akhaliq/flux-3-image-klein-angles-lora

FLUX 3 Image Klein — Multiple Angles LoRA

Camera-control LoRA for FLUX 3 Image Klein (base): re-render a subject from a new camera pose — 8 azimuths × 5 elevations × optional distance — using fal's <mva> prompt grammar (the one from the Qwen-Image multiple-angles LoRAs).

Full 360° orbit around a dog, a cat and a person, driven by the same <mva> azimuth prompts:

Usage (diffusers)

import torch
from diffusers.modular_pipelines import Flux3AutoBlocks

blocks = Flux3AutoBlocks()
blocks.sub_blocks.pop("prompt_upsampler", None)  # important: LoRA expects raw prompts
pipe = blocks.init_pipeline("black-forest-labs/flux-3-image-klein-base")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")
pipe.load_lora_weights("akhaliq/flux-3-image-klein-angles-lora")

image = pipe(
    prompt="<mva> right side view, elevated shot",
    image=input_image,            # your reference image (PIL)
    num_inference_steps=50,
    generator=torch.Generator().manual_seed(0),
    output="images",
)[0]

Prompt grammar

<mva> {azimuth} view, {elevation} shot[ close-up]

Azimuths	Elevations	Distance
front view	low-angle shot	(none)
front-right / front-left quarter view	eye-level shot	close-up
right / left side view	elevated shot
back-right / back-left (quarter) view	high-angle shot
back view	top-down shot

Training details

• Base model: black-forest-labs/flux-3-image-klein-base (undistilled, trained via the public mirror)
• Script: diffusers train_dreambooth_lora_flux3_klein_img2img.py (i2i DreamBooth)
• Dataset: akhaliq/flux3-multiple-angles-train-4k (4,000 synthetic render pairs, fal <mva> grammar)
• Config: rank 16 / alpha 16, bf16, 512px, lr 1e-4 constant, 8-bit AdamW, 2000 steps on A100 (checkpoints at 500/1000/1500/2000 in the repo)
• Trained 2026-10-16, the day after the FLUX 3 Image Klein open-weights release

Limitations

Trained on synthetic 3D renders, so it generalizes unevenly to real photos: people and animals re-angle reliably, but plain objects can be replaced by hallucinated subjects or scenes on some prompts. A v2 mixing in real multi-view photo data is planned.

Demo

Try it: akhaliq/flux3-klein-angles-demo

License

Trained on top of FLUX 3 Image Klein — weights subject to the FLUX Kommunity License.