Watch the tutorial video on how to build an anime-style short using this workflow:
A no-nonsense high level T2V / I2V / FFLF / REF2V workflow for Minimax H3 with lots of options & togglable quality of life features.
Toggle between the [T2V / I2V / FFLF] model & the [REF2VA] model easily
Added Forced Custom Audio: Lipsyncing to custom dialogue/music is now easy for T2V/I2V/REF2VA!
Temporal Upsampler 2nd Pass Option! [https://github.com/matlowai/ComfyUI-MAINodes]
New Image Loader allows cropping/setting max megapixels directly inside node: https://github.com/obvpm/comfyui-obvpm (Must install via unzipping to custom_nodes folder or through comfy-manager's Install via Git option, it's too new to be on the comfy-manager's index)
Kijai's Preview Override (put https://huggingface.co/Kijai/MiniMax-H3-TAE/blob/main/vae_approx/taeh3.safetensors in /vae_approx/ and set as the custom vae for non-pixelated previews)
Speedup: Sage-Attn + New Sol-Attn [https://github.com/kijai/ComfyUI-SolAttn_triton]
Speedup node: EasyCache / Spectrum [https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3]
Turbo LoRA (t2v/i2v but works with ref2va if you don't mind [some] audio degradation)
New (8/11/26) Lightx2v 1.0 release [https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main]
Film VFI Frame Interpolation (24 fps -> 48 fps)
Easily togglable reference fields: 4 pictures, 2 audio, 1 video
Use this workflow if you:
Are an AI Filmmaker who values character/scene consistency
Want to be able to use FFLF effectively, seamlessly extend videos, or use character reference sheets.
Want to edit videos, or copy & use motion from reference videos
Want fine-tuned control over your shots visuals and audio.
If you wish to generate simple one-off video clips like Will Smith eating spaghetti, the t2v/i2v model that you can toggle to in this workflow will do that for you.
Kijai's int8 convrot video vae: https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/minimax_h3_video_vae_int8_convrot.safetensors
UPDATE YOUR COMFY CUDA VERSION TO 13.0. If you start comfy and see cu130:
[INFO] pytorch version: 2.13.0+cu130Then you are good to go! But if you are using cu126, then ALL your gens with the best version of Minimax's model (INT8 convrot) will be 2x slower than they should be due to inefficient comfy-kitchen operations! Keep in mind if you update your cuda, you will need to reinstall Sage Attention! Download the correct wheel from here https://wildminder.github.io/AI-windows-whl/
ALL MODEL / NODE LINKS ARE IN THE WORKFLOW NOTES OR EASILY INSTALLABLE THROUGH COMFY-UI MANAGER.
If you appreciate what I'm doing, please consider following/subscribing on my patreon (free) which gets access to all my work early.
https://www.patreon.com/cw/foxfuressence
Also my youtube where I make AI filmmaking tutorials plz & thx:
Description
Edit: 8/16/26: Fixed bug where Audio 1 wasn't plugged in to H3 REF2VA node.
Adds Comfy-Kitchen Attention Speed Up
Adds Temporal Upsampling De-Rope (fixes fast motion blur/smudge) [https://github.com/matlowai/ComfyUI-MAINodes]
Replaces RIFE with built in comfy node Film VFI
Adds Model Shift node
This is a pretty large update. The Temporal Upsampling 2nd Pass option improves action scenes drastically.
FAQ
Comments (9)
Thanks for making this great workflow. i used to use it from the beginning,. it's greatly designed. and i would have the following suggestion. actually i modified it to add the following for my routine use. and i guess you might also interested.
1. Add an LTX 2.5 upscale 3 steps for FL2V with original prompt input before RFIE, it can efficiently upscale the output quickly. The overall time for high resolution output can be reduced,
2. Addition of optional (default off, it's too slow) SeedVR2 Video upscale (have modified INT8 node). This is very slow, but just in case someone want to output high resolution video without much loss of quality.
3. Addition of clean VRAM step and RTX upscale after RFIE, So the resolution can "Quickly" increase, this is fast, but nvidia card only. not for ROCm.
Anyway, thanks for updating the workflow to 2.0.
And i have 1 question. can F2V do lip sync with audio input? it doesn't seems work... but ref2va is ok.
I have thoroughly tested LTX 2.5 upscale and it is extremely bad for several reasons: 1. It is prone to the same problems LTX has (any high motion shots turn into smudge/screen tearing), 2. it requires an entire other model load which cuts down on gen speed overall, and 3. the prompts for minimax look so different than the prompts for LTX that you can't even really gain any guidance benefit.
I am also not a fan of SeedVR2's video upscale, having tested it thoroughly. It's jittery and even with large context sizes (batch size I think they call it?) it's so clearly unintelligent about what it thinks blurry objects are that the extra detail makes the output worse. Same with RTX upscale.
Ultimately if I add an upscale it will be what I am currently testing: Upscaling the latent directly after 1st pass finishes, and use Minimax to upscale with just a few extra steps.
And yea, using the Force Custom Audio option you can make both models lip sync really easily. Just toss the line of dialogue into the prompt between <d></d> tags, enable Audio 1, load your line of dialogue, and switch that Force Custom Audio option on. It overrides all other audio tho, so be aware of that.
One more suggestion. this is also one of my modification from your workflow.
Since the turbo LoRA for FL2VA and REF2VA will be two different LoRA. when i switch from FL2VA to REF2VA, i usually forget to switch the LoRA too.. i suggest to load the LoRA before the Anyswitch, so you can make two different LoRA node for different model.
I thought about that but ultimately didn't want to force the issue for anyone. Most people aren't swapping between models all that much, it just felt a little too hand-holdy for my taste.
It looks good. But all the necessary parameters are scattered in random places. And... I can't find the node to set the animation time for FFL2VA.
And it’s not clear whether FFL is a typo or is it different from FL?
It's all in the green group node area labeled "Video / Prompt Settings." The node is "Video Length (seconds)". I am aware this is a fairly complicated looking workflow at first. It isn't randomly scattered at all. Each color-coordinated group is one theme: Red on the left is optional functions, green is prompt/video settings, and yellow are user inputs/references. Everything is toggleable and flexible, so there's no avoiding SOME initial confusion. And yes, FFL is just a typo for FFLF, thank you for pointing that out.
Is it normal for reference to video while using a video reference to take SOOOO long to render? I used your last updated workflow so not this new one but I have an rtx 5090 and at 0.7mp trying to copy a 11sec video took me 25 minutes I believe 😭
Are you using the new Temporal Upsampler? If so then yeah, that adds a bunch of time to a gen because it's literally going back through it and regenerating the smudged/low-quality parts.
If not, V2V editing just straight up takes a long time with this model. If you're referencing a long video (over 24 frames) then yeah, it can take that long. Video edits of 96+ frames or more, easily.
@foxydits so there’s no way to really go around that then? Looking forward to trying your latest workflow as I have really liked your first 2, because the regular I2V is really fast