author: NoArtifact
baseModel: Hunyuan Video
hashes:
AutoV1: 9F93C0A4
AutoV2: C9E3BC8517
AutoV3: D94F4A3C9E19
BLAKE3: C484F3DC0B44B856C38FFB767AEEB285AEFBCB5CF19D18133D890D5598A45861
CRC32: CF30F7EE
SHA256: C9E3BC8517140F5D92439249B4FFC9476E20833B939B36F8CD9D1CE6D28881A9
metadata:
format: SafeTensor
modelPage: https://civitai.com/models/1186768?modelVersionId=1391442
preview:
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/1f8b2d17-7175-4668-8540-e23e7e85e2a6/width=450/56754539.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/127edb86-7cd4-4578-8535-76aa5e953a2f/width=450/56741006.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d5dc68a4-9242-4cf5-9641-af8c14df11e0/width=450/56741023.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/3326d2ab-36f8-42df-b3e3-c1720b275678/width=450/56741029.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d6045664-68c0-43bf-8f77-fc6affe620b5/width=450/56741059.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/cb83cbc8-3254-4d5d-a4fc-df04ea8f273e/width=450/56745203.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2abedce1-70c5-45b1-bb0d-3126d4d69e96/width=450/56748142.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c688fe8f-bce9-47b4-8516-8ad42394c299/width=450/56748388.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/0117a57c-d4b5-4980-87bd-4bc8a4aa00ff/width=450/56750525.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/727d99f3-a559-4076-9b76-a694e6190f4e/width=450/56755942.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8116bc87-1a3e-4995-a097-52d97078e63b/width=450/56757529.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d098a2fe-1f69-440f-b091-82e97504f7a4/width=450/56758928.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c9018d59-2f48-4464-ba71-673e9dbc1fa8/width=450/56762417.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/60383917-c330-4f9f-9e68-4c819c20a38a/width=450/56762455.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8328c7dc-af94-44f6-b0df-d1414b87b30b/width=450/56833087.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/a411e244-b112-4eac-a3f2-c61b519e898c/width=450/56842995.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/11724752-d134-4a1a-837a-dacea7010a42/width=450/57899572.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/5740c603-395e-45fd-984d-26875c80c4ba/width=450/57903293.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/8f49b8ca-f88c-42e8-8248-23b965a85f06/width=450/57907120.jpeg
- >-
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/410b263c-b09e-4dfa-9d68-555329b686a6/width=450/57923584.jpeg
website: Civitai
Trigger Words
8itchWalk4, 8itchM0unt, BAKw4lk, BAKM0unt
About this version
Epoch 20, Too many clip trained (about 70 clips, compared to less than 25 for V1.1), it lost consistency, can be somewhat managed through stronger prompting of certain aspects (for example if you want back view, make sure you clearly depict something that is veiwed from the back, don't just put back view it may not be enough and it will hallucinate front view into the back view... for example boobies in the back !
If I ever make another version I'll probably reduce clips to around 30 MAX, between 20 and 30
Move_Enhancer
V3 INFO below (I recommend reading the usage tips and extra tips) V3.0b is epoch 25 I just posted it, basically the same as V3.0 maybe a bit better, maybe not, it's just out there for you to test and experiment, feedback appreciated
V3/V3.0b (for V3.0b just read the about, it's basically the same with 7 more hours of training)
After more testing, I consider V3 is the best version, BUT, V1.1 still have some qualities, I wont say it's a clear upgrade if you just plug and play replace V1.1 with V3 in a test for example, I would say it's an upgrade if you understand that you need to make stronger emphasis on some details, like which way the person is facing, which way the person is going etc... Because V3 being trained on more than the double number of clips than V1.1 including much more back view (V1.1 had none) V3 can make confusion between front and back since it knows both, so be more precise in your prompt and I believe in this case V3 is an upgrade.
V3 Pros :
- Better knowledge of back view than any previous of version
- Better knowledge of Mounting (mostly Horse riding) animation than previous version
- Should have also better knowledge of front view, movement in general
- Should have better 'jiggling/bouncing/swaying/wobbling' understanding than previous version
- Have an extra vague understanding of water physics/interaction (water falling, like a bucked of water poured on someone for example, (don't expect miracles though)
V3 Cons :
Tend to hallucinate more than any previous version (but can be somewhat controlled by making stronger prompting to guide it), this is probably due to having being trained with much more varied clips, and more clips overall, period, than previous versions, it's the downside of this.
Have a bit of confusion between forward and backward motion sometimes (this is part of hallucinating but on the animatin side of things) again, you should be able to somewhat control it by making stronger emphasis on what you want in the prompt.
Running animations seems more hit and miss than V1.1 probably because I had more walking clips in the training this time (I also had running ones, but the proportion wasn't the same) in general as I already said, v3.0, while capable, is a touch more hit and miss overall anyway.
No men trained, (this is the case for every version, but since this one have the most clips it probably mean the model favor women the most, so, making male version of animation may generate very feminine walking/running men at best, less versatile on this aspect)
Like previous versions, this model have strong tendency to produce big to massive breasts, so if it's what you wanted fine, if you wanted more natural smaller breasts, you may want to reduce the weight of this lora and/or combining with other loras that produce smaller breasts. (I may include more veriety of breasts size in next version)
Usage TIPS :
Use BAKM0unt trigger, for back view mounting/riding video, this one is a bonus, and should be a bit hit and miss since there is only ONE video in the whole dataset that is like that (it's a test just in case it urn okay)
Use 8itchw4lk trigger at the beginning of your prompt for front view, walk/run and some 'iddle' animation, such as simple dance moves or jumping or things like that
Use 8itchM0unt trigger for any front facing Mouting/riding style animation (trained mostly on riding horses on a saddle also one Giant tiger, and of course try your luck at something more creative and see how it work)
Use BAKW4lk trigger for any back view style walking/running animation, and also some iddle dance/jump kind of animation
Notice, these trigger words are just basic guide, the real deal is your actual prompt that follow, these are just to give an extra hint to the model, be descriptive.
Extra tips (thanks to other users findings) for making those 'fun bags' moving and jiggle with more flair, combine one or many of these other loras at various weight (test by yourself) lowering or increasing the Move Enhancer one to achieve a good balance
- https://civitai.com/models/1052680?modelVersionId=1181194 Dancing with breasts bouncing
- https://civitai.com/models/1230603?modelVersionId=1393761 SingularUnity Beautiful Blondes
- https://civitai.com/models/1096716?modelVersionId=1231902 HunyuanVideo Reverse Cowgirl (this one might seems more surprinsing at first, but seems quite effective at making it bounce)
and most likely other loras achieve similar quality of 'breasts physics'
V2 INFO Below
V2 (Fail Edition actually) was a big failure, my training settings weren't good, I decided finally to upload it still, for testing purposes, mix it up at low weight (0.30 - 0.6 max) with v1.1 or other loras, and see if you may have some decent result, This is unfortunately NOT AN UPGRADE over 1.1, I will revert to previous settings and keep clips in slow motions for the next version (IF I EVER MAKE ONE that is!)
V2 (Fail_Edition) pros :
- Footage aren't automatically in slow motion anymore
- Some knowledge of the back view (make that behind movin, very hit and miss though)
V2 cons :
- Choppy output, unsteady output like camera have no stabilizer
- Less consistent
- trained on 45 frames clips, meaning animation tend to 'reset' more often, creating less harmonious walking/running animation, training time was significantly reduced though...
- It's simply a failure, and you are warned, don't need to tell me it's bad, just TEST or don't test, and see what you may get.
- v1.1 is clearly better (and that's the bottom line)
V1.1 and V1 INFO
V1.1 (decided to not call it V2) One step forward, two steps back ? Too early to say, I fixed the blocky, chunky overly low res output as I expected, problem is the motion seems now to be a bit less natural, a touch more 'stiff', and maybe also the final output looks a bit too much CGI ? Hard to say for certain now, but I'm not super happy with the result, sitting on the fence until I test more out of it, only good point is a much less blocky pixelated low res output which was the main problem of 1.0. If you guys having test V1 can test 1.1 and tell me your impression (there is only so much I can test by myself, many parameters can make output vary a lot) appreciated.
TIPS for V1.1 : I added two clip of woman riding a mount (a white horse, and a big tiger, so for consistency if you want a sure shot, go for white horse or tiger as her mount) add 8itchM0unt at the start of your prompt (or just before saying something like : She is riding on top of a [insert your mount of choice] if you want to go for something else than white horse or tiger, increase the strenght of your prompt for example (she is riding a giant Eagle mount:1.4) but it will certainly be hit and miss since only two clips on the entire training data are consisting of mount, best of luck.
Woman moving toward the viewer, it significantly enhance the realism and allure of walking/running/jumping woman.
Tested (with and without Lora) ended in a quick 6-0 knockout in favor of the lora. This was a quick test lora (that still took my overheating RTX3090 8hours to complete...) It has some flaws, but in term of enhancing movements it's quite good, surprisingly good in my few tests.
Strenght : Movement, overal, try it, you'll love it, possiblities are quite big, I even managed a somewhat believable skateboarding movement by lowering the lora weight, but it's was not amazing, just good enough considering no skateboarding at all in the training...
Weakness : Due to no images being trained, just low res video clips, the model tend to create a bit of a blurry output (try same prompt with and without the lora at equal resolution you will see quickly) but I feel the bad is vastly less crucial than the good which is movement
Second main weakness, ONLY Front view toward the camera movement trained (I had plan to make back view but wanted to be sure it went somewhere first with less training data) so you can still try to make it from rear, but don't expect as good of a result then
Lesser weakness : Almost ALL training material had MASSIVE BOOBIES ! I mean massive, so it tend to create slightly bigger than usual boobs even if not asked for... not the biggest of deal but still.
Future plan: I think you should all try (those with capacity to train, if possible with TWO PC so testing while the model is training, unlike me who need to wait before testing...) to train those kind of loras, but if for some strange reasons, someone don't come with an even better one soon, I have already ideas to improve it, hopefully vastly. Main issue would be to reduce the low res aspect, maybe by adding high res image in the mix, and/or increasing the resolution (I used 320 resolution, which is not huge, but I was already at 22.3 G out of 24 V ram used during training, so not so much headroom left
Usage tip : Workflow and prompt (some) in the still images, Got my best result while putting weight of 0.68 to 0.85 (I suspect it is a bit overtrained, seeing how it picked up the blurryness of the low res videos, maybe I should test epoch_30 too).
Fun fact, NOT A SINGLE training video included fully naked women, so don't be hesitant to prompt clothing as well, even if it mostly consisted of bikinis and very provocative stuff, it also kinda understand when it's not fully naked, that's why I mostly tested for naked I wanted to see if residual clothing was appaearing when trying naked output, overall it's okay
For those upper body movements, adding boucing, jiggling breasts, as well as realistic physics, tend to help
I just added the training data prompt.txt I used, and the toml files in case it can help, or if anyone have advice on them for those with ideas.