Hi everyone! I just open-sourced TrueRelight, a physical & directional relighting LoRA trained on black-forest-labs/flux-3-image-klein-base and evaluated on the 16-step distilled black-forest-labs/flux-3-image-klein.
Plain prompt-only relighting on distilled FLUX.3 Image Klein often behaves like a flat color wash (e.g., tinting existing shadows orange/yellow instead of reversing light direction). By training on 280 pixel-aligned <ref_image_0> pairs (rank=32, 1024px) and popping prompt_upsampler during 16-step inference, TrueRelight can:
pytorch_lora_weights.safetensors, 160.7 MB): https://huggingface.co/datasets/pandacooming/FLUX-3-Image-Klein-TrueRelight-Dataset``https://huggingface.co/pandacooming/FLUX-3-Image-Klein-TrueRelight-LoRA/blob/main/FLUX_3_Klein_TrueRelight_Demo.ipynb``diffusers)
Pop prompt_upsampler before running the 16-step pipeline so the <ref_image_0> grammar reaches the LoRA cleanly:
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")
(Also compatible with flux-3-image-klein-comfy at strength 0.8–1.0, 16 steps, CFG 1.0, euler/ simple, skipping the prompt upsampler.)
Would love feedback from anyone testing it on portraits, tricky shadows, or ComfyUI workflows!