Should work work without updating comfy, if your last update was march 7th. The RTX workflow might require a comfy update tho. I usually have a 2nd comfy install with no models, just to test comfy updates before updating my main comfy installation.

Hi there, if you want to use LTX-2.3 GGUF model, the only thing you need to do is download it and replace the checkpoint loader node to a unet loader node and let it load the GGUF model, anything else (vae, textencoder, loras, etc) remains the same.

The vae is only embedded in the LTX-2 version, for the LTX-2.3 version, everything is seperated, as per the download list.

Supposed to work with 16g vram. With rtx 5 series you probably can use a smaller fp4 gemma model. I think the comfy ltx template shows a dl link for that. But i doubt that is the real issue. What does comfy say when crashing?

@RandoWando the T2V workflow uses the same loader for textencoder like the I2V workflow. Maybe compare and double check the node show both gemma and text_projection are set. try rightclick and reload node.

Hi, you can see from the examples posted by me and other users, those do not look terrible or disastrous, you can even download the clip and throw it into comfy as those contain the workflow, prompt and settings in the metadata, if you get terrible results with it, your setup might have an issue …

@BocekAdam You can chose between distilled and dev model, both have pros and cons. Dist is faster, delivers better image quality, but lacks a little of variation, for me Dist wins due to speed, and I am ok with it´s quality, almost 95% of clips I posted are with Dist model.

Dev. model is more ore less the opposite. Both have a bit of issues with facial consistency. Think there are Loras that can help with that.

@BocekAdam If you use the workflow with RTX Video Super Resolution, it is required to also set the final resolution to the desired width/height there, it is the black node on the right above final output.

You can also switch RTX to "resize by multiplier" to like 2. With your 760x1280 you can then get double resolution as output.

Edit: The render process works with 2 sampler in sequence;

1. sampler 1 creates the motion, the scale_by factor defines at which resolution the render is done. 1 = uses the full resolution you have set in width/height (very slow). Scale_by 0.5 renders with half the resolution (=default).

2. sampler 2 upscales the video (always x2) in 3 steps with distilled lora. It is a spatial upscaler that has much more info for scaling as it works in latent space vs. the usual upscalers, that just upscale pixels.

3. RTX VSR upscales video to a final resolution very quick.

You enter a prompt and it is sent to ollama, the group on the left, that is labeled as Ollama.

There is a switch to toggle ollama off, so you can run you own prompt without ollama, it is on by default.

Hey. did not test a lot of Loras with PowerLoraloader, which is set after the detailer lora. You could try to load the detailer in powerloder, after your lora. If this still not works, then the detailer (which is a LTX-2 Lora), might not like additional loras...

Workaround, as you say, would be to switch detailer off on main process, when using other loras. Or maybe lower strength of lora.

meanwhile tested some loras and it seems like the detailer loras need to be off or with lower strength (<0.5) on both main process and upscaler, when using loras in the power lora loader node.

Also the upscaler process requires a lora processing too for some loras, which is currently not setup. To fix, easiest way is to add another power lora loader node after the detailer lora for the upscaler and connect it with the subgraph for the sampler.

Hi, good point. I think comfy will be updated to have a proper LTX2.3 preview.

Until then here is a workaround:

From Kijas nodes use "LTX2 Sampling Preview Override" and hook it up between powerlora loader output and Guider Input (within Dev or Distilled pipeline on the left), so connect model in/out. It will then extend the sampler subgraph with the preview, you might need to move the Prompt node below out of the way.

You can add this vae to have the preview less pixelated: https://github.com/madebyollin/taehv/blob/main/safetensors/taeltx2_3.safetensors

Here is a screenshot: https://github.com/kijai/ComfyUI-KJNodes/issues/566#issuecomment-4016594336