FLUX.2 Klein 9B + Consistency LoRA

Klein 9B image editing with a Consistency LoRA that stops pixel drift.

Upload an image, describe what to change, and Klein 9B rewrites those parts. The Consistency LoRA (by dx8152) keeps everything else locked. Backgrounds don't shift. Faces don't warp. Fine details stay aligned between before and after.
No trigger word. Pick a strength, write your prompt, run it.

How do you edit images with Flux 2 Klein 9B and the Consistency LoRA?

Upload your image, write a prompt describing the edit you want, and set the LoRA strength based on how much you want to preserve. The workflow handles the rest. Defaults are tuned so most users only touch two things: the prompt and the LoRA strength.
Edit Prompt Describe what you want to change in the image. Be specific about what to add, replace, or remove. For example: "replace the man's clothes with Roman knight armor, replace background with a medieval room, remove the laptop." The more precise your prompt, the better Klein 9B targets the edit. Avoid mood words like "cinematic" or "dramatic" since the model amplifies those and they become hard to undo.

LoRA Strength (default: 0.8) Controls how hard the Consistency LoRA works to preserve what you didn't ask to change.

• 0.2 to 0.4: Want big edits with more room for the model to reinterpret? Start here. Klein 9B has more freedom but may drift in areas you wanted untouched.

• 0.5 to 0.7: A middle ground. Edits come through clearly while most of the original stays stable.

• 0.8 to 1.0: Need backgrounds and faces to stay close to the original? Go here. At 1.0, the LoRA locks down hard and edits stay surgical.

Play with it on a test image to find your sweet spot.

Steps (default: 8) Number of sampling steps. 8 is the default and works well for most edits. Higher values add detail but slow things down. For quick tests, 6 to 8 is fine. For final output, try 10 to 12 and see if it sharpens the result.

CFG / Guidance (default: 1) Controls how closely the output follows your prompt. At 1, the model stays relaxed and natural. Higher values push harder toward your prompt but can introduce artifacts. For most edits, leave this at 1.

Seed Set to "randomize" by default. Lock the seed to a specific number when you want to compare the effect of different prompts or LoRA strengths on the same generation. Useful for A/B testing your settings.

Resolution Your input image gets scaled to 1 megapixel before processing. The output matches your original aspect ratio. No need to resize your images beforehand.

What is the Consistency LoRA good for with Klein 9B?

The Consistency LoRA is built for edits where most of the image needs to stay identical. It prevents the slow pixel drift that happens when you run multiple edits on the same base image, making it ideal for batch work, product shots, and character-consistent storytelling.

Product photography is the clearest win. Change a product's color, swap a background, add a prop, and the product shape, lighting, and reflections stay put. Run ten variations and they all look like they came from the same shoot.

Character editing is the other strong case. Swap outfits, change hairstyles, or adjust expressions while the face, pose, and environment stay locked. This makes it practical for storyboards, comics, or social media content where the same character appears across multiple panels or posts.

The LoRA also helps when you're iterating. Without it, Klein 9B can slowly deform fine structures (hair strands, fabric patterns, background textures) over repeated edits on the same image. The Consistency LoRA reins that in.

Doing a single quick edit where you don't care about preserving details? A regular Klein 9B workflow without the LoRA will be faster and gives the model more creative freedom.

FAQ

What LoRA strength should I use for Klein 9B Consistency edits?
Start at 0.8 (the default). If your edits feel too restricted and the model isn't making enough changes, drop to 0.4 to 0.6. If you're seeing drift in areas you wanted untouched, push toward 0.9 to 1.0. Test on one image before committing to a batch.

Do I need a trigger word for the Klein 9B Consistency LoRA?
No. Load the LoRA and set your strength. No special token or keyword in the prompt. Write your edit prompt the same way you would without the LoRA.

How many steps should I use for Klein 9B image editing?
8 steps is the default and handles most edits well. For quick previews, 6 works. For polished final output, try 10 to 12. Going above 15 rarely improves quality and adds generation time.

Can I use the Consistency LoRA for text-to-image with Klein 9B?
This workflow is built for image-to-image editing. The Consistency LoRA works by comparing your input to the output and reducing drift, so it needs a reference image. For text-to-image, use a standard Klein 9B workflow instead.

How do I run Flux 2 Klein 9B with the Consistency LoRA online?
You can run Flux 2 Klein 9B with the Consistency LoRA online through Floyo. No installation, no setup. Open the workflow in your browser, upload your image, and hit run. Free to try.