Are you experiencing the benefits of anchor nodes?

A 12-second video with a resolution of 768*1344 will be approximately 5 minutes long in total.

@colaholic Are you aware that AMD GPUs don't perform the same way as NVidia? You've probably implemented ROCm, and while it's improved compared to the past, the AI ​​world fundamentally operates on NVidia's CUDA, so you'll likely encounter problems unrelated to specifications. I have an RTX 5070Ti with 16GB VRAM and 64GB RAM, but even with an AMD 24GB VRAM and 64GB RAM, it probably wouldn't run smoothly.

However, if you're getting errors with VAE Decode, replacing the node with "VAE Decode (Tiled)" is effective. This splits the video for processing, reducing the probability of errors.

I think it might be possible by cleverly tweaking the prompts, but I'm not sure if the finicky ltx2.3 will readily comply.

Without a detailed error log, it's impossible to tell, and frankly, I can't answer even if you ask me. The best thing to do is to put the entire error log into an AI like GPT or Gemini as a ComfyUI error log. I'm just doing the same thing.

@aldis54 Your Kjnodes version might be outdated. Try updating it through the ComfyUI Manager. Of course, update ComfyUI itself if you haven't already.

It depends on the environment, but in my case, I think it takes about 7-8 minutes.

The most likely culprit is Lora, but what do you think? LTX 2.3 is quite sensitive to Lora's effects, so it's best to avoid using Lora if possible.

Also, in my experience, Dasiwa's checkpoints and the 10eros node I use for upscaling samplers are incompatible, resulting in artifacts.

However, the problem of video degradation during upscaling is quite troublesome, and I'm also struggling with a similar issue, albeit in a different workflow. It's probably a fundamental problem with LTX 2.3.

Regarding this workflow, if you use the recommended checkpoints and Lora has the appropriate weight, I think there won't be much degradation during upscaling.

@local306 The creator of 10Eros states that artifacts occur with elongated sizes exceeding 1080p resolution, so this workflow cannot handle those cases.

Unfortunately, I don't know the cause for any other reason.

Unfortunately, this workflow cannot be easily implemented as is. You will need to either fine-tune the Sigma value connected to the sampler or swap out the nodes.

We are still unsure of the appropriate number of steps and Sigma value. This will be a challenge for the future.

Let me preface this by saying I don't have a solution.

However, I have confirmed that LTX 2.3 causes significant image drift when zooming in with the camera, resulting in unrecognizable people and various other problems.

I'm not entirely sure about the intent of your second question, but Upscale Model typically upscales to 2x or 4x first, and then reduces it. Therefore, if you choose a 4x model, for example, and the original resolution is high, the resulting video will be even larger, so it's not surprising that it might get stuck. You might need to lower the initial resolution setting.

@severelyAI We're both using translation software, so don't worry about it. Maintaining facial consistency is a challenge with LTX 2.3, so I don't think there's a clear solution. (In other words, it's a problem with the LTX 2.3 model itself.) However, many people are trying various methods. With my generation method, I think this workflow strikes the best balance between consistency and generation speed, but I think everyone will have a completely different opinion. I recommend trying other workflows as well.

@severelyAI Regarding the first point, the 10eros node has a unique mechanism where it runs the sampler twice within a single node (which essentially serves the same function as a 3-sampler workflow). I believe this is why it appears the way you pointed out.

​As for the prompts, I didn't use an LLM. To be honest, I'm just reusing the prompts I previously made for Wan, and I haven't gotten around to experimenting with LTX-specific prompts yet. If you don't mind that, I have disclosed all the prompts for my videos, so please feel free to check them out on my profile.

@severelyAI You are very knowledgeable. I'm doing this with no knowledge at all, so if you have any good suggestions for improvement, please let me know.

Please update KJNodes to the latest version. If the problem persists, I cannot help you.

@xiayu2200629 Ah, that's the cause. The fact that there's no input to the left of the PatchsageAttention node means the model hasn't loaded correctly in the first place. This also means the LtxvImgtoVideoInplace node doesn't have a model input either.

Please check that the correct model is selected in the model area at the top left of the workflow, or that the connection isn't broken.

As it is, it is not compatible, so please replace the diffusion model and text encoder nodes in the upper left model area with those for GGUF.

Please do it yourself.