<h3 id="training-wan2.2-lora-with-16gb-of-vram">★Training Wan2.2 Lora with 16GB of VRAM</h3>!!! No download required! This is a technical showcase! !!!→→→Original / Training dataset containing configurations provider→→→ <a target="_blank" rel="ugc" href="https://civarchive.com/models/1944129?modelVersionId=2200388">https://civarchive.com/models/1944129?modelVersionId=2200388</a><h3 id="differences-from-the-original:">●Differences from the original:</h3><ul><li>4070 Ti Super 16GB / Mem 64GB</li><li>This is my first training (Really the first run)</li><li>(Replace musubi_tuner_<a target="_blank" rel="ugc" href="http://gui.py">gui.py</a>)(I forgot to do this so i2v didn't work)</li><li>t2v (Unchecked I2V Training)</li><li>The model was also changed from i2v to t2v</li><li>Changed the path of other models to my environment</li><li>video_directory and cache_directory in dataset_bounce_test.toml</li></ul><ul><li>Install Triton and Sageattention2 (SDPA / I'm not sure about the effect, but I think Xformers would be good too.)</li></ul>pip install -U "triton-windows&lt;3.3"python -s -m pip install .\triton-3.2.0-cp312-cp312-win_amd64.whl→→→Guide is here→→→<a target="_blank" rel="ugc" href="https://civarchive.com/articles/12848">https://civarchive.com/articles/12848</a><ul><li>Blocks to Swap (As far as I know, this model has 40 blocks)</li></ul>35 (original is 10)<h3 id="vram-required-for-training">●VRAM required for training</h3>30 percent of the time, I was using 15GB, and the rest of the time it was around 10GB.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/62d7f1bf-9d39-45b5-a806-b9ce59a55bc9/width=525/62d7f1bf-9d39-45b5-a806-b9ce59a55bc9.jpeg" />As I write this, I set it to 40 to see if I could train with 12GB, but I got an error."AssertionError: Cannot swap more than 39 blocks. Requested 40 blocks to swap."<h3 id="memory-required-for-training">●Memory required for training</h3>Approximately 29GB of system memory was used.So I think 32GB should be just enough.<h3 id="time-required-for-training">●Time required for training</h3>The high was 2 hours and 19 minutes, and Low was 2 hours and 17 minutes.I trained for 20 epochs, but it took 5 epochs to see results.So if I do it right, I can train in just over an hour. (Wow, I can't believe that!)<h3 id="training-resolution-(dataset_bounce_test.toml)">●Training Resolution (dataset_bounce_test.toml)</h3>I think it will learn based on this specification rather than the resolution of the material.resolution = [256, 512]<s>I just realized that I don't know what "source_fps = 64.0" means.</s>I see, the source video was really 64FPS.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/eba3d0c9-e5c9-4d43-a805-76da581d1b3e/width=525/eba3d0c9-e5c9-4d43-a805-76da581d1b3e.jpeg" /><h3 id="training-screen">●Training Screen</h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/3c472999-2494-409f-b9a6-2c2f48da7c51/width=525/3c472999-2494-409f-b9a6-2c2f48da7c51.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/26ee2b5c-6e07-4365-bd52-91bb49ab5ffd/width=525/26ee2b5c-6e07-4365-bd52-91bb49ab5ffd.jpeg" /><h3 id="configuration">●Configuration</h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/d0c57e6a-44f1-4b89-8974-efa98c7d679f/width=525/d0c57e6a-44f1-4b89-8974-efa98c7d679f.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/bc722ccd-515f-4c3b-a5bd-4e75888dbc11/width=525/bc722ccd-515f-4c3b-a5bd-4e75888dbc11.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e5b4043e-88bf-43cf-8682-c114068eb4dc/width=525/e5b4043e-88bf-43cf-8682-c114068eb4dc.jpeg" /><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/be30198d-0209-4ce0-905d-6a908ff248bf/width=525/be30198d-0209-4ce0-905d-6a908ff248bf.jpeg" /><h3 id="software-and-version">●Software and version</h3>PS D:\sd\git\PGCRT\musubi-tuner_Wan2.2_GUI\venv\Scripts&gt; .\activate(venv) PS D:\sd\git\PGCRT\musubi-tuner_Wan2.2_GUI\venv\Scripts&gt; pip listPackage Version Editable project location----------------------- ------------ ---------------------------------------absl-py 2.3.1accelerate 1.6.0av 14.0.1bitsandbytes 0.45.4certifi 2025.8.3charset-normalizer 3.4.3colorama 0.4.6contourpy 1.3.2cycler 0.12.1diffusers 0.32.1easydict 1.13einops 0.7.0filelock 3.13.1fonttools 4.60.0fsspec 2024.6.1ftfy 6.3.1grpcio 1.75.0huggingface-hub 0.34.3idna 3.10importlib_metadata 8.7.0Jinja2 3.1.4kiwisolver 1.4.9Markdown 3.9MarkupSafe 2.1.5matplotlib 3.10.6mpmath 1.3.0musubi-tuner 0.1.0 D:\sd\git\PGCRT\musubi-tuner_Wan2.2_GUInetworkx 3.3numpy 2.1.2nvidia-ml-py 13.580.82opencv-python 4.10.0.84packaging 25.0pillow 11.0.0pip 22.2.1protobuf 6.32.1psutil 7.0.0pynvml 13.0.1pyparsing 3.2.4python-dateutil 2.9.0.post0PyYAML 6.0.2regex 2025.9.1requests 2.32.5safetensors 0.4.5sageattention 2.2.0sentencepiece 0.2.0setuptools 63.2.0six 1.17.0sympy 1.13.1tensorboard 2.20.0tensorboard-data-server 0.7.2tokenizers 0.21.4toml 0.10.2torch 2.6.0+cu124torchvision 0.21.0+cu124tqdm 4.67.1transformers 4.54.1triton-windows 3.2.0.post19typing_extensions 4.12.2urllib3 2.5.0voluptuous 0.15.2wcwidth 0.2.13Werkzeug 3.1.3zipp 3.23.0<h3 id="how-many-clips-and-how-long-do-i-need">●How many clips and how long do I need?</h3><a target="_blank" rel="ugc" href="https://civarchive.com/models/1454728/blowjobs-man-in-frame">https://civarchive.com/models/1454728/blowjobs-man-in-frame</a>20 clips of 3 seconds, 16fps, 512px512px<a target="_blank" rel="ugc" href="https://civarchive.com/models/1962482/wan22-t2v-14b-prone-bone-sex">https://civarchive.com/models/1962482/wan22-t2v-14b-prone-bone-sex</a>12 clips, ranging from 3 seconds to 6 seconds RTX4080.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1954733/asshole-wan-22-t2vi2v-14b">https://civarchive.com/models/1954733/asshole-wan-22-t2vi2v-14b</a>259 * 512x512 videos of 3 seconds a single 5090<a target="_blank" rel="ugc" href="https://civarchive.com/models/1953874/twerking">https://civarchive.com/models/1953874/twerking</a>20 videos - 3 seconds.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1953467/licking-breasts">https://civarchive.com/models/1953467/licking-breasts</a>23 videos - 512*512 - 3 seconds.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1953632/reverseanalcowgirlwan22t2v14b">https://civarchive.com/models/1953632/reverseanalcowgirlwan22t2v14b</a>7 clip used of about 3-4 seconds each RTX 3090 and 32 GB ram<a target="_blank" rel="ugc" href="https://civarchive.com/models/1343431/bouncing-boobs-wan-14b">https://civarchive.com/models/1343431/bouncing-boobs-wan-14b</a>7 videos with about 4 second 4090 it was done in about 35 min<a target="_blank" rel="ugc" href="https://civarchive.com/models/1934246/standing-sex">https://civarchive.com/models/1934246/standing-sex</a>19 clips of 5 seconds, 16fps.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1930903/blowbang">https://civarchive.com/models/1930903/blowbang</a>21 clips of 5 seconds, 16fps.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1930239/missionary-anal-trans">https://civarchive.com/models/1930239/missionary-anal-trans</a>20 clips of 5 seconds, 16fps.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1929589/kissing-tongue-action">https://civarchive.com/models/1929589/kissing-tongue-action</a>25 clips of 5 seconds, 16fps.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1927742/side-sexspooning-trans">https://civarchive.com/models/1927742/side-sexspooning-trans</a>29 clips of 3 seconds, 16fps.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1916746/walking">https://civarchive.com/models/1916746/walking</a>10 secs ++<a target="_blank" rel="ugc" href="https://civarchive.com/models/1927612/reverse-cowgirl-trans">https://civarchive.com/models/1927612/reverse-cowgirl-trans</a>41 clips of 3 seconds, 24fps (should've done a 16fps), maybe for next time.<a target="_blank" rel="ugc" href="https://civarchive.com/models/1894970/wan-22-reverse-suspended-congress-i2vt2v">https://civarchive.com/models/1894970/wan-22-reverse-suspended-congress-i2vt2v</a>23 video clips of 3 seconds each, mainly 3D animation<a target="_blank" rel="ugc" href="https://civarchive.com/models/1869475/wan-22-anime-cumshot-aesthetics-precision-load-i2v-beta-version">https://civarchive.com/models/1869475/wan-22-anime-cumshot-aesthetics-precision-load-i2v-beta-version</a>39 original animated clips Length: 3 seconds each FPS: 16 ↑Training Details↑<a target="_blank" rel="ugc" href="https://civarchive.com/models/1944129/slop-bounce-wan-22-i2v">https://civarchive.com/models/1944129/slop-bounce-wan-22-i2v</a>7 of my old AI generated videos on my 3090 ↑Training data includes↑<a target="_blank" rel="ugc" href="https://civarchive.com/models/1941041/facefuck-t2v-wan22-video-lora">https://civarchive.com/models/1941041/facefuck-t2v-wan22-video-lora</a>12 clips of 256 x 256 pixels, 3sec to 8sec<a target="_blank" rel="ugc" href="https://civarchive.com/models/1874811/ultimate-deepthroat-i2v-wan22-video-lora-k3nk">https://civarchive.com/models/1874811/ultimate-deepthroat-i2v-wan22-video-lora-k3nk</a>593 clips with various resolution, 1sec to 3sec<a target="_blank" rel="ugc" href="https://civarchive.com/models/1852647/m4crom4sti4-huge-natural-breasts-physics-wan22-video-lora-k3nk">https://civarchive.com/models/1852647/m4crom4sti4-huge-natural-breasts-physics-wan22-video-lora-k3nk</a>211 res on a dataset of 889 clips 1sec to 3sec (several source videos at 16fps)<a target="_blank" rel="ugc" href="https://civarchive.com/models/1858645/facial-cumshot-wan-22-video-lora">https://civarchive.com/models/1858645/facial-cumshot-wan-22-video-lora</a>307 clips with 211 resolution, 1sec to 3sec<a target="_blank" rel="ugc" href="https://civarchive.com/models/1845306/sideview-deepthroat-wan22-video-lora">https://civarchive.com/models/1845306/sideview-deepthroat-wan22-video-lora</a>700 clips with various resolution, 1sec to 3sec<a target="_blank" rel="ugc" href="https://civarchive.com/models/1969272/4n4l-pl4y-i2v-anal-didlo-lora-wan22">https://civarchive.com/models/1969272/4n4l-pl4y-i2v-anal-didlo-lora-wan22</a>98 clips with 256x256 resolution, 1sec to 3sec<a target="_blank" rel="ugc" href="https://civarchive.com/models/1960102/self-nipple-sucking-lora-i2v-wan22-k3nk">https://civarchive.com/models/1960102/self-nipple-sucking-lora-i2v-wan22-k3nk</a>319 clips of 5 different scenes with various resolutions, 1sec to 5sec 211 res<a target="_blank" rel="ugc" href="https://civarchive.com/models/1954774/cunnilingus-pussy-licking-lora-i2v-wan22-k3nk">https://civarchive.com/models/1954774/cunnilingus-pussy-licking-lora-i2v-wan22-k3nk</a>121 clips with various resolution, 1sec to 5sec 256 res<h3 id="create-captions-(txt)-from-video-file-(analyze-one-frame-as-an-image)">●Create captions (txt) from video file (Analyze one frame as an image)</h3>I don't think this is the smartest way to do things, but I hope it helps someone.<s>You can load this image by downloading it and dropping it onto the Comfyui screen.</s>Oops, my workflow disappeared after uploading.The caption text will be saved with the same filename as the video file.<h3><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/dba14cf1-20ff-4144-885b-f52168493f5c/width=525/dba14cf1-20ff-4144-885b-f52168493f5c.jpeg" /></h3><h3 id="i-tried-the-next-training">●I tried the next Training</h3>32 videos, 20epocs, resolution = [512, 288], blockswap=36However, for some reason, 649 video caches were created, resulting in a final total of 12,340 steps... (39h54m / 1lora)<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e11fd126-3992-42bc-bdb5-9e2ac173d5d6/width=525/e11fd126-3992-42bc-bdb5-9e2ac173d5d6.jpeg" />

Wan2.2 16GB Training

bounce_test_HighNoise-000015a.safetensors

15epoch-High

5epoch-Low

5epoch-High

10epoch-Low

10epoch-High

15epoch-Low

20epoch-Low

20epoch-High

Files with SHA256: d61577eaa26704235cc3f803aa4837711986e2094ee96832a672c5636b6492ff