I've decided to create and share a minimal NSFW workflow for LTX 2.3 that I've been using for the videos I've recently produced.
All required models, LoRAs, and VAEs will be linked in the notes, along with workflow tips, recommended LoRAs, and example videos created using this setup.
The default settings are optimized for a system with 16GB VRAM and 64GB RAM. However, the memory footprint can be reduced further by lowering the image compression and resolution settings.
V1.4 - Made some adjustments to the workflow to make it easier to use. A list of the patch notes can be seen on the right for version details. Keep in mind the default steps have increased from 8 to 14 so if you see an increase in gen times that's why.
V1.3 - Required Custom Nodes have been reduced, FPS has been set to it's default value 24 FPS, and Additional Lora and Video Examples have been added,.
For additional background audio and effects I'd suggest checking out my other workflow for MMAudio. https://civarchive.com/models/2452337/slop-demons-mmaudio-studio
Description
Rearranged and color coordinated nodes.
Added Final Width and Height nodes to see what the final resolution will be after the Img Compression node.
Added a control panel section for frequently changed nodes.
Swapped out the manual sigma node for the LTXVScheduler node for easier changes to steps default step count is set to 14 up from 8.
Renamed/subgraphed some nodes to help keep things tightly organized.
Replaced Lora strength examples with video example links.
FAQ
Comments (12)
Can you also post the workflow/prompts for the actual starting image that you use?
The workflow for image generation is just a standard Anima workflow using the Lora's I created for it. Nothing special going on for it. For the image example I used all 3 of my Lora's at a 0.75 strength.
This workflow has a problem with its file referencing method. It attempts to reference the LTX 2.3 checkpoint (from \Models\DiffusionModels), the Audio VAE (from \Models\VAE), and ltx-2.3-22b-distilled-fp8.safetensors (from \Models\Lola) all from within the \Models\StableDiffusion folder. This makes it very inconvenient to use, as these files cannot be mixed into the StableDiffusion folder. Therefore, in my environment (ComfyUI v0.23.0), I have decided to give up using this workflow.
The workflow utilizes the same standard nodes used from ComfyUI's default workflow for LTX 2.3. The LTX 2.3 models go into Checkpoints both ltx-2.3-22b-distilled-fp8.safetensors and 10Eros are LTX 2.3 models. Audio VAE in the VAE folder is standard. If these are not appearing in your folder structure my guess is you're not using ComfyUI portable or added the folder StableDiffusion as it's not a standard folder with ComfyUI portable.
@SlopDemon
Thank you for the reply.
I understand that in the default ComfyUI Portable structure it may work fine. However, the workflow contained custom nodes (LTXVAudioVAELoader and LTXAVTextEncoderLoader) that had hardcoded relative paths such as LTX 2.3\VAE\... and LTX 2.3\.... This caused the models to not be found when using an organized folder structure outside of the standard models/checkpoints and models/vae.
I have already resolved the issue by manually replacing those nodes with UNETLoader, VAE Loader KJ, and Dual Clip Loader.
@hot_coffee Glad you were able to resolve it. LTXVAudioVAELoader and LTXAVTextEncoderLoader are Core ComfyUI nodes if you intend on using them in the future. The LTX2.3 you're referring to in the folder structure is just a folder I added into the Checkpoints folder. So it's Models/Checkpoints/LTX2.3/ltx-2.3-22b-distilled-fp8.safetensors for example.
is this most complicate workflow
Where is VRAM Limiter if I want to use all 24gb?
Thx <3
will be there something similar for minimax by any chance ?
At the moment there's not much to add to Minimax outside of speed nodes which hinder the output. If there's anything significant that requires its own workflow outside of ComfyUI's default one I'll create it.