QI_2.1_AnyAngle Camera Angle Control LoRA: 48GB+ VRAM, File List

QI_2.1_AnyAngle Camera Angle Control LoRA: 48GB+ VRAM, File List

Sample outputs are available on the model card.

At a Glance

Item Value
Repository lilylilith/QI_2.1_AnyAngle
Family guide Qwen-Image-2.1 guide (3 articles)
Publisher guide Alibaba (Qwen): models and licenses
Published 2026-09-28
License apache-2.0
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

In AI image generation, specifying camera angles in prompts often fails to yield the intended composition, a common challenge faced by many engineers. While existing image editing models (such as Qwen Image Edit and FLUX.2) are beginning to make it possible to change camera angles in existing images, most of them have limited azimuth and elevation options, falling short of the free control creators demand. Furthermore, it has been reported that attempting dramatic angle changes causes the image style to deviate significantly from the original.

“QI_2.1_AnyAngle" is a LoRA designed to overcome these challenges. This model inputs the original image and a “coarse" image generated from a 3D model or Gaussian Splat as references into Qwen Image 2.1, transforming the image to any camera angle while maintaining style consistency.

The specific workflow is as follows. First, a Gaussian Splat or 3D model is generated from the original image using tools such as Tripo Splat, Trellis2, or Pixal3D. Next, the model is imported into Blender, adjusted to the desired camera angle, and rendered. Finally, this rendered image and the original image are provided to Qwen Image 2.1 with the AnyAngle LoRA applied, along with the instruction prompt “Change the camera angle from to .", to achieve the target angle.

Specifications

  • Base Model: Qwen/Qwen-Image-2.1 (7B parameters for visual generation component)
  • Architecture: LoRA
  • License: apache-2.0
  • Recommended Settings:
    • LoRA Strength: 1 – CFG Scale: 3.0 – Steps: 20 steps or more
  • Supplementary: If fast inference is required for applications such as shot planning, a turbo lora can be used to reduce latency at the cost of a slight drop in quality.

Performance and Quality

Regarding the training method of this model, efforts have been made to maximize style consistency. The training uses photorealistic rendered images generated via Blender with diverse styles, as well as MiniMax H3, a SOTA video model for digital illustrations. In the process using Blender rendering data, pairs of frames from a camera angle different from the original image (anchor) are created (target). Training is repeated by generating a Gaussian Splat or 3D model from the target image and using it as a “coarse render" in the control image.

Furthermore, it is designed so that consistency can be maintained and manipulated even in styles such as illustrations and sketches. This utilizes MiniMax H3’s image-to-video to generate “orbit" renderings where the subject remains completely stationary while rotating around it. This achieves high consistency where the style does not break even when the angle changes.

However, generation quality depends on the accuracy of the input 3D data. It is reported that if the generated Gaussian Splat or 3D model does not accurately recognize space or if the modeling is too coarse, it may cause misalignments in item placement or distortion in facial features. Such issues are particularly prone to occurring when the anatomical structure is extremely distorted or when generated from low-resolution images.

Strengths and Use Cases

This model excels at freely changing only the camera angle while maintaining the style (art style) of the existing image. Specifically, the following use cases are anticipated:

  • Advanced angle manipulation from realistic images using Blender rendering data
  • Angle changes maintaining style consistency using orbit rendering by MiniMax H3 for digital illustrations and sketches
  • Fast inference for the purpose of rapid composition checking, such as shot planning (using turbo lora)

It is tagged with image-to-image and is intended to be used by incorporating it into advanced image editing workflows via 3D models or Gaussian Splat.

How It Differences from Similar Models

Differences from other models based on Qwen-Image-2.1 are as follows:

  • “Qwen-Image-2.1-viggle-turbo" 4-Step Fast Image Generation and Editing Model: Required VRAM 48GB+ is a distilled model that enables fast generation in 4 steps using DMD, whereas this model aims for freer camera angle control and style preservation by using 3D models or Gaussian Splats as references.
  • “Qwen-Image-2.1-Uncensored-GGUF" Image Generation Model: Required VRAM 16GB+ is a model focused on weight reduction through quantization, while this model functions as a LoRA specialized for a specific editing function (camera angle manipulation).

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 7.1B parameters (taken from the base model Qwen/Qwen-Image-2.1)

Your VRAM Quantization File size Est. memory needed
48GB (RTX 6000 Ada / A6000, etc.) BF16 30.8GB 37.0GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. This release is an adapter (LoRA etc.); the table shows what the base model Qwen/Qwen-Image-2.1 needs. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Distributed Files

Weight files published in lilylilith/QI_2.1_AnyAngle, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
QI2.1_AnyAngle.safetensors 120MB

How to Get It

Related Articles

What to Read Next

Sources