i2v GGUF works me incerrectly
Hello,
I have problem with i2v GGUF model. It behaves like IP adapter not i2v. This means that produced video is only "some inspiration" of input image and not starting with input image as first frame like in case of original BF16 model. I tested it with workflow from https://comfyui-wiki.com/en/tutorial/advanced/hunyuan-image-to-video-workflow-guide-and-example
GGUF model interest me because this can be used with UNET DisTorch loader that dramatically reduces VRAM requirement. For example in case of HD and 49 frames I'm able to render only with slightly less then 15 GB VRAM requirement.
Any idea how to fix it?
Thanks
Make sure you're using the correct workflow. The models in this repo are the v2 "replace" ones, I assume the workflow you're using is for the v1 "original" model (also available in this repo, as mentioned in the readme).
You can try the workflow from the bottom of the official example page here, I believe this one should work: https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/#v2-replace
Thank you. I downloaded v1 (original) model Q4_0 for testing purpose and it's really look that the workflow that I'm using is for this old v1 version. The eesulting video is very close to the input image. Difference betweet the result and the input is propably caused that Q4_0 is less precise than BF16 model.
Unfortunatelly I don't know how to fix the workflou to be usable for v2 models. The link that you provided is only for safetenssors models, not GGUF.
In theory, it should be fairly simple to adapt any workflow to the GGUF nodes. All you have to do is replace the model loader node (Load Diffusion Model) with the GGUF equivalent (Unet Loader (GGUF)) and then connect it to the same things it was connected to before by dragging the model output of the new node to the target node(s).
You can bring up the node by double-clicking anywhere and typing "gguf", or by taking your existing workflow and selecting the node, then just copying it with ctrl+c and pasting it in the other workflow lol.