I'm currently doing it as we speak. Will be published in maximum 10 minutes.

I kept the same system as the official LTXV2.3 workflow had at the time, it was using the image loader to define the latent size of the T2I output, so even in T2I you need an image. I didn't change that because I'm not using T2I, I didn't think about this sorry. I don't know If I can give a full control only on the T2I latent with this configuration, but I'll try do do it on the next update.

About SageAttention, if I kept it active it's a mistake, sorry again. You can deactivate it inside the subgraph on the checkpoint loader.

So, I checked and I think I can give you full control of the T2I latent size quite easily. Would you mind to come into my Discord server to test that function for the next version?

@laogetmd862 On my profile, under the banner

@laogetmd862 @jjd9 I added the file to download there on this page, you have now a config file directly under the archive.

I did an important correction, I forgot to divide latent size for first pass, you'll maybe want to download again the config file.

I don't know about your error, I need to precise that I upgraded everything in ComfyUI befor releasing the workflow.

But I can tell you wich node is doing that. Try to delete the RAM et VRAM cleaner inside the 2 and 3 stages subgraph, and also the pair behind the final video output. If the error stays, delete the "Soft Full Clean RAM and VRAM [LP]" in the "⚡ Save Prompt" group.

Sometimes it's because of the Clip, try another version of Gemma3.

@Renessance Set the language source and target in English, prompt in English and when you want someone to speak, prompt "She or he says in Russian: "A phrase in Russian". It works like this. Otherwise delete the translator and replace it with a basic text box.

@fatestaynighthyoga84848 Hi! It's because you try to load a diffusion model through the checkpoint loader. Some models are checkpoint merged (with VAEs and Clip inside), and others are just diffusion models needing to load VAE and clips aside.

If you did download the 10Eros v1 at the release, redownload it, I think TenStrip changed the fp8 file.

Quote: "I'm uploading a new FP8 version. No split files, both load as a checkpoint like normal. New FP8 version is very similar to bf16 with minimal loss." From TenSTrip 5 days after release.

I don't understand your statement. You don't need Sulphur dev if you only use the GGUF group.

Sometimes subgraph have some conflict issues with other nodes, if you have problems running the groups, expose the subgraph taking out all nodes and try to run again.

If you don't need the Safetensor group, delete it with the switches inside the subgraph.

@Sgt_Fartz gemma-3-12b-it-abliterated_lora_rank64_bf16: Allows you to use a Non Ablitareted Gemma clip and "uncensor" it with this LoRA. This is in a case you don't use a Abliterated version of the Clip.

ltx-2.3-22b-distilled-lora-384: Official distilled LoRAs from Lightricks, good enough but from my experience, it can drift with faces and textures.

ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe: Custom distilled LoRAs from TenStrip (Who shares the Eros model on Civitai). More stable with my usage, no drifts using them at 0.5 on both pass. Quote from Tenstrip:

!!!!! The LTX2.3 Distilled LORAs are detrimental to your outputs. I worked on and fixed this problem before you guys experienced it by severely reranking and gutting the distilled loras. There are many long complex reasons why, but simply put: it was most likely trained on T2V latents produced by the base model. Now, this is a finetune I2V producing model, that distillation has the opposite effect. You've probably noticed it many times and thought it was just the model. The 384 1.1 official version and even the other reranked versions reduce trained anatomy accuracy, reduce sexual motions or any trained motions not in the base model, and they add motion blur/choppy edges by disturbing feed-forward layers (that is a serious issue Lightricks needs to fix). They are also insanely overkill for what they need to do. My reranks are linked there, especially the cond_safe versions which have conditioning and feed-forward zeroed out letting them produce much better with sharper motion and upscales with a 1.0 first pass and 0.5 upscale pass, or 0.65 single pass.. Use those reranks with this model and Sulphur T2V and I2V... or don't read any of that and use the normal distilled lora and then leave a comment about all the issues it causes.

@maumau22 Sadly not for the moment. It was the first thing I tried to do with LTXV but I encoutered 2 major problems:

1. I was not able to find a way to batch and concat audio through each loop correctly.

2. The point was to loop 25-30 seconds segments together, but more then 2 segments is too heavy to keep image and audio in memory during the loop.

I'll do it at some point if I find solutions.

@raflrafl This can be many different things:

First of all, try to change your distilled LoRAs, sometimes I get stuttered audio with TenStrip cond_safe distilled LoRAs.

I also suggest now to run at 25fps and bypass the slicer group (the slicer group have currently a math logic error running at 25fps).

Run without LoRAs to check (other than the distilled ones), sometimes some LoRAs have bad audio.

The next version of the workflow will allow users to choose between VAE in the checkpoint or seperate ones to have more control.

@ericshaler902 Probably nodes conflict. To check, disable all custom nodes. Restart, enable only workflow's nodes. Restart, check.

---> Sliders display correctly? ---> Nodes conflict.

----> Still not displaying correctly? ---> Update as many thing as you can before retry.

I updated everything in my Venv (ComfyUI, custom nodes, etc...) before deploying the workflow.

@ericshaler902 Oh right, I forget about that, I don't use Nodes 2.0 cause some creators said that they never will make a node update for that. Nice to know thank you !

@Pohtaetoes your error doesn't come from the clip loader, it's because of the subgraph. Take your GGUF group outside the subgraph and it works normally. I'm still trying to figure it out why this behave like that.

@AnonNameMetoo I didn't update for now but if you have the same problem, expand the model subgraph to take out the nodes, it will resolve the error. Then it's just less esthetic but it's working.

@tauroi748

To be honest, I never thought about generating without audio on LTXV2.3; the workflow wasn't designed for that purpose. Unfortunately, I don't know if it's possible to generate without audio, but you can simply exclude it from the video:

1. Replace the video player with VHS Combine.

2. Retrieve the video link from the "Create a video" node (the audio link has to be connected to a dummy node) and connect it to VHS Combine.

3. Delete the "Create a Video" node.

In this configuration, you'll only have the video without the audio.

But wouldn't it be simpler to use WAN2.2 for your situation?