I just finished training a new Hunyuan LoRA for Optimus, and the results are incredible. The quality and prompt adherence are unlike anything I’ve worked with before. This model is perfect for creating cinematic, high-detail scenes with ease, trained on flux outputs.
If you're curious about my AI work or want to see examples of what I do, check out my YouTube channel. I'd love to hear what you think or see how you use it.
Follow me on X for more details about my workflows and content.
example prompt: A side-view shot focuses only on Optimusk’s sleek, black-and-white mechanical legs as he walks along a peaceful, sunlit path. His precise, rhythmic steps contrast with the soft, natural world around him. The ground beneath him is a blend of smooth stone and scattered leaves, dappled with golden sunlight filtering through the trees. A gentle breeze stirs the grass at the edges of the frame, and faint ripples shimmer in a nearby stream. The camera lingers on the motion—subtle, deliberate, almost poetic—capturing the contrast between machine and nature. The atmosphere is quiet, reflective, and deeply cinematic.
note: if you get optimus prime you may not be loading the lora properly.
Description
FAQ
Comments (23)
It is impressive to see what Hunyuan can do with loras. atm we don't have a lot of cool creative ones and i personally can't really dedicate a machine to running them locally, not when i am using my PC all day and so far it seems creating lora's online is too costly. but i hope this becomes easier. I seen a video about an experimental lora trainer that is built into comfyui module that could potentially make a lora in a couple hours. Though no idea how good it is or anything yet, i am giving more time for the tools to mature before i try. Not a fan of using linux or having to do some virtual machine or any of that nonsense. But i look forward to the future.
With image2video coming soon you won't have to worry about all that. That's hunyuans biggest weakness rn so loras are like a band aid fix atm. It's in their open-source plan confirmed to release soon. I have this running on windows 10 locally w/o any optimizations and it's not too shabby. However I do agree they should optimize it more for windows bc nobody wants configure all that themselves. It is definitely in it's early stages for sure, the future is bright though : )
@bhopping I personally don't think img to video will solve that. I think you'll still need lora's. I've already seen what img2video looks like a little bit with a couple of the models out there wit the free trial. It works for generic things but soon as you ask it to do something it's not trained on it falls apart. Being able to start with an initial image will be amazing but i think it'll still need a lora to guide it. At least if your doing anything base model isn't good at. If your just doing basic things sure. I don't know, i'll hold full judgement for when img2video comes out and see. I am sure i'll use it, i just don't we can give it any image and action we want and it'll do it without training.
@zengrath Yeah I agree there really hasn't been really any control on any video generations in general on the model themselves anywhere. When I initially used hunyuan raw when it came out, it was very annoying to control. It was just randomness. The loras will still be needed for sure for motion/style. I just dislike having to use t2v bc it leads to a lot of inconsistencies, or having to download/create character loras. It would be crazy one day to not have to use any loras at all
It will become esier very soon I think. As for now tou need to be proficient on phyton
@bhopping image to video will be great but loras are way better for characters if you really looking for continuity to make something with a story.
by the way image to video is supposed to come out on the Chinese new year. in 4 days.
@mckenna wow that's very soon. thanks for letting me know, I can't wait to test it out. I can already think of a workflow for it. I'm hoping having a character lora for both img and video will synergize with i2v
@mckenna Great. I assume it won't be immediately available on Comfyui until someone ports it over. but look forward to playing with it, thanks for the info.
@zengrath it usually its up and running in about a day or so :)they may not release on that day to this is based on something they posted on x but it may change
Curious, did you train this on Kling/Runway generations?
no. it was trained on stills. the pictures where created with flux with my custom Optimus lora.
Well then, I wrote Optimus Optimusk and I guess it understood Optimus prime transformer instead lol
Can I see how you've formulated your prompts for 1 of your video? Looks really good!
make sure lora is working well for you, that happen to me when I had the wrong config and lora was not loading. but yes use Optimusk... example prompt just in case: A side-view shot focuses only on Optimusk’s sleek, black-and-white mechanical legs as he walks along a peaceful, sunlit path. His precise, rhythmic steps contrast with the soft, natural world around him. The ground beneath him is a blend of smooth stone and scattered leaves, dappled with golden sunlight filtering through the trees. A gentle breeze stirs the grass at the edges of the frame, and faint ripples shimmer in a nearby stream. The camera lingers on the motion—subtle, deliberate, almost poetic—capturing the contrast between machine and nature. The atmosphere is quiet, reflective, and deeply cinematic.
@mckenna You made me realise that I was using the wrong Hunyuan checkpoint model from not getting anything good from your prompt. I was using the BF16 version instead of FP8. From what I understand there are versions specifically for Kijai nodes. Getting way better results now!
@Catz glad to see bro I will share the workflow I am using when I have the time but I recommend quantized version of fast hunyuan. I have it running at very high quality.
@Catz also thanks for sharing your outputs. for better results avoid describing the motion (not good at that yet) think of it like an animated still image.
@mckenna I initially started using this workflow and it was very well optimized. Rendering a 208x368 video then vid2vid and then upscale. https://civitai.com/models/1092466/hunyuan-2step-t2v-and-upscale
It worked well for generic txt2vid results, but had issues with some loras when it was at the vid2vid part.
Currently using this as it's 512x768 and give much greater results, but I get out of vram (24GB) if I try to last longer than 3sec / 73frames
https://civitai.com/models/1007385/hunyuan-allinone-fast-tips
I'm curious what you're using. TeaCache and Wavespeed nodes seems to be best of what I know.
@Catz hey bro my workflow is out come and try it https://civitai.com/models/1219744/hunyuan-triple-lora-fast-high-definition-optimized-for-3090-orby-bizarro?modelVersionId=1374109 if you have a slow card you may need to lower resolution use a different gguf.
@mckenna Ill give it a shot tonight! Btw, used your optimusk on musk lol https://civitai.com/images/56114049
There's an issue with linking resource on that model, but will do when its fixed
@mckenna Hey man, I've tried your workflow and it's extremely fast! I've just been obsessed with this other workflow these past days and give so much more quality.
https://civitai.com/models/1007385/hunyuan-allinone-fast-tips
Yours with the fast gguf does make good references for quick ideas.
bro, what is the optimum Lora Strength to use?
0.9 is good but depends what you are trying to do
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.