Since I have seen some decent interest (and had a few DM conversations with other ES members) around 3D conversions, I thought I would make a post outlining some of the basics for those who may want to try this themself or even just understand the process.
This will not be a fully comprehensive walkthrough like some others I have written in the past. I will assume you have basic knowledge of windows (or linux) terminal and can follow the instructions on the 3D conversion github in addition to my outline here.
Let’s get started!
What you will need
A fairly powerful PC with an Nvidia GPU (Must be Nvidia for CUDA support) that has at least 6GB (ideally 8gb) VRAM.
Knowledge of Windows/Linux terminal and related terms
Follow the instructions for Windows or Linux linked on the main github page to extract the files and run the appropriate install scripts (install.bat for windows).
Get new AI model for IW3
Follow the instructions on the IW3 subpage (nunif/iw3 at master · nagadomi/nunif · GitHub) to obtain the new “Distill” model (you will want the “L” version unless your video card has low VRAM, in which case try the “S” or “B” variants. Place it in the IW3 pretrained models folder and name it appropriately.
Setup IW3 and Start 3D Converting
Launch IW3 (iw3-gui.bat on windows) and use the following settings, leaving everything else at default:
3D Strength: 4.0*
Depth Model: Distill_Any_L
Depth Resolution: 512
Edge Fix: 2 (Check the box as well)
Flicker Reduction: 0.99 (Check the box as well)
Max FPS: 1000
Video Format: MP4 or MKV
Video Codec: hevc_nvenc
CRF: 24
Preset: slow
*: Occasionally you may need to lower 3D strength in some content if there is excessive warping or exaggeration. 4.0 works for almost everything however and is a good baseline. If you get excessive warping you can also try increasing Foreground Scale to a positive value (around 2) which will have the model not try and create as much 3D effect in the background, which can help reduce distortions at the cost of a little realism.
Here is a screenshot of my full settings at time of writing:
POV - the effect is most convincing for POV videos, but you can use it on anything
Clear/Sharp video - Anime/animation converts extremely well thanks to this. Use 4k sources when possible, or upscale your sources to a minimum 1440p before converting.
Simple Backgrounds / Minimal Camera Movement/Effects - The newer (distill) model handles camera movement a lot better, but it can still make some distortion during panning/zooming. Simpler backgrounds help hide any warping effects from “popping out” the subject.
Text, whether subtitles or UI elements for Fap Hero/etc. can be hit or miss with distortion. If it bothers you try and source “Raw” versions without subs/UI
As methods evolve I will try and keep this post updated. If you have any questions or want to talk about your experiences doing 3D conversions please feel free to drop a reply in thread!
this is what i am currently doing, can i follow your settings but instead of using your Distill_Any_L, Can i use the ZoeD_Any_N depth model instead? also let me know if i have to change the settings to pair with the ZoeD_Any_N depth model
I highly recommend against the ZoeD models as they are effectively “Generation 1” - the 3d effect is still there but artifacts are much worse and the 3d effect is less realistic.
The Any_V2 model is effectively Gen 2 (or 1.5) and a bit better than ZoeD.
The new Distill model is a significant upgrade from both, and is worth acquiring, effectively a Generation 3.
Here is the portion of the guide above that links to the appropriate section of the IW3 github. They have a full explanation of where to acquire the additional models (AnyV2 or Distill - you only need Distill) - as well as what to name the files and where to place them.
Step 2: place in C:/ or where ever you downloaded the IW3, find the folder it should be like this: \nunif-windows\nunif\iw3\pretrained_models\hub\checkpoints
Step 3: once it is there YOU MUST RENAME * Rename the downloaded .safetensors file based on your chosen model:
For Distill_Any_L: Rename to distill_any_depth_vitl.safetensors
For Distill_Any_B: Rename to distill_any_depth_vitb.safetensors
I believe I used #2 - I do not remember if I found reference to this or simply assumed because the other was labeled “demo” - but that should work.
I have not compared against Owl3d lately. When I first started this, I did a demo with Owl3D and DepthAnyV2 (the Gen 1.5/2 model) in iw3. The results were very close. The new Distill model is better than DepthAnyV2 and IW3 now supports single-eye only conversion (mirroring functionality in Owl3D) so in my estimation it is just as good.
Definitely not normal, it sounds like your PC may not be using your GPU properly - do you see your 4060ti shown as the processor in IW3 or is it maybe trying to use CPU?
There is also a discord for IW3 if you want to ask there as well, since im not a developer for the project I only have limited insight on technical issues: Iw3 2d to 3d
