Oh, I hadn't noticed there's a new version, thanks. It seems that the last hotfix was yerterday though, I will wait a few days before updating.

The patched workflow is ready for downloading.

This errors are related to LTX Director new version. The new version released very recently and I wasn't aware of the update, I'm currently hot fixing the LTX Director workflow to upload a hot-fix. It won't take full advantage of the new features of the node, but it will work as it was working before, until I upload an updated version adapted for the new version of the custom node.

The patched workflow is ready for downloading.

What workflow are you using?

@recovry777524 Make sure that you turn custom audio off in the LTX Director node, if the audio is muted, maybe you have it turned on, if the option is on and no custom audio is added to the LTX Director node time line, the final audio is completely muted. To use ID LoRA in fact, no audio should be added to the timeline, and custom audio should be off. Hope this helps.

@recovry777524 Can you share the original generated video here in CivitAI? The original video will contain the workflow metadata, and I can check it out to see what's wrong

@recovry777524 You have "Lipsync enhancer" enabled, that group is only usefull to extract the voice from the custom audio you add to the LTX Director timeline, if no custom audio is used, this group should be disabled. Sometimes for example, you will use a custom audio that has noise or background music and the voice part won't be very clear, then the model could struggle doing a clean lipsync, this is the kind of scenario where "Lipsync enhancer" can help you out, because it will try to separate the voice alone and use that to create a clean audio latent. So for this specific case, it won't work. Since you are not using any custom audio, an empty wave is passed to the voice separator, and, as a result, a muted audio latent is created (not empty, but muted), since that audio latent is muted, the model creates a muted video as a result. So try disabling "Lipsync enhancer" and see if that fixes your problem.

I'm working on the hotfix, it is almost ready.

@ywen35951247 

The patched workflow is ready for downloading.

Yes, sorry, the LTX Director node got updated and the workflow got broken, I got the fixed workflow almost ready.

The patched workflow is ready for downloading.

This is an all in one workflow, as more features are directly supported within the LTX Director node it self in the future, I will be getting rid off the redundant parts. For now, not everything is supported withing the node, like ID LoRA, two pass processing, or LTX Detailer.

Thanks for the comments. Sorry, I don't have any kind of guide on the workflow, I will consider doing one though, thanks for the feedback.

@viniandrieu845 Unfortunately, there is no ID LoRA for that yet. To generate videos with 2 specific voices, I recommend creating the dialog separately with a dedicated text to speech model and then use that as a custom audio.

In the downloadable zip file, there are 3 workflows, one legacy workflow (that works without the LTXDirector node), and 2 LTX Director workflows:

- One works with the 1.3 version
- The other works with the new v2 version
Choose whichever fits your current setup.

For what you say, I will assume that you are talking about the LTX Director updated node, so try the file named "LTX23LTXDirector2.json", that's the workflow working with the new version.
Another thing, I would recommend avoiding installing custom nodes directly from their github repos, it is better to use ComfyUI manager instead. When possible, avoid git clone, and only use it when the node is not available in the ComfyUI Manager node list.

Now, if the problem you have is not being able to load old videos with the new update, I'm sorry, that's an issue with comfyUI and not my workflow, it is just how updates work. What you can do is inspect the json metadata from the video and check the settings there, or extract it using ffmpeg and upload it to your favorite LLM to ask for the settings you need.

Use the resolution master node to input a custom width and height. If your resolution is not an standard 16:9 video, then, in the "Decoding" group, there's a "Resize Image v2" node, where you can adjust the final look of the video, by default is set to "keep_proportion" crop, try stretch and see if that fixes the stretched look. This is an issue I need to properly fix to make it automatic, but for now, I hope this manual fix helps.

You need to vary your dialogs, avoid caps, don't do anything that can resemble a movie script, do something more natural, descriptive and casual as if you were telling a friend a story, like:

The man enters the room and he says "I really like this place, it gives me good vibes", then he heads to the bar and says to the bar tender "A cold beer pleace".

@Arcadeath You are right, It is already in the negative prompt (at least if they are running the workflow as it is uploaded). The suggestions I made are based on my own experience, I generate dialog heavy prompts and I have had the exact same issue in the past, most of the time is the structure of the dialogs.

What workflow are you using? what do you mean with "Keyframe 1 nodes" (in plural), there's only one Keyframe 1 node in the legacy workflow. Maybe you can upload the original video here to civitai and I check the metadata to see what's going on.

Check your global prompt, if you are using the old version (1.3), the global prompt might be hidden but active. If it has something there, delete it or replace it with your own prompt.

@AzextonHunter This sounds like the text encoder is not being capable of understanding the system prompt correctly, just for future reference, what text encoder are you using?

You need to use the right workflow for the right version of the LTX Director node you have, 1.3 or v2, both are included in the zip file, but you need to use the right one.

Thanks for the comment.

That's actually not a bad idea hehe Thanks for the Buzz by the way :)

It is not possible, LTX needs the audio latent no matter what. You can artificially create a muted audio latent, but it could mess with your video, since the video is created on top of the audio latent, for example, if your video contains characters "talking", they will look like mimes or lifeless dolls, besides, it won't accelerate your generation, since even a muted audio latent will be processed regardless. If you want to generate only video, I can suggest the Wan2.2 model. Or, you can disconnect the "audio" input of the "Video combine" node from the "Save video" group to save your videos without audio, but again, this won't accelerate your generations, since the model will still process the audio anyway.

Unfortunately, this is a known issue with LTX2.3. The two most effective ways of forcing character consistency are:
1. Training a character LoRA
2. Using keyframes to lock the face in several parts of the video, at least in the first and last frames.

What variant of the workflow are you using?

@elmejor9369 I'm not sure what variant of the workflow you are using, but for the LTX Director versions, lip enhancer should only be used with custom audio ON, and with a custom audio added into the timeline. If custom audio is on, but no custom audio is added, the resulting audio is an empty track. Lip sync enhancer should be used only for that, lipsync enhancement, it is only useful to extract the voice from the custom audio you add to the LTX Director timeline, if no custom audio is used, this group should be disabled. Sometimes for example, you will use a custom audio that has noise or background music and the voice part won't be very clear, then the model could struggle doing a clean lipsync, this is the kind of scenario where "Lipsync enhancer" can help you out, because it will try to separate the voice alone and use that to create a clean audio latent.

Don't know what's going on, I got no experience in Runpod. All I can tell you is that in my own 4090 system, you can generate 50 seconds videos at 7020p using half res and the 2x upscaler. Granted, there's tons of offloading going on, in some instances both my RAM and VRAM reach peak usage.

@rahulr1199 No, I don't use those 'cause I don't need them. My system has 64GB of RAM so offloading is enough in may case. The 4090 alone is not enough for models like for example 10Eros, the new version is more than 30GB in size, so extra RAM is a most. You can run models bigger than 24GB but you need the extra RAM to avoid the OOM erros. Have you tried the GGUF version?

@rahulr1199 Yes, a really good one is this one:

LTX-2.3 Riding pov (I2V)
This LoRA improves movement and jiggle physics, despite its name, you can use it as a general purpose NSFW LoRA, with values between 0.4 and 0.8.

@rahulr1199 LTX sucks at character consistency, two ways of somewhat solving that issue:
- Training a character LoRA
- Use keyframes to lock the face in several parts of the video, or at least first and last frames.

Regarding controlnet, I honestly don't use controlnet for nsfw content, so I can't recommend a good source for that.

@rahulr1199 Sorry, I got no idea how runpod works.

@rahulr1199 Unfortunately, the movement and character interactions lack behind the 10Eros model, you can push it a little bit with LoRAs, but they won't make it as strong in that field as the 10Eros model. Whatever 10Eros version you are using now, try the other one and see if that one does better.

What variant are you using?

@Paradantoz It could be that the segment lengths are too short for the model to smoothly fill the gap between them and/or the segment prompts are not specific enough on how to transition from scene to scene. I recommend watching the author's tutorial:
https://www.youtube.com/watch?v=vM60pJJqqEI

My source images are high resolution, minimum 1080p but many are in 1440p or 4k. Then the videos are rendered in a base resolution of 720p and upscaled to 1080p. But the quality comes mostly from the reference image generation itself, bad quality images will result in bad quality videos.

Yes, the "Base prompt" is the "Global prompt", so use it like that, like the global prompt, and use the timeline for your segment prompts. For basic videos that have no segment prompts, like a basic text to video or a basic image to video, you can use the Base prompt alone.

@PrimeInMe Exactly, the global prompt (Base prompt) is the main guider of your video, that's why in most cases, you only need the global prompt to create your videos.

My workflows are not designed to run on online platforms, sorry.

As a general rule of thumb, you should look into the size of the safetensors file and see if it fits into your VRAM size, so choose the file that is under 12GB in size. This is not the most correct way of calculating things, but it will be enough till you learn how offloading works and what's the best model for your setup. ComfyUI's offloading system allows you to load bigger models than your VRAM, but to keep things simple, follow that rule for now.

For the dev version you need to add the distilled LoRA into the model pipe line, just like the 10Eros model. Note that in the list of suggested models, I recomend the distilled version, so you can use ltx-2.3-22b-distilled-1.1_transformer_only_int8_convrot.safetensors if you don't want to change anything in the workflow.

@drfaker911219 The segment is too short for the action imo, you could get it but you need to try different seeds, normally I generate 20 videos at a time and then choose the best. Yes I've noticed the audio issue, for now I fixed it by adding a normalization node, the issue is that the audio is saturated, so if you reduce the LUFS it sounds normal.

@drfaker911219 Not at the moment, the node I use to reduce the LUFs is the "Normalize Audio Loudness" node from the Wan video wrapper pack, and I use a value of -34 to -14 depending on how loud the final audio is.

@drfaker911219 You can connect it right before the final "Video Combine 🎥🅥🅗🅢" node, in the "Save video" group, between the "FINAL_AUDIO" get node and that video combine node.

Here's the LTX2.3 official prompting guide:
https://ltx.io/blog/ltx-2-3-prompt-guide

I don't know who she is so no hehe

Not in the current state, but you can add the preview model into the pipeline if you need it, then the samplers will show the preview, that's how other workflows do it. LTX is very fast though, so normally, you don't need it, since you can simply general several videos in one go at half res and choose the best of them for further upscaling and interpolation.