LTX 2.3 I2V | LLM Enhancer | T2I -> I2V | Detailer/Upscale
Features:
Turn any Text to Image to a Image 2 Video.
Z-Image T2V:
Uses an initial relative simple prompt
Enhances the prompt via Ollama
Image selection (if more then 1 image is generated as draft)
Details and Upscales (via SeedVR2) the selected image
LTX 2.3 I2V:
Image and (short context) pro,mpt is sent to Ollama for enhanced prompt.
Upscale sampler
Ollama part can be skipped by flipping the true/false switches. Same goes for the T2I part if you want to supply your own image.
Runs without issues on a 12GB GPU with a Q8 GGUF model.
Note: For both the T2V and I2V Ollama enhancement, I'm using the qwen3-vl-abliterated-8b model myself (still need to evaluate a Gemma based model).
Description
FAQ
Comments (8)
Are these examples using the full model? fp8/Q8? I assume not distilled because the quality is too good.
All with regular Q8 GGUF (not distilled). Most with the Vantage one (from https://huggingface.co/vantagewithai/LTX-2.3-GGUF/tree/main/dev), some others with Unsloth.
@dutchit288 That's a relief. I grabbed the Unsloth one. Do you think it's about the same as Vantage? or are you team Vantage now? Thanks for answering these questions btw.
@Jellai The Vantage one does seem to be more consistent on longer (30 sec) videos, but haven't really compared enough to say that reliably.
Can I ask how 12 gb vram system manage to run around 24gb Q8 gguf model?
also, how long does it take to make a video? (also provide length and resolution info pls)
You'd probably need to ask ComfyUI (and/or LTX) developers for the first one :), but it runs quite fine. Newer ComfyUI utilizes VRAM offloading.
As far as generation time, length/resolution: Usually between 10 and 14 minutes for up to 30 seconds with 1408x769 and close to 20/25 minutes when pushing it to 40 seconds (any longer and quality seriously degrades in my experience).
@dutchit288 can you tell me which exact models do you use in this workflow including z-image model if it is no problem? I will download and test workflow with the same settings you posted with your workflow. For example; I couldnt find Z image model that shows in the workflow (pornmaster one)
For the Z-Image model it shouldn't matter much, but I https://civitai.red/models/2270401?modelVersionId=2700412 , https://civitai.red/models/1609320/intorealism?modelVersionId=2835157, and https://civitai.red/models/2384856?modelVersionId=2868056 are my usual choices. For the LTX GGUF I mainly use the Vantage Q8 from https://huggingface.co/vantagewithai/LTX-2.3-GGUF/tree/main/dev.