Descriptions:
This tutorial from the Veteran AI channel explores a multi-angle LoRA for the Qwen Image Edit 2.5D model, trained on over 3,000 pairs of Gaussian Splatting data. The video explains why Gaussian Splatting-derived training data produces particularly strong results: because it reconstructs full 3D scenes, every object within becomes spatially adjustable, yielding geometrically accurate multi-perspective outputs that conventional training data cannot match.
The LoRA supports 96 distinct camera positions built from three coordinate systems: three distance levels (Far, Medium, Near), eight directional angles (Front, Back, Left, Right, and four 45-degree diagonals), and four height elevations (-30°, 0°, 30°, and 60°). The video walks through the exact prompt structure required—using the trigger word `<|sks|>` followed by azimuth, elevation, and distance keywords in that order—and recommends setting the LoRA weight between 0.8 and 1.0, with 0.9 as a safe default.
The full workflow is demonstrated on RunningHub’s ComfyUI platform using the Qwen 2.5D BF16 main model chained with an acceleration LoRA and the multi-angle LoRA at 1280×720 resolution. Practical tests show that single-subject scenes produce the most interpretable directional results, while two-subject scenes can introduce ambiguity in front/back assignments. The capability is presented as a significant tool for AI filmmakers looking to add visual variety and cinematic depth to generated content.
📺 Source: Veteran AI · Published January 09, 2026
🏷️ Format: Tutorial Demo
