Holli Would
Hey all! This is a character I've thought of doing for a while and I figured I'd give her a shot myself. I actually watched the movie and went frame by frame in parts to grab as much of a dataset as possible and ended up with over 200 images! She's very diverse with facial expressions and poses that I can see with testing. Let me know what you all think! She was fun!
Trained with NoobAI-vPred but works perfectly fine with illustrious checkpoints
Grabbed some dancing and I Found it easier to pull those up by prompting (Dancing:1.2) if you're interested
Some nice facial expressions came up so she's pretty expressive with: Angry, annoyed, and naughty face
Let me know what you all think and post images! NSFW welcome :D
Description
FAQ
Comments (8)
Thanks!
Yo man, your Holli stuff is unreal 🔥 I've been recreating some of it to learn but my stuff looks like dog compared to what you've put togehter.
I think I got the Anima checkpoint + your Holli LoRA working and animating the stills with Wan i2v. Stills come out great, but my videos are coming out a little jumpy/stuttery and I'm trying to get that smooth look yours have.
Civitai didn't save the gen data on your videos (external generator), so I'm flying a bit blind on the Wan side 😅 If you're down to share, would massively appreciate:
Which Wan are you on — 2.1 or 2.2, and the 5B or 14B model?
Steps / CFG / sampler+scheduler for the video pass?
How many frames + what fps, and do you interpolate (RIFE/FILM) to smooth it out?
Do you upscale the video after, and with what?
Thanks fam 🙏
@bacuss Yo, glad you're digging it 🙏 Here's my exact Wan setup.
For the model I'm on Wan 2.2, the 14B one. Specifically it's the SmoothMix Wan 2.2 I2V gguf merge (Q6_K), and it uses the dual high noise + low noise models. Text encoder is umt5-xxl Q5_K_M gguf, the VAE is the wan 2.1 vae since 2.2 still reuses it, and clip vision is clip_vision_h.
For the sampler pass I run 6 steps total at CFG 1, euler_ancestral with the simple scheduler. It's actually split across two KSamplerAdvanced nodes. The high model runs steps 0 to 3 with add_noise on and leftover noise returned, then the low model picks up and runs 3 to 6 with add_noise off. ModelSamplingSD3 shift is 8. The reason 6 steps even works is because SmoothMix is a distilled/accelerated merge, so the speed is baked in and you don't need a separate speed LoRA. I also run the SmoothMix Animation and Style LoRAs at 0.2 each on both high and low.
Now here's the part that'll fix your stutter. I generate 81 frames (about 5 seconds) at a 16fps base, and then I run RIFE interpolation (rife49.pth) at 2x and output the final video at 32fps. That interpolation pass is where almost all of the smoothness comes from. If you're just playing back the raw 16fps render with no RIFE, low motion clips like these are always gonna look choppy because there's so little movement to begin with. Just make sure your multiplier and output fps line up. RIFE at 2x should pair with VideoCombine at 32fps (16 times 2). If you do 2x but leave it at 16 you get smooth slow motion, and if you set 32 without interpolating you get fast and stuttery.
For upscaling it's nothing fancy, just a 2x lanczos ImageScaleBy on the frames before the RIFE pass. No model upscaler or anything.
So if your stills are clean but the motion stutters, adding that RIFE 2x to 32fps step is basically 90% of the smooth look. 🤙
@DeskGrenade you're a genius. How did you figure that smoothening technique for the video?
So the video is looking so much better now but your Holli stills look way crisper than what I'm getting and I think I'm missing a step. A few Q's if you're cool sharing:
1. For the image checkpoint — are you on the ANIMA-V1-Turbo-AIO straight up, or do you stack anything on top of it?
2. Your gen data shows "22 nodes" — looks like there's more than a basic txt2img in there. Are you running a hires-fix / upscale pass (latent upscale + 2nd sampler)? If so, what upscaler, denoise, and steps?
3. Those {__sensualposes3__} / {__randomangles__} wildcards — are those just pose/angle text lists, or is there quality stuff baked in?
4. Any detail/quality LoRAs beyond the Holli one?
Trying to match that clean look — your detail is unreal 🔥 thanks 🙏
I'll share a post with my updates!
@bacuss haha those are my wildcards I use for comfyui you can ignore those , if you wanna know my latest workflow go to my newer image posts cuz I simplified my image generation workflow a little bit. For the videos you don't need to worry about that. for the holli one yes I used ANIMA-V1-Turbo-AIO I just used cfg 1 and and 8-10 steps it works for this checkpoint trust me 😆
You can also use my latest images and put that in comfyui just change the checkpoint to ANIMA-V1-Turbo-AIO and set the CFG to 1 and steps to 8-10
@DeskGrenade that worked like a charm!! You're the big dog.
The Turbo-AIO tip you gave me was spot on. Running ANIMA-V1-Turbo-AIO at CFG 1 / 8 steps cleaned up my Holli stills a ton, exactly like you said. 🙏. I posted a new post with details.
Quick update on what I've been building off it: I got her talking now — feeding a still + audio into Wan 2.2 S2V (sound-to-video), with the voice done locally on an open-source TTS so no API. Also been doing 2× upscales on the stills with RealESRGAN (anime model) to post the high-res versions.
One thing I'm still chasing though — your stills have this extra detail/richness to them that mine don't quite hit, even with the Turbo-AIO. For example around the ass and vagina. How are you adding that? Is it a hires-fix / refiner pass (and if so, what denoise + upscaler are you running), a detailer, or something in the prompt/wildcards? Would love to know how you're getting that final layer of polish.
Appreciate you!
@bacuss I'm glad that's working! It might be the checkpoint because I don't use the TURBO one lately I use my own I made https://civitai.red/models/2705138/cosmic-grenade-mix-anima i'm not sure if ANIMA-V1-Turbo-AIO has that built in? But no, I'm not using a high rez fix or anything on that. Just as long as you put like "pussy" or "ass" in the prompt depending on what you're looking for it should focus on it more