Introduction:
First of all big thanks to @portraitman @dogarrowtype and other admins of furry diffusion channel for their encouragement and support in helping animation creations, other good fellows here have also been very helpful and creative in animation gens.
It’s been a while since the release of animatediff, but creations so far (both swf/nsfw) are mostly scene change, dancing, character transformations which are generated through vid2vid or openpose controlnet without exploiting the full potential of animatediff. After testing with different combination of base models, motion models and prompts, there are indeed decent animations which can be also generated with the simplest setup, providing pleasant fapping experience.
Installation and preparations:
These two guides cover almost all the details, for both webui and comfyui users.
https://civitai.com/articles/2877/how-i-make-animation-via-animatediff-tutorial
https://education.civitai.com/beginners-guide-to-animatediff
Workflow summary (myself):
Choose the suitable base models
Choose the suitable motion models
Use the right prompts
quick preview (Use LCM) to the above steps combinations
Set up parameters (and latent upscale), generate and edit
1. Choose the suitable base models.
The compatible base models discovered so far (will keep updating):
Newdawn7.4
can create realistic animations
Indigofurrymix 7.5,8.0
can create with various actions
Yiffymix 3.4
can create various actions and facial expression
Purrfectmerge
can create various actions
Bb95furrymix 12.0
can create various actions, needs more testing
Examples 2 thanks to @aurion
Fluffyrock
can create animations, needs more testing
examples 1 thanks to @eevee_k
2. Choose the suitable motion model
The motion model is CRUCIAL in creating decent animations, here’s the list of motion models discovered so far (will keep updating):
Default motion models
the basic motion model, can produce simple body actions like walking, waving hands.
Examples 2 thanks to @aurion
Motion module experiments
The most suitable motion model so far for nsfw actions like vaginal sex, anal sex, cumshot, etc. Can switch between v0.1 and v0.3 for better results (v0.3 makes the output more anime style).
Examples 1 (keywords: anal sex, cowgirl position, anal penetration)
Examples 2 (keywords: penetration, penis thrusting very hard, vagina moving up and down, cum, creampie, massive cum)
Examples 3 (keywords: autofellatio, cum, messy)
Examples 4 thanks to @dogarrowtype (keywords: masturbation, cum, messy)
Improved human motions
needs more testing
3. Use the right prompts
The animatediff prompts are no big difference from pic generation prompts. The main goal is to explain the animation as short as possible, longer prompts would lead to lower quality gens.
4. Quick Preview
Images size of 512X512 with 16 frames are recommended for this step.
With the latest LCM Lora online, sample steps and generation time can be dramatically reduced.
https://github.com/0xbitches/sd-webui-lcm
https://mybyways.com/blog/generating-animations-videos-with-sdxl-lcm
But animation output with LCM enabled has color loss issue, so it can be suitable for previews.
If the preview is messy and distorted, it’s very likely the action prompts isn’t compatible with current base model or motion model, rewind to step1~3 and change the combinations. If the output is mostly correct, proceed to next step.
5. Set up parameters, generate and edit.
Setting up proper parameters for gens after preview confirmation is also important, the size of images better not exceed 768X768 depending on creators’ gpu capabilities. Longer frame number like 32 to 48 is recommend since 16 is too short for big movements.
thanks to@eevee_k some samplers should be avoided.
The upscale can be achieved by fixing the seed thanks to @eldragon0.
Since closed loop option doesn’t affect longer animations output, frame edit programs like Photoshop is recommended to rearrange the frame.