This media is not supported in your browser
VIEW IN TELEGRAM
Load Video UI - Custom Node to Trim, Resize, and Preview Videos in Realtime
https://redd.it/1t3x1ec
@rStableDiffusion
https://redd.it/1t3x1ec
@rStableDiffusion
It's the 24th century. How is there still no actually good porn model?
https://redd.it/1t445jw
@rStableDiffusion
https://redd.it/1t445jw
@rStableDiffusion
Converting 2D animations to 3D with LTX 2.3 Lora
https://www.youtube.com/watch?v=g88lmbYmZWs
https://redd.it/1t4a0r5
@rStableDiffusion
https://www.youtube.com/watch?v=g88lmbYmZWs
https://redd.it/1t4a0r5
@rStableDiffusion
YouTube
2D to 3D Animations
This is my first attempt at converting 2D animations to 3D. These clips are the result of an on going experiment.
All of my AI demos:
https://www.youtube.com/playlist?list=PLe3OBqR7FeRhZM6SNoIWibQ1PA2JREYtL
All of my AI demos:
https://www.youtube.com/playlist?list=PLe3OBqR7FeRhZM6SNoIWibQ1PA2JREYtL
My LTX 2.3 LoRA Training Journey: Fighting for VRAM even with a 5090
I recently completed a training run for an LTX 2.3 LoRA and wanted to share my settings and findings for those working with similar hardware. I’m running an RTX 5090 with 32GB of VRAM.
1. Tooling & Troubleshooting
AI-Toolkit: I initially tried using AI-Toolkit, but it was a frustrating experience. It suffered from frequent, random freezes with no clear way to debug or recover.
Official Trainer: I eventually switched to the official Trainer scripts. Since the official scripts can be a bit finicky to set up, I used AI agents like Claude to help debug and refine the scripts. This made the transition much smoother and allowed me to get the environment running properly.
2. VRAM & Stability (Avoiding OOM)
To fit the training within 32GB VRAM, a few adjustments were necessary:
Disable Audio Module: This is a mandatory step to prevent Out of Memory (OOM) errors.
Resolution: I settled on 512x512x49. Anything beyond these dimensions proved unstable on my setup.
Other Settings: Followed the official recommended configurations.
3. Performance Metrics
Speed: \~0.58 steps/second.
Total Duration: 1500 steps took approximately 40 minutes.
https://preview.redd.it/ktmt9cljoazg1.png?width=1039&format=png&auto=webp&s=d2ac1f8234c5d822ffe0f479ca9937a1bf1ce3cd
4. Results & Conclusion
The primary goal of this LoRA was to capture specific repeating motions in 2D animation.
The results were very satisfying. While the base LTX model didn't naturally produce these specific movements, adding the LoRA successfully introduced the intended motion patterns. Interestingly, even though I trained at a lower resolution/frame count (512px, 49 frames), the LoRA generalized perfectly to high-resolution inference at 121 frames.
https://redd.it/1t4bbsi
@rStableDiffusion
I recently completed a training run for an LTX 2.3 LoRA and wanted to share my settings and findings for those working with similar hardware. I’m running an RTX 5090 with 32GB of VRAM.
1. Tooling & Troubleshooting
AI-Toolkit: I initially tried using AI-Toolkit, but it was a frustrating experience. It suffered from frequent, random freezes with no clear way to debug or recover.
Official Trainer: I eventually switched to the official Trainer scripts. Since the official scripts can be a bit finicky to set up, I used AI agents like Claude to help debug and refine the scripts. This made the transition much smoother and allowed me to get the environment running properly.
2. VRAM & Stability (Avoiding OOM)
To fit the training within 32GB VRAM, a few adjustments were necessary:
Disable Audio Module: This is a mandatory step to prevent Out of Memory (OOM) errors.
Resolution: I settled on 512x512x49. Anything beyond these dimensions proved unstable on my setup.
Other Settings: Followed the official recommended configurations.
3. Performance Metrics
Speed: \~0.58 steps/second.
Total Duration: 1500 steps took approximately 40 minutes.
https://preview.redd.it/ktmt9cljoazg1.png?width=1039&format=png&auto=webp&s=d2ac1f8234c5d822ffe0f479ca9937a1bf1ce3cd
4. Results & Conclusion
The primary goal of this LoRA was to capture specific repeating motions in 2D animation.
The results were very satisfying. While the base LTX model didn't naturally produce these specific movements, adding the LoRA successfully introduced the intended motion patterns. Interestingly, even though I trained at a lower resolution/frame count (512px, 49 frames), the LoRA generalized perfectly to high-resolution inference at 121 frames.
https://redd.it/1t4bbsi
@rStableDiffusion
Badass professional workflow - How High-Effort AI Usage Looks
https://youtu.be/--LJZeuN2PE?si=aps7FTS480hVcavu
The video shows how to create the initial and final frames of an animation, starting from the manual creation of an original robot to the creation of environments and 3D meshes to guide the various AI steps.
https://redd.it/1t49nyt
@rStableDiffusion
https://youtu.be/--LJZeuN2PE?si=aps7FTS480hVcavu
The video shows how to create the initial and final frames of an animation, starting from the manual creation of an original robot to the creation of environments and 3D meshes to guide the various AI steps.
https://redd.it/1t49nyt
@rStableDiffusion
YouTube
Gen AI Workflow: How to Actually Maintain Creative Control
In this video, I am going to be demonstrating a workflow that is an alternative to a lot of the AI videos you see online that make big claims, but in reality are just simple text-prompt-to-video generation that anyone with a keyboard can create.
I will show…
I will show…
This media is not supported in your browser
VIEW IN TELEGRAM
GTA 70s - Teaser Trailer: Z-Image Turbo - Flux Klein 9b - Wan 2.2
https://redd.it/1t4gjfj
@rStableDiffusion
https://redd.it/1t4gjfj
@rStableDiffusion
I built a dual-monitor image curator for sorting large Stable Diffusion output folders (looking for feedback)
Hey all,
After generating way too many images and struggling to sort through them, I ended up building a small desktop tool to handle it.
The goal was to make reviewing large output folders faster without breaking workflow.
Right now it lets you:
\- Tag images as favorites, junk, or other preset categories
\- Filter and isolate specific groups quickly
\- Jump through large batches (10 / 25 / 50 / 100 at a time)
\- Use a dual-monitor setup so one screen stays clean for viewing
I mainly built it because going through thousands of images in file explorer or basic viewers was getting painful.
It’s all local, no cloud stuff, just meant to be fast and simple.
Curious how other people are currently managing their image libraries and whether something like this would actually fit into your workflow.
Happy to share it if anyone wants to try it out.
https://redd.it/1t4rx5r
@rStableDiffusion
Hey all,
After generating way too many images and struggling to sort through them, I ended up building a small desktop tool to handle it.
The goal was to make reviewing large output folders faster without breaking workflow.
Right now it lets you:
\- Tag images as favorites, junk, or other preset categories
\- Filter and isolate specific groups quickly
\- Jump through large batches (10 / 25 / 50 / 100 at a time)
\- Use a dual-monitor setup so one screen stays clean for viewing
I mainly built it because going through thousands of images in file explorer or basic viewers was getting painful.
It’s all local, no cloud stuff, just meant to be fast and simple.
Curious how other people are currently managing their image libraries and whether something like this would actually fit into your workflow.
Happy to share it if anyone wants to try it out.
https://redd.it/1t4rx5r
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit
Explore this post and more from the StableDiffusion community