comfyUI is standard just use a wan2.1 workflow, you need alot of system recourses to run the 14B though at least 12gb VRAM and you might be able to get away with a minimum of 32gb of ram but most likely you need at least 64gb of ram. Wan 14b if you want it to make good generations you need to use the fp16 versions and that will use over 64gb of ram for inference. I tried many models on the standard fp8 people generally have by default and 1 in 10 generations would be somewhat ok, swap to fp16 almost all are good if your prompting correctly.

@basedbase @mylo1337  Thank you both for the replies. I'm fairly sure i have the system resources but i don't use any of the things listed by Mylo. sigh And I don't have time to re-learn entire new systems like comfy.

Is there anything out there that allows you to make animations, atleast short loops, that doesnt want to HARVEST your data, or require you to NOT make NSFW like that one who's name i forget, and preferably run it locally, that can be used on Forge?

I only ask because i HAVE looked but i find nothing other than the ones that do the things mentioned above.

@mylo1337 Thanks for the info, and sorry for the late replies. Had DND all weekend.

@mylo1337 really 15 seconds per step? Im surprised, im training the 1.3b with a batch size of 2 and gradient accumulation steps of 2, 480x480 res and up to 91 frames my time per step is 40 seconds on a rtx 3090 makes me wonder if im bottlenecked somewhere but I don't see how I would be running a 7800x3d and 128gb of DDR5 even have 3.5GB of VRAM to spare so its not like im at 99% VRAM usage.

@mylo1337 Also what was gradient accumulation steps? Trying to get the settings down for furry content since wan its difficult to train it correctly into wan.

@mylo1337 How many training videos were in your dataset for version 1 and 1.1? Did you change video frame rate to be the same across all videos?

@mylo1337 Thank you so much for the training info you provided, and, ofc, the lora itself. It is fantastic! I've noticed a lot of other lora's I've tried introduce various problems such as uncontrollable slow motion or fast forward (due to lack of normalization of frame rate during training), degradation of animation quality into shitty low poly 3d look (reason unsure), melting faces (and other parts), desaturation and change of style into gritty realistic (reason unknown) - none of which are present in your lora. Not only does it not introduce problems, but it improves the output quality of other loras when included alongside them. I think there is much learning potential here in your training method because obviously you are doing it right. Did you include various drawing style in the training data (like realistic, anime, 3d, etc.), is that why it works so well and doesn't change the style of the given image into something else like other loras?

@mylo1337 How did you caption the dataset? Besides the obvious motion info, did you caption subject appearance, background, facial expressions, drawing style (e.g. 3d, anime, realistic), quality? (Asking because you are obviously doing something right so I wish to learn from your success.)

@mylo1337 Was v 1.1 a checkpoint you resumed from with new training data/lr

@mylo1337 Thank you so much for sharing. 😘 I'd love to see your captioning tool! Currently, there are very few options available for video captioning and doing it by hand by creating txt files and then opening them in windows explorer is... uh... tedious to put it mildly. Even video cropping, which should be a basic function for a video editor, is either not possible or very inconvenient in most video editors. Shocking, isn't it? I only found out when I actually wanted to crop videos for training purposes. Most video editors require you to set a resolution of the project (which you can't do because you don't know the target resolution beforehand) and you can't crop it by hand like you would easily crop an image in Photoshop. I've yet to find a simple Photoshop-like video cropping app. Some video editors don't let you change framerate at will but only offer some more commonly used fps which doesn't include wan's 16. It's a giant mess making it very difficult to prepare a proper dataset.

@mylo1337 I also just did the same with my new single NSFW pose lora recaptioned 5 videos each extended to 5 minutes each, used dense captions I had qwen3 30b a3b rewrite and trained 8 epochs for a much better end result. Used a very low LR for tuning so it would not suffer from catastrophic forgetting LR was 4e-6 for the tune and I used LR of 5e-5 for the base lora. Each used a batch size of 4.

@Remiel Try captioning and editing a 170 video dataset it takes a very long time lol, going forward i'll be having a LLM rewrite all my captions to be denser and more descriptive, in my testing that helps alot with the quality of the outputs the lora makes

@Remiel Just recaptioned and retrained a dataset to use much more descriptive text files and specify if the video is 3D, 2.5D, or 2D and increased batch size from 4 to 12 and used square root LR scaling and got the most consistent lora I've trained to date. Proper dataset captioning is paramount even my tensorboard loss chart shows a consistent drop in loss for train/loss vs all previous runs where it just bounced all over the place.

@Remiel I trained the 1.3b though so if you scale batch size that much on wan 14b you will need at least 3 a6000 48GB GPU's, though you likely don't need to scale that much on the 14B since it learns much better per optimiser step than the 1.3b since you know, 1.3b parameters vs 14b parameters.

Civitai only recently started to support slightly more than extremely basic and limited ComfyUI metadata, and apparently the on-site gen uses ComfyUI for the backend, so you can likely guess how well the site will handle anything less common.

As long as it's the 14b model. You can pick any quantization, or scaled version.

Just so you know, Q8 is just as good as FP16. I've use them both and the difference is almost non existent.

@mylo1337 How much time would you say it takes per video and length? My 2080 with the GGUF at length 33 steps 20 took 18 minutes based on one time test. It uses about 17.5 GB of RAM (27GB of RAM all together). Using non GGUF takes 40GB of RAM, including OS and other stuff. I'm also curious about the wattage on the 4090 and 3060 when using AI. I would estimate that the 4090 total system power would be 500 watts from the wall. I get around 260 watts on average. I was looking forward towards the 5090, but it just sucks so much power and has issues, let alone the price. My room already blows the breaker when my AC and computer are on at the same time (Happens rarely).

Thats Wan 14B in general, you can use block_swap but I don't have issues running at full fp16 precision on a single 3090 its still slow at least 10 minutes for a higher res generation with any respectable frame length. Try using dpm_2 for the sampler and ddim_uniform for the scheduler that produces the most consistent and best motion generations from all the sampler combos ive tried. Use around 20 steps for the best quality but 6 steps is good for a quicker generation. Samplers and schedulers matter alot when aiming for quality UniPc is honestly not that great.