(More demo vids inbound)

Turns expressions into a transparent-animated AVIF files - works with SillyTavern. This is an extension of my SillyTavern Anima Workflow.
(This workflow is designed for the more knowledgeable ComfyUI user.)

Update Notes (0.3):

Qwen Image 2.1 Turbo (6-step) in combo with BiRefNet (Toonout) background remover now used. Qwen IMG is super clean, but time consuming, even with TURBO. So, I've also added frameskip hacks. It actually works fairly decent up to 3 frameskips. 4 frameskips and things start to morph, 8 frameskips is pretty rough but good for testing. Qwen Image is great for background removal, like really good, but not perfect, still leaves edges, and sometimes just doesn't work, so we use it in conjunction with a safer background removal tool. In the end, this method, even with frameskips, still doubles or triples the time of version 0.2. To help with time saving, I now recommend using DavidAU's Qwen38TurboFusion735 (Linked below), it's less accurate for roleplay, but it's faster and for our purposes, does it's job. On my hardware v0.2 completes in ~5 minutes, v0.3 takes about ~13 minutes with frameskip set to 2, but the result is significantly cleaner.

Requirements:

• SillyTavern Character Expression Generation Workflow

• Soft requirements (In other words requires tweaking workflow to avoid using these):
  - KJ's FP8 MiniMax H3 (Ada+Blackwell generation only)
  - Minimax-h3-Turbo
  (Use the int 8 version of Minimax and it's turbo lora if Ampere)
  - Better Human Motion Minimax (SFW)
  - Slop Bounce (SFW & NSFW)
  - 2D Anime Style (NSFW)
  - Mystic XXX (NSFW)
  - Faster! Harder! (NSFW)

• DavidAU's Qwen38TTurboFableFusion735 + mmproj-BF16 (Or your own LLM and mmproj) - Place in ~/.../ComfyUI/models/LLM/GGUF

• Disawa Nodes - On Nodes 2.0, this requires either lowering your security level or git cloning it.

#!/bin/bash
cd ~/.../ComfyUI/custom_nodes
git clone https://github.com/darksidewalker/dasiwa-comfyui-installer.git

• v0.3 and Newer: Viggle's Sigma node: [ Download ] & Viggle's turbo 6 step (best version is int 8 convrot) [ Download ]

#!/bin/bash
cd ~/.../ComfyUI/custom_nodes
wget https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/raw/main/comfyui/viggle_turbo.py

Operation:

Anima workflow, select Mode 8
Then run
Take the picture it kicks out, with the minimax prompt and plug it into this workflow.
(Minimax promp can also be found in *.json 'negative prompt' section - if doing multiple)

Should look something like:
Lewdness: Low

Expression: Joy
Character Tags: Elf, blonde, tunic
Character Actions: laughing, smiling

(Feel free to tweak the input as you like, but leave lewdness and Expression alone, it's used in the logic of this workflow)

Technical Note:

This uses ComfyUI-LLM-text-processor node set. Means you need llama.cpp installed inside of your ComfyUI environment. (And other things like CUDA ## but the llama cpp might give you a fit)

#!bin/bash
cd ~/.../ComfyUI/custom_nodes
source ./venv/bin/activate
python -m pip install llama-cpp-python
# Run ComfyUI
python main.py --listen --port 8188 --enable-manager

Linux Users (Like myself): LLM node might point to the wrong binary, paste:
llama_cli.py & llama_binary.py
from:
.../ComfyUI/custom_nodes/ComfyUI-LLM-text-processor
Into your favorite AI bot and ask it to fix it to work with your distro of linux. Personally, I had to SED a few lines to get mine to work.

Two, alternatively, you can opt for another node set to replace the 'LLM' node, such as the QwenVL Nodes, which come self-contained with auto downloaders (But may require you to figure out your own system instructions and Setup).

Second Note: Interpolation is done by tensorRT, CUDA acceleration. AMD users, you can swap these out with other ones if you need to. This workflow is yours to tweak to your specs.