TrueRelight
16-step photo relighting for FLUX.3 Klein
2 min readupdated
TrueRelight is a LoRA (a lightweight add-on for a base model) by pandacooming that changes the lighting of existing images based on text prompts. Trained on black-forest-labs/flux-3-image-klein-base, the adapter is optimized for fast 16-step generation on the distilled flux-3-image-klein version. Unlike standard prompt edits that just tint the frame like a color filter, this model builds physically plausible directional light, rim highlights, and cast shadows without altering the subject's features, framing, or background.
Key features
- Precise directional control: lets you rotate the light source, remove harsh shadows, or add dual-tone lighting.
- Detail preservation: keeps the original background, composition, and subject appearance intact when explicitly stated in the prompt.
- 16-step inference: optimized for fast generation on the distilled
flux-3-image-kleinmodel. - Two checkpoints to choose from: the main
step 1000checkpoint for a natural balance, andstep 1500for deeper, more dramatic contrast. - ComfyUI compatibility: works via the standard LoRA Loader node on
flux-3-image-klein-comfy(BF16 or INT8).
What you can do
- Rescue harsh outdoor portraits: remove a hard diagonal shadow across a face and replace it with soft, diffused daylight.
- Change the time of day: add warm sunset rim lighting or golden highlights while keeping the original background scenery.
- Apply cinematic color lighting: illuminate a subject with neon lights or a warm orange and cool teal combination.
Example
from diffusers import ModularPipeline
from diffusers.utils import load_image
blocks = ModularPipeline.from_pretrained("black-forest-labs/flux-3-image-klein").blocks
blocks.sub_blocks.pop("prompt_upsampler", None)
pipe = blocks.init_pipeline("black-forest-labs/flux-3-image-klein").to("cuda")
pipe.load_lora_weights("pandacooming/FLUX-3-Image-Klein-TrueRelight-LoRA")
img = pipe(prompt="Relight : warm sunset light from the right", image=load_image("in.jpg"), num_inference_steps=16, output="images")[0]
Read more
DetailsRecommended LoRA strength: 0.75–1.0 in diffusers and 0.8–1.0 in ComfyUI (16 steps, CFG 1.0, euler / simple sampler).
Training: rank=32, lora_alpha=32, BF16, 1500 steps on an A100 80GB GPU using a dataset of 280 image pairs at 1024px resolution.
Make sure to disable prompt_upsampler and start your prompt with the
Relight : token.Limitations: words like "studio" or "night street" can redraw the background, and subtle brightness tweaks without a clear light direction are less effective.
License: FLUX Kommunity License v.1.0.










Comments