eric-venti-seeds/Sun-Direction-Lora-Flux2Klein9B

Hugging Face Models Trending Models

Summary

A LoRA model for Flux 2 Klein 9B that enables controlling sun direction in exterior images by using a reference ball image and an overcast intermediate step. Includes a custom ComfyUI node for rendering sphere light references.

Task: image-to-image Tags: Flux2Klein, Sun, I2I, lora, lighting, ball, direction, overcast, image2image, LoRA, image-to-image, en, base_model:black-forest-labs/FLUX.2-klein-9B, base_model:adapter:black-forest-labs/FLUX.2-klein-9B, license:apache-2.0, region:us
Original Article
View Cached Full Text

Cached at: 07/06/26, 10:35 PM

eric-venti-seeds/Sun-Direction-Lora-Flux2Klein9B · Hugging Face

Source: https://huggingface.co/eric-venti-seeds/Sun-Direction-Lora-Flux2Klein9B

https://huggingface.co/eric-venti-seeds/Sun-Direction-Lora-Flux2Klein9B#sun-direction-lora-flux-2-klein-9b-v1Sun Direction LoRA Flux 2 Klein 9B v1

Trigger word:match the sun direction from the reference

LoRA trained to be able to move the sun to a specific elevation and rotation angle relative to the camera.

This lora mainly works for exteriors. If the image is not an exterior, I can’t assure it will work correctly as it hasn’t been trained with any interior images.

MAKE SURE THE IMAGE YOU ARE TRYING TO CHANGE THE SUN DIRECTION DOESN’T HAVE A CLEAR LIGHT DIRECTION OR MAKE IT OVERCAST FIRST (the workflow takes care of this).

Basically add your OVERCAST image and the ball image as latent reference and you have it!

I am already working in a v2 where this intermediate overcast step is not necessary, and other stuff like light hardness, color, intensity, etc etc


This LoRA uses a ball reference image, so I created a Sphere Render Node to make easier to specify where the light comes from.

If you don’t want to use the node, I added the blender scene to render the ball reference images in case you need it.

Download the Sphere Light Render Node from here:

sphere

https://huggingface.co/eric-venti-seeds/Sun-Direction-Lora-Flux2Klein9B#examplesExamples

https://huggingface.co/eric-venti-seeds/Sun-Direction-Lora-Flux2Klein9B#how-to-useHow to use

The trigger sentence is:

match the sun direction from the reference

Feel free to add any other words needed after the trigger sentence like: “add a clear blue sky” or “keep the cloudy sky”

It works way way better if the image doesn’t have a clear light direction, like an overcast day.

If the image you want to add or change the sun direction is already overcast or without a clear light direction you can deactivate the top part of the workflow.

The top part of the workflow generates an overcast image to be ready to apply the sun in bottom part of the workflow.

If you wanna do a timelapse with different light directions, generate the relight images with a fixed seed to avoid some flickering between the images.

Hope you like it, already working on a better v2!

Eric Venti.

Similar Articles

fal/LTX-2.3-3DREAL-LoRA

Hugging Face Models Trending

A LoRA adapter for LTX-2.3 that converts rough 3D viewport animations (from Blender, game engines) into photorealistic video while preserving composition and camera movement.

FLUX.2 [pro]

Replicate Explore

Black Forest Labs has released Flux 2 Pro, a new image generation and editing model with improved text rendering, photorealism, and character consistency. The model is available via API on Replicate.

RuneXX/LTX-2.3-Workflows

Hugging Face Models Trending

This Hugging Face repository provides workflows and model downloads for Lightricks' LTX-2.3 video generation model, designed for use with ComfyUI, including split models, GGUF versions, and required custom nodes.

Alissonerdx/LTX-Best-Face-ID

Hugging Face Models Trending

A LoRA for LTX-2 (22B) that enables identity-preserving reference-to-video generation using overlap reference conditioning, TASS-RoPE, and ArcFace identity loss, available for ComfyUI via BFS Nodes.

larryvrh/MiniMax-H3-Turbo-Lora

Hugging Face Models Trending

An early-preview LoRA for MiniMax-H3 that enables joint video and synchronized audio generation in 4 sampling steps instead of ~20, offering roughly 5x faster sampling, with ComfyUI custom nodes and three bf16 checkpoints.