Hi, it should be compatible with the GGUF version of the model, but 8Gb of vram seems very low, I would say that 16gb is the minimum, even 24 is low for Hunyuan unfortunately, But I recommend you to try LTXv, it's fast and not so demanding on vram

Hi, no, I didn’t train anything sexual. It’s essentially a LoRA focused on refining the "look" style, face, and upper body. You can check the Flux version images for a better understanding of what it does. The dataset used for the Hunyuan LoRA is roughly the same but much smaller approximately 1,500 images compared to around 5,600 for the Flux version. I’m currently testing the addition of videos and motions, but that will be a separate model.

The goal is to showcase the overall look, so I’m fine with 2–3-second videos as long as they have decent quality and motion. I have a 4090, and generating 249 frames at this resolution is not feasible, at least with the FP8 version of the model. The official FP8-scaled version doesn’t support LoRA yet.* Not using FastHunyan here*

Also, 20 seconds for 249 frames equals 12 FPS, which is kinda slow for the motion. I prefer short videos that look real-time rather than long, slow ones.

I use Enhance-A-Video and Sage Attention, yes. But since the model isn’t restricted by size or length, there are no "best settings" to recommend, sky "vram" is the limit.

As for long video generation, I’m not interested in that at the moment. The current methods aren’t very good. Using the last frames of the first video to generate the second and so on, creates a kind of self-convolution, and the results degrade progressively over time.

1/5 is AI images generated with the Flux version using prompts of the original dataset,

4/5 is the original dataset (but smaller), only composed of real photoshoot with DSLR on 70-200mm lenses

Hi, I share all of my workflows, you can find them on my profile page, but you can have the same results with a default Hunyuan workflow. The videos I posted are upscaled and interpolated to 60 fps with topaz video 3.3.3

The point of posting all that? I think you don't see the amount of training, and rendering time to create these videos, I made about 200 at high resolution each of them renders in about 320seconds. More than 15 hours of extra computing just to give you guys some examples...

I posted 10000 of these prompts two monh ago on the first iteration of the model, flux 1.0. Is it enough or you want more?

@StanleyPain in the download as training dataset there is a .txt file that contains 10k prompts, this is all the dataset captions merged

Hi, I rendered most of my last posts at 800x1184 without teacache, no additional post processing, then upscaled using topaz video 3.3.3, x2 with frame gen to 60fps

I misclicked and deleted the whole page instead of an updated unpublished version, I asked the moderation to revive it, fortunately it was still possible

@makiaevelio543 This is correct, the thing is, when you train a vast amount of images with various characters/poses/environments/night or day/framing etc. with long LLaVA or Florence2 captions, it's hard to clearly describe what it does and what the model retain the most from training, this is why I called it secret sauce it's not a gatekeeper thing

Almost every face was different so I don't know what could be the cause for faces that defaults to the same woman

I resumed the training from 15K steps (B3) to 30K steps (C2), but I didn't have the time to test it and do comparisons yet

I removed the v1 few days ago, this was not the full training "images" data, it was the captions.

This is the combined captions for the B3 version

https://drive.google.com/file/d/1cb-1H_kWnjFtSMcVPvTAPw7PrZUVN7oe/view?usp=sharing

For the most part yes, but there is a small part of flux synthethic data augmentation

Hi, it just looks sharper with its almost imperceptible, i'm not stopping you to use your own upscale solution, I do things for me first then I share if it can be useful for others, you hate it or not i don't give a single F to be honnest

@OFIA Hi, I didn't use anything else than topaz and comfyUI, you can use the upscale node from comfyUI with the upscale model of your choice, you can also use frame interpolation node to have a more fluid video, from 24 to 48 fps. On topaz the frame interpolation is not limited by a multiplier value, you can set the desired output fps instead, I use 60fps.

I ran a batch to generate all these videos over the night, I prefered using an external upscaler to only upscale the videos that I wanted to keep, this way I didn't spent extra power on videos that was not good enough.

It was trained with diffusion-pipe, I have a 4090 but I didn't remember the peak VRAM usage during the training, I think 16go is enough but i'm not sure. To train videos, 512x512px 33frames rank 32 barely fits on a 24go GPU, so it's not possible if you don't have a 3090/4090 or less (to train videos)

It was trained with diffusion-pipe yes, I have a 4090 but I didn't remember the peak VRAM usage during the training, I think 16go is enough but i'm not sure. To train videos, 512x512px 33frames rank 32 barely fits on a 24go GPU, so it's not possible if you don't have a 3090/4090 or less

If you say so