This workflow is an adaptation of the one shared by @tenstrip, the creator of Ltx2.3 10Eros. I want to express my sincere gratitude to them for releasing such an incredible checkpoint and workflow!
Update History
v1.1 - I changed the default OmniNFT Lora weights. I haven't changed the workflow itself. The First - 2.0 and Upscale - 0.7 values were excessive and restricting functionality, so I changed it to First & Upscale - 0.8. I had changed the values from @tenstrip's original workflow, so I reverted them.
When I first tried the official Ltx2.3, I wasn’t very impressed. However, I’ve found that using this 10Eros checkpoint and workflow allows you to maintain facial consistency remarkably well. While the character movement might feel a bit less dynamic compared to wan2.2, in my experience, it’s actually a huge step up from wan2.2's SVI. Plus, it generates video with audio, features faster generation times, and supports video lengths up to 20 seconds. Because of all this, I’ve decided to officially dive headfirst into generating with Ltx2.3!
That being said, even a tiny change in settings can completely alter the results. What I’m sharing below is simply the specific configuration that happened to work for me this time. I haven't actually tested LTX 2.3 extensively yet. I'm putting this workflow out there in hopes that the community can test it, play around with it, and help improve it further. Current video generation presents a dilemma: maintaining consistency and creating dynamic movement are often at odds. The goal is to strike a good balance.
Just a heads-up: I am by no means an AI generation expert. Everything here is the result of pure trial and error, based on hands-on testing and digging through a ton of community information. Because of that, I won't be able to answer highly technical questions. I really only know what I’ve personally experienced. While I have some experience with WAN 2.2, I’m still a complete novice when it comes to LTX 2.3. If you run into issues, I'd highly appreciate it if you could troubleshoot them on your own.
【Reference Workflow】
10Eros_10SNodes_LikenessGuideHelper_I2V_v3.2.json
https://huggingface.co/TenStrip/LTX2.3-10Eros_Workflows/tree/main
【10S-Comfy-nodes】
https://github.com/TenStrip/10S-Comfy-nodes
In addition to this node, several other nodes may not be installable due to ComfyUI Manager's security level. In that case, please lower the security level of ComfyUI Manager or obtain it from GitHub.
●The safetensors model required for this workflow
【Checkpoint (VAE included)】
10Eros_v1-fp8mixed_learned.safetensors
https://huggingface.co/TenStrip/LTX2.3-10Eros/tree/main
【Text Encoder】
Gemma_3_12B_it_fp8_scaled.safetensors
https://huggingface.co/Comfy-Org/ltx-2/tree/main/split_files/text_encoders
【Upscaler】(Be sure to select "1.1". "1.0" has a bug.)
Ltx-2.3-spatial-upscaler-x2-1.1.safetensors
https://huggingface.co/Lightricks/LTX-2.3/tree/main
【Distilled Lora】
Ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors
https://huggingface.co/TenStrip/LTX2.3_Distilled_Lora_1.1_Experiments/tree/main
【OmniNFT Lora】
OmniNFT_converted_lora.safetensors
https://huggingface.co/VasiliyWeb/OmniNFT_ComfyUI/tree/main
【Edit Anything Lora】
Ltx23_edit_anything_global_rank128_v1_9000steps_adamw.safetensors
https://huggingface.co/Alissonerdx/LTX-LoRAs/tree/main
Downloading all the files will bring the total size to over 40GB. To run Ltx2.3 smoothly, your PC needs at least 64GB of RAM unless you're using a top-tier, expensive GPU. Right now, ComfyUI relies more on system RAM than GPU VRAM because it offloads any VRAM overflow to the RAM. Even with that offloading, it likely won't run on anything less than 32GB."
In fact, I hit OOM errors myself depending on the resolution and video length I tried.
Quick Tips:
Resolution Matters: Unlike wan2.2, Ltx2.3 struggles to maintain facial consistency if the reference image resolution is too high. In my tests, 720p to 1080p yielded the best results, whereas 1440p (2K) didn't work well.
Be Specific with Prompts: Your prompt descriptions need to be much more detailed than what you'd use for wan2.2. While audio is obviously a requirement, the model tends to completely ignore anything not explicitly written in the prompt, making it less flexible. Vague instructions won't cut it here. @tenstrip has put together a great summary on this, so I highly recommend checking out their guide for more details!
Troubleshooting Node Errors: In recent versions of ComfyUI, you might run into occasional node bugs. For instance, even core nodes like "Load Image" can sometimes show up with a red border and be marked as "Missing." In this workflow, you might encounter a persistent red-border error with nodes like "LTX2 LoRA Loader Advanced," even after you've installed them. If this happens, please delete the error-producing node and simply recreate/reconnect it manually. Replaced nodes should function properly and resolve the issue.
Description
v1.1 - I changed the default OmniNFT Lora weights. I haven't changed the workflow itself. The First - 2.0 and Upscale - 0.7 values were excessive and restricting functionality, so I changed it to First & Upscale - 0.8. I had changed the values from @tenstrip's original workflow, so I reverted them.
v1.0 - This workflow is an adaptation of @tenstrip's workflow. Maintaining facial consistency is the top priority.
FAQ
Comments (26)
Thanks! Is there a reason you like to use the edit anything lora in this case?
It's mentioned on Civitai's 10Eros page.
From what I've researched, these editing Lora versions are very good at preserving the structure between frames and maintaining the "space and camera motion of the original video," so this property is utilized.
However, I'm not sure if this is the correct way to use it. I'm using it this way because the results seemed good.
@kenpechi Aaah interesting, yeah sounds like worth experimenting with. Thanks.
I need to ask what the OmniNFT lora is for? What does it do?
OmniNFT LORA is a LoRA implementation developed to improve the video and audio quality of Ltx2.3, as well as their synchronization performance (synchronization rate).
In other words, its main objectives are to correct the discrepancy between audio and video, improve audio quality, and reduce degradation over time.
The reason the First Pass is set to 2.0 is because this usage was introduced in the community, not because I was absolutely certain of it.
kenpechi has entered the battle field ! 🎺
I came to the joint research lab. lol
@kenpechi If you want to research more, Dr34ml4y also shared some workflows, that I found useful.
https://civitai.red/models/1811313/dr34ml4y-all-in-one-nsfw-wanltx2
Oh, this is very heavy for my 5060ti 8gb. I used the workflow from the creator of 10eros, and my graphics card was able to generate it in 5-6 minutes. However, this workflow takes 20+ minutes to generate, which is very time-consuming. I am not very familiar with creating workflows, so could you please guide me on how to configure the preview display? It's just that when it takes 20 minutes to generate, I want to be sure that the output will be as specified in the prompt, and I want to skip if something goes wrong. You might say, "What's the problem with waiting for the first pass?" Well, even the first pass takes about 10 minutes.
I think you need to add a node related to the preview, but sorry, I don't know. Try looking it up yourself.
Enter first pass subgraph and add LTX2 Sampling preview as shown in the example. https://imgur.com/a/t6YwBYl
It is a heavy workflow, so this will make it slower too. Preview rate of 8 is kind of fine, but usually people do around 12-16 to have a better view. Of course that would make the workflow heavier. You can add the same node on the second pass too.
I've tried installing the LTX node both through the ComfyUI manager and manually, but it never installs. I always get "import failed" and "try fix". Could someone please help me?
You probably need to do a complete reinstall. It is the way of the Comfyui.
@jbear2447 Hi, thanks for replying. I've already tried that several times, but it didn't work. I'm using Easy Installer, and it's the latest version.
That's the extent of my knowledge buddy...did you try turning it off and on? LOL
Go into github repo, click on code and copy the URL.
Delete the custom node folder. Open cmd and git clone the repo.
Run comfyui, check for errors and if you have any, just give it to chatgpt. Sometimes it's just a requirement that wasn't added to the list
LTX nodes often have dependency and version conflicts, so it might be easier to start with a fresh ComfyUI installation and set up the environment again from scratch
I'm using StabilityMatrix with a ComfyUI venv setup.
ComfyUI version: 0.20.1
Python: 3.12.10
PyTorch: 2.12.0 + CUDA 13.0
If your having problems with that module try this in your venv environment :
pip uninstall -y kornia
pip install kornia==0.6.12
it worked with my install of Stability Matrix
@Trollmale101 This worked! I think the latest version of Kornia doesn't work with this node.
@lolbleach001584 Yeah looks like it, glad it worked
Optimized Fast creation speed It's impressive that it delivers smooth image quality and vivid colors without facial distortion
You're the best, Kenpechi!
Great job!
Too slow to be usable. Runs like absolute shit on my RTX 5060 Ti 16GB. Takes a year to make a low-res clip.
The fact that it's unusable due to insufficient system RAM is probably the reason, but it certainly does take longer than the standard workflow. I'll take this into consideration for future improvements.
Thanks for your feedback! Please use a different workflow.
@kenpechi 64GB of DDR5 at 6000 MT/s. Not the RAM. It takes like 2x-3x as long as a regular workflow. Not worth, imo, but that's subjective.
@aurelius Oh, you're pretty strict. Well, you don't have to use mine, you know.