Can you post a comfyui template that has settings for the ltx-2.5-22b-dev-transformer-bf16
First I want to thank you for releasing this model for the community, I've studied this for the past week, as newbie trying to figure these video models, and each day I've found something new about it, like this model can take a reference card with characters, objects and locations and use them to create a video. nice.
but I have a very big problem that I can't figure out how to solve it. I've even tried using and LLM to figure this out but unfortunately it gets bugged down in how the github examples are build and doesn't have the correct advice.
HOW do I make the full ltx-2.5-22b-dev-transformer-bf16 work, what are the settings? or even better just give us a full power workflow for that.
What I've figured out myself till now is that I need to switch the manual sigma to the LTXVScheduler (that is intended to work for the DEV model) and then figure out the settings for that node and also structure the rest of the comfyui to not bottleneck the entire thing.
Attempting to use the github distilled workflows as a base had been a failure for me. I understand why those templates exists, those are great for 4090s for example, or 5090s but can you please post a full power comfyui that gets as input a full HD frame and creates the video, regardless of VRAM limitation? I'm assuming an RTX Pro 6000 with 96 Gb of vram could at least let you generate a minimum amount of seconds.
Thank You!
I can't help with the scheduler settings, but here is the memory side.
The bf16 dev transformer is about 42 GB, so with working memory it needs about 47 GB of VRAM.
It does not fit fully even on a 32 GB card. ComfyUI can still run it by keeping part of it in system RAM,
but then you need a lot of RAM (around 64-96 GB) and it gets slower.
On 24 GB or less, a GGUF file is the easier way. The GGUFs I know of are made from the distilled model, not the dev one.
Sizes and what fits each GPU: https://vramready.com/models/ltx-2-5/
my main problem the full dev model doesn't uses the default fixed sigma values but a separate node (is in the comfyui node list) LTXVScheduler and some settings I don't know yet. Also it may be that there are other nodes or some other way to connect it. Basically the workflows I've found are for the other two models, the smaller 21 gb and the 42 gb distilled.
Regardless of VRAM, just assume I have 200+gb vram that is the thing I want. the vram doesn't matter for this question. I've tested the 42gb on 5090 and it worked, it likely it will work with the full dev one but will just take more time. I've also tested the existing workflows on rtx 6000, that one has enough ram to have all full in memory. Basically I just want the option to have full quality, actually I'm trying to find a solution to distorted "fish lips" when people are talking, maybe the full version can animate the face better that is my goal.
Thanks for the details. I haven't run the dev model myself, so I don't have tested scheduler values.
This is what I found in the official docs:
- The fixed 8-step manual sigmas are only for the distilled model. The docs say the LTXVNormalizingSampler is "NOT tuned for the full model".
- The full model needs CFG above 1. The MultimodalGuider node does that (their suggested start: modality 1, skip step 1). It is slower.
- The advanced example workflows are here: https://github.com/Lightricks/ComfyUI-LTXVideo/tree/master/example_workflows/2.5
- Node docs: https://docs.ltx.io/open-source-model/integration-tools/ltx-comfy-ui-nodes
If someone from Lightricks can share the dev settings, I would like to know too.