Trained with OSTRIS AI-Toolkit @ runpod.io
Flux Klein:
Seamfix LoRa:
Comfyui 360 Workflow
Seamless 360° Image Generation with Flux Klein
One year after my first experiments generating 360-degree images with Flux, I'm now taking a fresh approach using the klein 9b model. The results have improved significantly compared to my earlier attempts.
To achieve truly seamless 360° images, I trained a second LoRA specifically designed to support seamless inpainting. This additional LoRA helps repair the seams that typically appear where the image wraps around, ensuring a smooth transition across the entire panorama.
With this comfryui workflow—Flux Klein 360 Lora plus the seamfix LoRA—it's now possible to generate mostly seamless 360-degree images consistently.
Flux Kontext:
Comfyui Workflow:
Transform any input image into immersive 360-degree panoramic views with this specialized LoRA model.
Model Description
This Kontext LoRA has been specifically trained to convert regular input images into full 360-degree panoramic representations. The model leverages advanced training techniques to understand spatial relationships and generate contextually appropriate wraparound imagery.
Training Details
Dataset: Custom before/after image pairs showcasing the transformation from standard images to 360-degree panoramas
Training Framework: Ostris AI Toolkit
Model Type: Kontext LoRA for Flux
Purpose: Image-to-360° panorama conversion
Usage Notes
Base Quality: The raw output without upscaling may appear less refined
Recommended Enhancement: Use tiled diffusion upscaling for significantly improved results
Workflow: A complete ComfyUI workflow for creating seamless 360-degree images will be uploaded soon
Applications
Perfect for creating immersive content, virtual environments, VR experiences, and panoramic artwork from standard input images.
----------------------------------------------------------------------------------------------------
Flux Version:
The LoRA can create 360-degree panoramic images, and there is a high probability that the textures at the edges will be seamless.
However, as is often the case with AI-generated images, it doesn't always work perfectly.
But it works ;)
Here’s a tip for creating the images:
Choose a 2:1 format, such as 1536x768 pixels, for the first sampler. Then, use the Ultimate Upscaler to do a 2x or 4x upscaling. This should help maintain quality while generating 360-degree panoramic images.
The 'SCHNELL' version is trained with Rank 64 and SCHNELL as a Base Model, making it a bit more resource-efficient. 'Dev' offers the best image quality with Rank 128, but if speed is the priority, 'SCHNELL' is ideal.
Description
FAQ
Comments (4)
Can you train something like Qwen Edit's Multiple Angles for Klein?
Note that it is not rotation of a subject in place, but rather rotation of the entire scene, given a reference image.
This has the potential to make way for Multi-Angle LoRAs like with Qwen Edit. I think it could already be used as such if you chain multiple inputs and outputs together... Will be experimenting.
No, for multi-angle you need a different dataset. It’s quite time-consuming to create, and as a 3D beginner it’s too complex for me.
If Louis ever shares his dataset, I could train it.
@denrakeiw This can already be used in a hacky way for multi angle by creating a panorama with the subject removed, projecting the panorama and viewing the scene at a different angle, rotating the subject separately, and reinserting them into the scene. But it's a lot of steps and generation.
For multi-angle: if I make a dataset for you, would you train it?