Do character LoRAs still matter?
last edited
Yes, character LoRAs still matter, but for fewer jobs than before. Reference-edit models do handle a one-off identity transfer from a few photos. A trained LoRA wins on likeness across many shots, video, and licensing.
What do reference-edit models do well?
MiniMax-H3-Image edits with up to 16 reference images in one checkpoint. Qwen-Image-2.1 takes up to 10. Qwen-Image-Edit-2511 lists better character consistency as a headline fix. Nothing is trained, so a new character costs minutes, and for a single image that speed matters. Limits per tool sit on character consistency and iteration.
Where does a character LoRA still matter most?
| need | reference edit | character lora |
|---|---|---|
| same face over many shots | drifts, rerolls needed | weights hold the face |
| new character today | 1 to 16 photos | dataset first |
| video clips | h3 video takes 9 images | swap loras act on clips |
| offline disk | 68.65 GB transformer | small file on the base |
A tester on the model discussion board found likeness slightly worse than the H3 video model. Another user said faces missed most of the time, yet two video LoRA files still loaded on the image model. Stacking is covered in multi-LoRA workflows.
Does the license matter?
| model | license | identity claim |
|---|---|---|
| MiniMax-H3-Image | territory-limited | 16 references |
| Qwen-Image-Edit-2511 | apache 2.0 | better consistency |
| Qwen-Image-2.1 | qwen research license | 10 references |
The H3-Image license text excludes the US, EU, UK, and one more market. It needs written approval above 1 million USD revenue or for hosted services. A character LoRA trained on a permissively licensed base avoids that, with a small file.
Sources
- model card: reference limit, checkpoint sizes, distilled design.
- user reports: community posts on likeness versus the video model.
- edit model card: consistency claims and license.
- Qwen-Image-2.1 card: ten-reference editing and the research license.