@sexgod1979 Ah, man, I appreciate the reply and help for the novices. v1 was awesome, and I look forward to trying the update. The basic LTX template suited my needs just fine. I would love to get your new workflow working, but I just don't have the knowledge for the final missing nodes:

comfyui_custom_switch (1)

comfyui_layerstyle (7)

comfyui-various (5)

Let me know if there is anything I can do, but my knowledge is sadly lacking.

@sexgod1979 Ahh, thanks SG, I'm downloading again now. I'll try your new custom workflow again once I figure out the pip commands. I'll have to get ChatGPT to help me with that lol. I'm glad I could contribute in some small way. If there is a way to get you a coffee, let me know! edit: I just see the updated help files with git nodes.. legend

@sexgod1979 v1.3 isn't performing nearly as well as v1 for me. The audio is better, but NSFW content isn't nearly as animated (in terms of movement) as before. Also getting lots of weird morphing with the genitalia. When I go back to v1, the results are consistently better. I wouldn't have thought it was the workflow, who knows.

@sexgod1979 The direct download link also failed for me in HF, so I went with the XET link instead. It's all good, I have it now and look forward to testing it.

if you use Cloudflare or a vpn, disable that. that was causing my civit downloads to fail

I used the Comfyui-INT8-Fast nodes to convert the bf16 file myself and it works fine. I know comfyui has native support now, but I get much faster results using his load int8 node somehow. Hope the node author doesn't abandon his work just because comfyui implemented poorer native support.

Convrot int8 was just released half an hour ago, and gguf is a completely separate quantisation technique from the base model (bf16).

You might be able to use multigpu (https://github.com/pollockjj/ComfyUI-MultiGPU) on a gguf quant with distorch and vram capping of the unet, but problem is the clip; I haven't found any gguf of the embeddings and its dual clip node doesn't seem to work with gguf + non gguf.

Or this workflow. https://civitai.red/models/2443867/ltx-23-22b-gguf-workflows-12gb-vram?modelVersionId=2879736

@sexgod1979 I know two repos, city's and ferretduck's. City's is well established but long unmaintained, dunno if ltx works correctly.

https://github.com/city96/ComfyUI-GGUF/tree/main/tools

https://github.com/qskousen/ggufy (https://civitai.red/user/ferretduck)

There's also the grandparent llama.cpp, never tried it directly.

https://www.geeksforgeeks.org/machine-learning/how-to-convert-a-huggingface-model-to-gguf-file-format/

There is also this space, looks super simple but seems to not work on your repo. Probably only intended for llms. Suspect llama.cpp might give the same result.

https://huggingface.co/spaces/ggml-org/gguf-my-repo

```

Error converting to fp16: INFO:hf-to-gguf:Loading model: PinkCherry_NSFW_LTX23 WARNING:hf-to-gguf:Failed to load model config from PinkCherry_NSFW_LTX23: The checkpoint you are trying to load has model type `ltx-video` but Transformers does not recognize this architecture. This could be because of an issue with the checkpoint, or because your version of Transformers is out of date. You can update Transformers with the command `pip install --upgrade transformers`. If this does not work, and the checkpoint is very new, then there may not be a release version that supports this model yet. In this case, you can get the most up-to-date code by installing Transformers from source with the command `pip install git+https://github.com/huggingface/transformers.git` WARNING:hf-to-gguf:Trying to load config.json instead INFO:hf-to-gguf:Model architecture: LTXForVideoGeneration ERROR:hf-to-gguf:Model LTXForVideoGeneration is not supported

```

I didn't even realize that I could do this on my 4090. Been running all LTX 2.3 generations with tiled VAE. Switched it out and it worked fine. Quality is better, you are right! Thanks!

@sugarfreewafers Which VAE should be used? Please provide the link so I can download it.

@sexgod1979 Which VAE should be used? Please provide the link so I can download it.

@animart920 it's a native comfyui node (already included), you don't download anything. Just search for "VAE Decode" inside comfyui.

@animart920 - Use the VAE Decode comfy node instead of the VAE Decode (Tiled) node. Use LoadCheckpoint, load the sexgod-pinkcherry-v15*.safetensors checkpoint, and plumb the VAE from there to the VAE Decode node.

@sugarfreewafers ok... tanks

@sexgod1979 I've gone through most other image models and krea2 works fantastic. Prompt adherence, trainability, high def gens, and familiarity with most anatomy out of the box. SNOFS is my go to lora for krea. Use krea2_raw bf16 and a turbo lora if you can.

@sexgod1979 Would you consider training a similar lora for Krea2? Lots of good stuff in this one you've trained for LTX and based on my experience, LTX is much more difficult to train. Of course, I'm sure one difficulty is that your training data is likely mostly videos, not images...