<span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Please note that the original source of this Lora is CivitAI. (<a target="_blank" rel="ugc" href="https://civarchive.com/user/worgman">civitai.com/user/worgman</a><span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">). It is available for download for free and should not be published elsewhere.This is a concept LoRA to manipulate the camera angle. It provides a fixed perspective shot of the image. Intended for NSFW content but is not limited to that. Please do post pictures below, as I'm interested to see what everyone is able to make with this tool.Note that Adetailer and inpainting small regions does not play well with my perspective LoRA due to the emphasis on camera angle manipulation. My recommendation is to disable the Lora, significantly reduce weight usage, or lower the inpainting denoise to mitigate this.Adetailer inpainting denoise of 0.25-0.30 mostly fixes this at the cost of a small amount of detail.<h1 id="spyview">Spyview</h1>This is a supplemental concept to the fixed perspective LoRA. It is completely separate, from Fixed Perspective V1 - V3 as of V1.0. This could change, but right now the plan is to keep them separate but continue updates for each within this page as they both are camera manipulation LoRA.As of Spyview V1 the style impact has been reduced, and there has been more variation in what you can prompt to be in the foreground. V1 is far from perfect, however. Due to how much flexibility I allowed in the LoRA you will run into bad outputs occasionally. Good prompting helps to avoid this.You do have some control over how far the subjects in the image are from the camera through the use of "out of frame" tags. If you call for a large amount of a subject to be out of frame the LoRA should bring the subject closer to the foreground. Alternatively, calling for Full Body in the prompt should do the opposite as the LoRA needs to output the entire image which makes the subject(s) appear further away.What does V1 do well?<ul><li>Low style impact compared to most of my V1 LoRA.</li><li>Very flexible<ul><li>You can prompt for combinations of items within the foreground. Note not everything works, but if you experiment, you'll find some interesting combinations can be done.</li></ul></li></ul>What does V1 do okay?<ul><li>The various trigger words are mostly differentiated from one another, but there is some overlap.</li><li>Sliding Door and Door foreground overlap a fair amount.</li><li>Door foreground tends to overpower other concepts</li><li>Window foreground could use some work. I tried to give too much control (indoor and outdoor looking in) and it caused issues.</li><li>Prompting for material type with door (for example glass door foreground) actually does work, despite no images of glass doors being used in the dataset. I'd like to continue this direction and add that type of flexibility into the LoRA. Will see how that goes over time.</li></ul>Where does V1 fail?<ul><li>Accuracy of LoRA to capture the prompter's intent is low. Certain tags within your prompt can cause wild variations in the output image. This is good for flexibility but not great for reproducibility of a specific idea. Testing was very difficult with this version as there were times when the images were not what I had anticipated with my prompt.</li><li>Certain concept tags are too diluted to work well. Object foreground is only about 3% of the overall dataset, so don't expect good results from using it.</li><li>The "architecture" of what is put in the foreground is nonsense depending on what is prompted. Door foreground usually shows the door at an unrealistic angle, or with multiple doorholes or handles, one on the wall and one on the door.</li><li>Keyhole foreground causes the model to put a keyhole on the image. I need to change the trigger word for this to fix it. It seems to either go at the very top of the keyhold cutout or gets put on the subject. Annoying, about 40% of the tests I did were fine, but it ruined the output of the ones that it showed up on. You can put keyhole in negative prompt to help some.</li></ul><h1 id="fixed-perspective-view">Fixed Perspective View</h1><h2 id="v1">V1</h2>My first attempt at creating a concept LoRA and something I have been collecting images for a while as I come across them. Felt like I had enough to give it a shot and this is the result. I will need to collect more to update it in the future. Requires ADetailer and HiResFix to get good outputs. If prompting more than one subject inpainting will likely be required to fix it up.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/91afcb5a-c2b2-4c15-8962-5164d2e8dd3b/original=true,quality=90/62546126.jpeg" /><h2 id="v2">V2</h2>I have expanded the size of the dataset used for training and have also categorized a section of the images to attempt to provide trigger words to manipulate the camera further. I was rough with it due to the sheer quantity of images used and limited time to spend pruning and cleaning up the tags. Due to that it could be improved, but I did manage to get part of the control that I wanted.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/3d219d9d-c5f9-4a3e-86ff-ef98a4b8448d/width=525/3d219d9d-c5f9-4a3e-86ff-ef98a4b8448d.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/119e4312-1692-4a87-8ea9-db098eb8d086/original=true,quality=90/00345.jpeg" alt="00345.png" /><h2 id="v3">V3</h2>Removed some images and added new sources to training dataset to attempt to have the LoRA have less of a style impact on the output as much as V2 or V1. Toyed around with tag weights to try and add further control. V3 is not strictly better than V2.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/20eaac95-2606-44df-bf5e-e9359a525cfa/width=525/20eaac95-2606-44df-bf5e-e9359a525cfa.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8085fc0d-57bf-431d-9112-75efd7d52c0c/width=525/8085fc0d-57bf-431d-9112-75efd7d52c0c.jpeg" />

Fixed Perspective View

Quadrupled training dataset size and added more variety to the dataset. Added additional trigger words to manipulate image output further. Increased step count in v2.

Files with SHA256: 9d960c951c959447ea5698eadbff079abfd348032195d313ad8a8a0d414ac4b9