ausboss/FLUX-3-Image-Klein-Consistency-LoRA

FLUX 3 Image Klein Consistency LoRA

Early release, only lightly tested. People asked for it, so here it is as it came out of training. I trained it the same way as my Qwen Image 2.1 Consistency LoRA: edit pairs where the edited picture is lined back up with the original, so the LoRA learns to make the edit without moving the picture. Try it and let me know how it does. The two quick tests I ran so far are below.

FLUX 3 Image Klein edits can shift the picture, most of all restyles like a watercolor or an anime version: the scene comes back a little bigger, or nudged sideways. This LoRA is meant to keep the edit on the original's frame. There is no trigger word. Load it and write your edit as usual.

How to use it

• Put it on the distilled model at strength 1.0: 16 steps, CFG 1, euler / simple. It was trained on the base model.
• Your picture is the reference (Set Reference Multi Latent, resolution keep resolution), and the KSampler starts from the latent that node returns.
• Write the change, then what stays: "Change her hair color to copper red. Keep her face, her pose, the framing and everything else the same."

What I checked so far

Everyday edits. On 15 everyday edits that I rated blind (weather, light, clothes, outfits, hair color, poses, two people), it helped weather and light, and hurt hair color and pose. Outfits were mixed. Against plain FLUX 3 Image Klein it was better on 4, worse on 5 and the same on 6.

Restyles. This was a quick automated check: 9 edits on pictures that are not in the training set (6 restyles, 3 clothing changes), one seed each, with and without the LoRA. "Moved" is how far the picture shifted at its worst corner.

	plain FLUX 3 Image Klein	with the LoRA
restyles: moved, median	25.8 px	12.4 px
restyles: moved, range	6.8 to 37.5 px	2.8 to 18.5 px
clothing changes: moved	0.2 to 0.5 px	0.2 to 0.5 px

So far it roughly halves the shift on restyles but does not stop it, and the clothing changes stayed in place with or without it. The edits still happened in all 9.

Files

file	notes
flux-3-image-klein-consistency.safetensors	The LoRA. Step 3,000 of a 3,000-step run, rank 32, 161 MB. Loads onto the FLUX 3 Image Klein files in ComfyUI.
LICENSE.md, NOTICE.txt	The FLUX Kommunity License and the notice it asks for. See License.

ComfyUI

You need ComfyUI 0.41 or newer, which is the first version with FLUX 3. The FLUX 3 Image Klein files are on Hugging Face behind a license click: black-forest-labs/flux-3-image-klein-comfy.

Also for FLUX 3 Image Klein: my Outpaint LoRA and the Single & Multi-Image Edit workflow. Videos on YouTube @ausbossdev.

How it was made

• ostris/ai-toolkit, arch: flux3, on FLUX 3 Image Klein Base in BF16 without quantization. Rank 32, alpha 32, AdamW 8-bit, learning rate 1e-4 constant, batch 1, 3,000 steps with a checkpoint every 250, resolution 1408 buckets, caption dropout 0, no flips, shift timesteps. About 1 h 45 min on one RTX PRO 6000.
• 778 pairs from 389 edits of 231 pictures, the same source pictures and edit instructions as my Qwen Image 2.1 Consistency LoRA, minus a few I left out. This time every edit was made by FLUX 3 Image Klein itself: recolors, relighting, paintings, seasons, outfits, hair, accessories, anime / 3D / comic looks, photo looks, objects and backgrounds.
• Each edit was lined back up with its picture, so the target sits on the source's frame. Every edit gives a forward pair (picture to edit) and a reverse pair (edit back to picture). Local edits keep the original's pixels outside the thing that changed. Pairs that were still off after the fix were left out.
• Captions are the edit instructions.

License

This LoRA is a modified version of FLUX 3 Image Klein: an adapter I trained on FLUX 3 Image Klein Base. It is shared under the FLUX Kommunity License v.1.0, the same license as the base model, and I added no terms of my own. The rights to use FLUX 3 Image Klein and this adapter come to you directly from Black Forest Labs under that license. Whether you may use the pictures you make commercially depends on section 2 of it, so read it first.

Black Forest Labs did not make, endorse or approve this adapter. It is provided as is, with no warranty and no liability, as in sections 5 and 6 of the license.

The notice the license asks for, which is also in NOTICE.txt:

This FLUX Model is licensed by Black Forest Labs Inc. under the FLUX Kommunity License. Copyright Black Forest Labs Inc. IN NO EVENT SHALL BLACK FOREST LABS INC. BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.

I changed the FLUX Model: this file is a LoRA adapter trained by ausboss.