| Drift measurement | None | None | Continuity score out of 100 on every take |
What each one is
LoRA and Reference images
A LoRA is a small set of weights trained on photos of one person. It gives the highest lock on likeness, but training is slow and each file works with one model. Reference images are photos you attach to each generation. They need no training and are fast, but you re-attach them every time.
.char
A .char is one portable file that holds reference photos, an identity fingerprint, a locked description, an optional trained adapter and an optional voice clip. It works in the OmniChar app, the ComfyUI node and OmniChar Cloud. FLUX.2 and H3 R2V use the references directly; Krea 2 and other H3 nodes use an adapter trained once. Every take gets a continuity score out of 100.
Strengths
Where LoRA and Reference images wins
- LoRA: the highest likeness lock on the model it was trained for
- Reference images: fast, with nothing to train
Where .char wins
- No need to pick a method for each model
- One file instead of a LoRA per model plus loose photos
- Carries body, wardrobe and voice too
- Scores every take out of 100 for drift
When to use which
Can you use both?
Yes. A .char can carry a trained adapter, such as a LoRA, alongside its reference photos. OmniChar uses the adapter on models that need it, such as Krea 2 and other H3 nodes, and the references on models that read them directly, such as FLUX.2 and H3 R2V.
Questions
What is the difference between a LoRA and reference images?
A LoRA is trained on photos and saved as a file for one model. Reference images are attached to each generation with no training.
Is a LoRA more accurate?
On the model it was trained for, a LoRA gives the highest lock. It does not carry over to other models.
Does .char replace LoRA?
No. A .char can hold one. It uses the adapter where it helps and the references where the model reads them directly.
Which models use references directly?
FLUX.2 and H3 R2V use references directly. Krea 2 and other H3 nodes use an adapter trained once.
How do I know if a take drifted?
Every take made with a .char gets a continuity score out of 100.