FLUX-3-Image-Klein-TrueRelight-LoRA

TrueRelight is an instruction-based relighting LoRA (rank 32) trained by community author pandacooming on FLUX.3 Image Klein Base. It is recommended to pair it with the 16-step distilled version of Klein: prompts should start with "Relight <ref_image_0>:" and clearly specify the direction and color of the light, while explicitly stating that the background should remain unchanged. The training data consists of 280 pairs of pixel-aligned before-and-after relit images, which have been made public on Hugging Face; the model card provides diffusers code and ComfyUI integration instructions, along with weights for a more contrast-enhanced 1500-step version.

TrueRelight is an instruction-based relighting LoRA (rank 32) trained by community author pandacooming on FLUX.3 Image Klein Base. It is recommended to pair it with the 16-step distilled version of Klein: prompts should start with "Relight <ref_image_0>:" and clearly specify the direction and color of the light, while explicitly stating that the background should remain unchanged. The training data consists of 280 pairs of pixel-aligned before-and-after relit images, which have been made public on Hugging Face; the model card provides diffusers code and ComfyUI integration instructions, along with weights for a more contrast-enhanced 1500-step version.

Why it matters Relying solely on prompts to change lighting often results in an effect similar to applying a filter to the entire image; this model can generate directional key lights, rim lights, and shadows, and it works even on the few-step distilled version of Klein, with an online demo included.

Hugging Face likes 3 · Hugging Face trending score 3 · Hugging Face downloads 156

License: Non-commercial · ComfyUI: Mentioned in the model card · VRAM: Not specified

Reasons

  • The model card includes 5 sets of comparison images with the same seed, showing results with and without the LoRA, and notes cases where v1 performs poorly
  • An online Hugging Face demo and a Colab notebook are available

Usage

  • In ComfyUI, load black-forest-labs/flux-3-image-klein-comfy and connect it using a standard LoRA Loader (strength 0.8–1.0); connect the input image to the reference latent, skip prompt expansion, use 16 steps, CFG 1.0, and the euler / simple sampler; example prompt: "Relight <ref_image_0>: relight with warm sunlight from the right, keeping the exact background unchanged"

Risks

  • FLUX Kommunity License, commercial use is prohibited
  • Very hard shadows on faces and necks can only be partially mitigated; the background may also change if scene-specific terms like "studio" or "night street" appear in the prompt

Open

抱脸 image-to-image 新模型

Explore the latest AI Radar updates