Same training photos · three matched prompts · six outputs

Comparing FLUX.1 [dev] vs FLUX.2 [dev] for character LoRAs

We trained FLUX.2 Dev on the exact 46-photo archive behind our working FLUX.1 Dev Mila LoRA, then ran both adapters on the same three prompts, seeds, trigger word, and 3:4 framing. FLUX.2 made polished images, but it did not preserve Mila's identity.

Training photos

Exact original archive

Training steps

1,000

Same step count

Matched outputs

3 pairs

Same prompts and seeds

The result

FLUX.1 remains our Mila baseline.

Both adapters completed the same three requests without a safety block. The difference was identity: the FLUX.1 controls stayed closer to the established Mila look, while all three FLUX.2 results depicted a different-looking person. We rejected the FLUX.2 adapter and made no product change.

The paired outputs

Same prompt and seed, different identity

These are all six unedited outputs. Each row shares one exact prompt, seed, trigger word, and 768×1024 frame. None triggered the provider's NSFW signal, and no alternates were generated.

Prompt 1

Close-up by a window

View exact shared prompt

Editorial close-up portrait photo of Mila, an adult woman, seated beside a large apartment window in soft natural morning light, relaxed direct gaze, subtle natural smile, realistic skin texture, crisp eyes, shallow depth of field, 85mm lens, candid high-end photography.

Prompt 2

Half-body cafe portrait

View exact shared prompt

Realistic half-body lifestyle photo of Mila, an adult woman, standing outside a modern city cafe in a fitted black jacket and simple white top, one hand lightly holding a takeaway coffee, natural posture, overcast daylight, street photography, 50mm lens, detailed hands and face.

Prompt 3

Full-body rooftop portrait

View exact shared prompt

Photorealistic full-body fashion photo of Mila, an adult woman, standing on a city rooftop at golden hour in an elegant red midi dress and black heels, both hands visible, relaxed balanced pose, skyline behind her, natural proportions, cinematic editorial photography, 35mm lens.

What was controlled

A controlled character migration test

The FLUX.2 trainer received the exact archived 46-image Mila dataset used for the original FLUX.1 LoRA. Both runs used the trigger word “Mila” and 1,000 training steps. Every reference had passed adult age review before reuse.

We then reran the established FLUX.1 adapter with the exact three FLUX.2 prompt hashes, seeds, trigger word, aspect ratio, and 28 inference steps. Each model used its production guidance setting: 3.5 for FLUX.1 and 2.5 for FLUX.2. This is still one character and three prompts, so it supports a Mila migration decision—not a universal ranking of the models.

Practical takeaway

Newer did not mean safer to migrate.

Keep the known-good FLUX.1 LoRA in production.
Evaluate identity before judging texture or aesthetics.
Use a small pilot before moving a full character library.
Reject a migration when the character is no longer recognizable.

What about FLUX.3 Dev?

Announced, but not available for this test yet

As of August 21, 2026, Black Forest Labs has announced FLUX.3 and made FLUX.3 Video generally available. The company lists FLUX.3 Image and the open-weight FLUX.3 Dev variant on its roadmap; it has not released a public FLUX.3 Dev image checkpoint or character-LoRA trainer that we can use for an equivalent test.

That makes FLUX.3 the next test to watch, not a current replacement candidate. See the official FLUX.3 announcement and the current release roadmap.

Bottom line

FLUX.2 produced attractive portraits, but attractive is not enough for a character LoRA. When the person changes, the model has failed the job. In this one-character, same-dataset pilot, FLUX.1 won by retaining the identity we actually needed.

See more model tests Train a character model