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
Adds Kijai's Overide Preview as main sampler preview
Adds Video/Audio Shift as optional node
Changes the Load Image node to a custom node that allows cropping/resizing options easily within comfyUI: https://github.com/obvpm/comfyui-obvpm
Fixes a bug or two where efficiency nodes weren't connected correctly
Defaults to Kijai's int8 convrot vae now that it works with updated Comfy.
Adds 2nd Audio ref (because sometimes you need two voice references, amiright?)
FAQ
Comments (55)
Hi! Where does the "LoadImageCrop" node come from? Can't find it anywhere. Thanks!
In the past 5 minutes I've posted it it in all of the descriptions here on civit, but originally it's in the workflow itself, in the big bright red note labeled "Custom Node Links" on the left side.
@foxydits Ah I completely missed that note! My bad. This workflow is absolutely amazing though, thank you!
@ldpz67 No worries, I need to add node links everywhere because everyone looks in different places. Hehe, just how it is.
where can we find the new LoadImageCrop node as mentioned in your changes?:
"Changes the Load Image node to a custom node that allows cropping/resizing options easily within comfyUI"
its not showing up in the install missing nodes
Not sure how you're not seeing it tbh! It's in the description of this workflow page, it's in the big bright red note labeled "Custom Node Links" in the workflow itself, and it's in the version information you were quoting from! It's also right here: https://github.com/obvpm/comfyui-obvpm
@foxydits thank you , not sure how I missed it!
I’ve always wondered about the hype surrounding Minimax H3. Now that I’ve tried out your workflow, I’ve finally been able to generate something decent in 2K. Thanks for your efforts.
First off, thank you for making this workflow. You make it easy for those of us that are less technically advanced. That being said, any possibility of incorporating some sort of grid for reference images or a director node? Theres info on reddit where you can add 9 individual reference images to a grid, and then use that grid as a single reference image to the "MiniMax H3 Reference to Video" node. Im sure I could figure that out myself, just wondering where the next versions you make are headed. Thanks again!
Meaning, you could theoretically input more than the 9 image reference limit, since the grid is a single reference source (containing 9 individual references. 9x9 = 81 reference images)
You can do that yourself of course, just by photo-editing your images together. I don't see the purpose of that functionality for adding to my workflow; 9 image refs is already a ton, and a single ref image can only be so big in dimension before it slows the model (2048 is maximum short-side resolution). I'm not sure about the director nodes, either.. I haven't seen one yet that seems better than the old fashioned way. If you do let me know.
I'd really like to give this one a shot, but am struggling to get all the nodes working.
It claims missing node packs for 'comfy-core' which includes the MiniMaxH3ImageToVideo and the MiniMaxH3ReferenceToVideo nodes.
The RIFE Interpolation Node and SolAttnPatch are also missing.
Searching the manager does not give me anything, and while I've tried multiple times to git clone KJnodes, it doesn't seem to change anything.
It doesn't help that it says kijai/comfyUI SolAttn_triton is already installed but the node still isn't working.
Any direction helps.
GPU: NVIDIA GeForce RTX 5090
PyTorch: 2.11.0+cu130
CUDA: 13.0
OS: Windows
ComfyUI: Windows Desktop
I've only ever used the Portable comfy so I'm not best to help you. However, it sounds like you have bigger issues if you're missing comfy-core nodes. Update your comfyUI, because minimax IS built into comfyUI and your version is saying it isn't. All of the github links to other nodes are in the big red workflow note on the left side labeled Custom Node Links.
Past your errors to chatgpt - i found out i had many issues this way.
I only encountered comfy-core issues in one of two scenarios
1. for some reason, sometimes the new comfy update breaks workflows when switching between them (perhaps other triggers as well, not sure) and then it breaks
2. modifying the blueprints in some ways broke it as well and after a restart or refresh it was giving me comfy core issues
@foxydits I'll try updating. It was a little odd, because I originally downloaded the base version of Comfy, and then when it updated, it changed to desktop. The opening UI, version launching, and all that was quite surprising. If the update doesn't work, I might have to try a fresh 'non-desktop' install.
Excellent work. Thank you. It does a great job with video continuation. Is it possible to add an option to append the generated video to the source video?
Hey thanks! But you owe it to yourself to use a proper video editor like Davinci Resolve (free) to do such things. It will let you finetune the continuation's lighting/framing for even better accuracy. This is workflow designed for serious AI filmmaking, so it's all about the genning part, not editing. Hope that makes sense.
Could you also add a last frame feature? I'm terrible at this and I want to make long videos using last frame.
It already has that option. <Picture 2> is last frame when you're using the T2V/I2V model option. Just make sure to turn on the Picture 1 and Picture 2 resize so it crops to the correct aspect ratio if your pics aren't already sized properly for your output video aspect.
I know u have said it several times, about the red brigh box, but I cloned those repositories and they still arent detected by comfyui, for example obvmp says thatit doesnt need further installation after cloning but ComfyUI doesnt detect it, it says that it is still missing... hope you can help me
And you're sure you're cloning it to the correct /custom_nodes/ folder yes? If comfyUI's not detecting it, my first suggestion is to always update to the newest comfy version. The obvmp Load Image node doesn't rely on any python dependencies or anything (it's a really simple node) so if installed correctly there can't really be too much that goes wrong. If when you start comfy you see in the logs ANY comment about it, let me know. Might be something like (FAILED TO IMPORT).
Thank you very much for your workflow.
a question(maybe i missed in the youtube tutorial):
if i have voice audio ref, and change it to a completely different voice actor (male to female for example) while keeping same dialogue?
If you have audio references and want to use someone's voice, there's a few things you'll need to include in the prompt.
in subject_definitions, after you describe your <Subject 1>, write:
<Audio 1> is the voice-timbre reference for <Subject 1>
in retention_analysis:
<Audio 1>: reference - its vocal timbre guides the vocal delivery of <Subject 1> without copying the original signal.
And for good measure, in detailed_description:
<Subject 1> speaks using <Audio 1> as their voice-timbre reference: <d>Gee, those muffins sure were great!</d>
(Or whatever your goal is)
Voice cloning in minimax is OK at best, but it's what I used for my entire short that I posted. With it, you can change any character's voice to anyone though, which is nice.
@foxydits Have you found a way to "weakly" sample audio, and have a voice "similar to but not exactly" like the reference? I was reading the official prompting guide, and it does seem to have a weakly_reference option, but so far it only seems to work for visual reference, not audio. I would like to take a voice sample, and have it come out either younger/older, or higher/lower, or with a different accent, but prompting it as such has a VERY weak effect, when using audio reference. It'll still sound 98% like the reference.
Is it possible to take two audio refs and blend them together into a single style/voice? That could give me the effect I want I think.
@lynxo I'm afraid I haven't attempted that. I will say that I have found the voice cloning to be 'not exact' enough that you could probably modulate it with prompting.<Audio 1>: reference - its vocal timbre guides the more youthful and cheerful vocal delivery of <Subject 1> without copying the original signal.
Try modifying that line to denote the changes you want.
hi guys update: managed to do it. (even with a VERY language that a lot models fail, and it did perfectly, but you need nikud in the text prompt, if you know you know what's the lang)
recorded myself (audio 1)
voice clone (audio 2)
image ref (Image 1)
(didn't test t2v though , i assume will work also):
subject_definitions:
<Subject 1> [describe your subject from image1] in <Picture 1>
<Audio 1> is the dialogue-content reference for <Subject 1> (S1), providing the exact spoken words, phrasing, and delivery pacing.
<Audio 2> is the voice-timbre reference for <Subject 1> (S1), containing a spoken [English NOTE this is the clone voice, so if its talking other than english you should tell him what lang the voice is] vocal layer with [describe the voice of character - pitchy, low and bass etc, if its famous character, noting it i think the model will know about it].
summary:
[reference generation + audio reference] <Subject 1> delivers the line carried by <Audio 1> in the vocal timbre of <Audio 2>, [one clause on the setting and the beat of the shot].
retention_analysis:
<Subject 1> (appears in [Shot 1]): fully_preserved - the facial identity, hairstyle, and wardrobe from <Picture 1> are retained.
<Audio 1>: reference - its dialogue content and delivery pacing are reperformed without copying the original signal.
<Audio 2>: reference - its vocal timbre guides the spoken voice of <Subject 1> without copying the original signal.
detailed_description:
The target video is in a [describe scene background].
[Shot 1] [type of shot medium/full/birdview etc] on <Subject 1> (S1), [action he/she perform], [add relevant if need new character, i didn't do subject 2 and it worked good]. [some description if camera dolly's in out/pan acrorss etc] Speaking in the vocal timbre referenced from <Audio 2> and reperforming the dialogue content of <Audio 1>, <Subject 1> (S1) says with [expression: joy, anger, fear], <d>[Spanish]Papá, perdón, Solo quería echar una mano </d> [describe the end of scene, the closure].
overall_soundscape:
[whats going on with background sound]
non_diegetic_music:
N/A [or some music you prefer]
Loving your workflow, currently my main workflow. Hoping that you would add many more features and optimizations in your workflow. Keep up the good work!
I plan to! Thanks for supporting.
Amazing!!
None of the turbo loras seem to be working for me. I thought you needed a special lora loader for the experimental turbo loras to work? Would power lora loader really be enough?
They work just fine. I've seen no evidence that you need a special LoRA loader. Just use Lightx2v's new 1.0 release (came out 1-2 days ago). It's the top one linked in my wf/civit description. I even used it with ref2va for my most recent short. 8-10 steps, 1.0 strength, euler/beta57.
Too many unknows nodes in this workflow. Based on Comfyui I have all the necessary Node Packs but still keeps throwing node related errors.
Sorry, they're all necessary or provide too much value to leave out. In the end I don't consider this a beginner friendly workflow; some of the nodes must be manually installed since they're not in the comfy registry yet due to the newness of the model.
@foxydits can you provide a guide to install all of them, I really want to tried but as a newbie, there is too much error on every nodes :(
ITS BCUZ U ARE NOOB. IM USING THE WORKFLOW 100% PERFECTLY, NO ERRORS, NO PROBLEMS. LEARN HOW TO USE COMFY AND HOW TO DO MANUAL INSTALATIONS.
@trankhacvinh1991 I'm not sure which nodes you're struggling with. This is a cutting edge newly released model and an advanced workflow so without specific error logs I don't know how to help. I will if I can but without anything to go off of.. you know.
@foxydits I think these are the missing nodes. LoadImageCrop
SpectrumApplyMiniMaxH3 .
I too am having issues with LoadImageCrop, ComfyUI says the pack is unknown so we cannot go to the Install Missing Packs in the Manager.
Same here with LoadImageCrop. I've temporary replaced it by a simple load image, but I'd like to know where we can find this node.
@KinkReaper it's right here https://github.com/obvpm/comfyui-obvpm -- the link is pasted everywhere I could think to paste it.. description, inside the workflow, version notes
@YaFellowDegenerate As explained in the description of this workflow & in the workflow itself in the "Custom Node Links" note, you install it manually by unzipping the contents of the node into your /custom_nodes/ folder. https://github.com/obvpm/comfyui-obvpm
@foxydits Thanks, I really didn't see that. I've been working in front of the computer all day, my eyes feel like I stared straight into the eclipse. 🙈
AMAZING WORKFLOW AND CONTENT!!!! WORKING 100%
I've been using this since the first day you posted the first version, but am now finding 4 images to be limiting for two-character scenes, to have the right hair/clothes etc for both of them. Can you make the next version allow say, 6 images for reference?
Is the 'ref image size' supposed to be set to 'max' as default on this workflow in the main reference to video node? Seems like the only way I can get this workflow to run anymore is if I set that to 'match.' 16 GB vram wouldn't be enough to run 'max?'
Thx.
Depends on what size your references are. If your reference images are really large, or you have several of them, or you're using video reference too, things get a lot slower. 'Match' just downscales everything to your latent size. That can be okay for certain things, but 'max' with proper image size control is better. That's why the new Load Image node is useful, just drag on the image itself and crop unnecessary pixel space out if you can, and set MP to 1 or 1.5.
Thanks for the WF, I'm using 1.6 and got some general questions.
1. I'm using a reference character design and with the prompt guide, I manage to have the character show up correctly. But at the first frame, the reference image shows up for a split second, is this normal? The image is just a 9:16 of a person with a white background.
2. Is there a way to add more reference image to the WF?
It sounds like you're accidentally using the t2v/i2v node, which is programmed to make the first-frame input be the first frame of the shot no matter what. Make sure ref2va model is toggled on.
Yeah, if you need more than 4 ref images, just copy one of the Load Image nodes and then drag the link to the next open available slot on the ref2va minimax node, should be something like ref_image_4.
Also you should update to version 1.8!
@foxydits Thanks for the quick respond. I'll try using REF mode in my next test. I'm also testing video extension. While the new video took the last frame from the reference video correctly, the length is not correct, I was expecting 10 seconds + 10 seconds but it only generated 10 seconds instead of creating an extended video. I did set the length to 10 seconds and use video continuation as the task description. Any suggestions?
@Cowgary The videos don't stitch automatically, such tasks are not comfyUI's purview. Not to say that there won't be specific workflows out there which will do it, but they will be gimmicky and just for that sole purpose. [video extension] from the model's perspective is simply only an instruction to continue from X amount of frames in your reference video, generating what your prompt describes as happening next. Afterwards you would need to use a proper video editor to merge the two clips together seamlessly.
@foxydits Ah thanks for clearing that up. WAN 2.2 had a workflow that would extend an existing video so I thought H3 had that built-in. May actually have to sit down and learn how to merge videos correctly or wait for someone to build a workflow.
@foxydits Just noticed there's a H3 Infinite Video Continuation WF pack. I took a look but it doesn't have a REF mode. I'm wondering if you will adapt a similar approach used in Herrgotts' WF for your next version?
For anyone who is having an issue with black screen outputs - per Kijai,
"int8_convrot VAE needs ComfyUI 0.31.0 or you get black outputs"
Your workflow is the best one I've found thus far & it's super easy to add my own tweaks to it. Thanks so much for your hard work!