Flux family
# Qwen Multiangle Advanced camera-aware image editing and perspective manipulation. ## Overview Qwen Multiangle is a specialized AI model designed to bridge the gap between static image editing and dynamic camera control. Built upon the robust Qwen-Image-Edit-2509 architecture, this model integrates a custom multi-angle LoRA to allow users to perform complex spatial transformations on existing assets. Whether you need to adjust the focal length, tilt the camera perspective, or zoom into specific details, Qwen Multiangle provides a precise, instruction-based interface for professional-grade image manipulation. Unlike standard generative models that simply repaint pixels, Qwen Multiangle understands the underlying geometry of the scene. By leveraging the 'go_fast' preset system and fine-tuned LoRA weights, it maintains high structural fidelity while applying requested camera movements. This makes it an essential tool for marketing teams looking to repurpose existing product photography or lifestyle assets into new, dynamic formats without the need for expensive reshoots or complex 3D rendering workflows. ## Use Cases * **Product Photography Repurposing:** Transform a standard product shot into a close-up or wide-angle view for social media banners. * **Dynamic Ad Creative:** Adjust the vertical tilt of an image to create a more dramatic or immersive perspective for digital advertisements. * **Consistent Brand Assets:** Rotate and reframe existing assets to fit specific layout requirements across multiple marketing channels. * **Visual Storytelling:** Use the zoom and tilt features to create a sense of movement and depth in static lifestyle imagery. * **Rapid A/B Testing:** Quickly generate multiple camera angles of the same product to test which perspective drives higher engagement. ## Parameters This model offers a comprehensive suite of controls for fine-tuning your output. Please refer to the Input Parameters section in the Pixloop dashboard for detailed descriptions of settings including seed, image, prompt, go_fast, lora_scale, aspect_ratio, lora_weights, move_forward, output_format, vertical_tilt, output_quality, rotate_degrees, use_wide_angle, num_inference_steps, true_guidance_scale, and disable_safety_checker. ## Tips for Best Results * **Start with High-Quality Inputs:** The model performs best when the source image is clear and well-lit; avoid heavily compressed or blurry uploads. * **Use 'go_fast' for Iteration:** Keep 'go_fast' set to true during the initial creative exploration phase to save time, then switch to false for final, high-fidelity renders. * **Combine Parameters:** Don't be afraid to combine 'move_forward' with 'vertical_tilt' to achieve complex, cinematic camera movements. * **Adjust LoRA Scale:** If the edit feels too aggressive or deviates too far from the original, try lowering the 'lora_scale' from the default 1.25 to 1.0 or 0.8. * **Use Descriptive Prompts:** Even though the model is camera-aware, providing a clear text prompt helps the model understand the context of the edit you are requesting. ## About Qwen Multiangle is powered by the Qwen-Image-Edit-2509 model, enhanced with a specialized multi-angle LoRA (dx8152/Qwen-Edit-2509-Multiple-angles) to provide superior control over spatial perspective. This model is part of the broader Qwen multimodal research initiative, which explores the intersection of complex reasoning, vision, and language. For more technical details on the underlying architecture, please refer to the research paper at https://arxiv.org/abs/2508.02324.
Model input reference derived from preset schema.
| Name | Type | Required | Default | Description |
|---|---|---|---|---|
| seed | number | No | null | Random seed. Set for reproducible generation. |
| image | image | Yes | "" | Image file uploaded to Cog (jpeg, png, gif, or webp). |
| prompt | string | No | "" | Optional text instruction appended after the camera directive. |
| go_fast | boolean | No | true | If num_inference_steps is omitted, true runs the 4-step fast preset and false runs the 40-step detailed preset. |
| lora_scale | number | No | 1.25 | Strength applied to the selected LoRA. |
| aspect_ratio | select | No | "match_input_image" | Aspect ratio for the generated image |
| lora_weights | string | No | "dx8152/Qwen-Edit-2509-Multiple-angles" | LoRA weights to apply. Pass a Hugging Face repo slug like 'dx8152/Qwen-Edit-2509-Multiple-angles' or a direct .safetensors/zip/tar URL (for example, 'https://huggingface.co/flymy-ai/qwen-image-lora/resolve/main/pytorch_lora_weights.safetensors', 'https://example.com/lora_weights.tar.gz', or 'https://example.com/lora_weights.zip'). |
| move_forward | number | No | 0 | Move the camera forward (zoom in). Higher values push toward a close-up framing. |
| output_format | select | No | "png" | Format of the output images. |
| vertical_tilt | number | No | 0 | Vertical camera tilt. -1 = bird's-eye, 0 = level, 1 = worm's-eye. |
| output_quality | number | No | 95 | Quality when saving the output images, from 0 to 100. 100 is best quality, 0 is lowest quality. Not relevant for .png outputs. |
| rotate_degrees | number | No | 0 | Camera rotation in degrees. Positive rotates left, negative rotates right. |
| use_wide_angle | boolean | No | false | Switch to a wide-angle lens instruction. |
| num_inference_steps | number | No | null | Explicit denoising step count (1-40). Leave blank to use the go_fast presets (4 or 40 steps). |
| true_guidance_scale | number | No | 1 | True classifier-free guidance scale passed to the pipeline. |
| disable_safety_checker | boolean | No | false | Disable safety checker for generated images. |