Video - https://youtu.be/rZnZcwobcR8

Overview

MiniMax H3 is super fun and currently the BEST OPEN-SOURCE video model you can run at home, for free on your own computer. Because it is SO AWESOME, I've been busy working away and plugging in new features to power-up my Epic MiniMax H3 Workflow!

New Sparse Linear Attention (SLA) LoRAs are now available from Lightx2v, and whilst I don't have an Nvidia 5090 to test them I'm sure some of you do! After a fair amount of testing, I found SLA was indeed faster on my 3090 - and you don't actually have to limit yourself to using the SLA LoRAs. Thus, for a quick and easy speed boost you can choose to use SLA. Seems faster than using plain comfy kitchen or sage alone!

For SLA, I'm using https://github.com/PlagueKind/ComfyUI-PlagueKind-Nodes which also provides a node to fixes issues others might have had related to "ERROR lora ... adaln_proj". I've not seen such issues myself, but then I've only been testing the various Turbo LoRAs and not any others. Either way, they have kindly provided nodes for both SLA and AdaLN.

There is also a new v1.1 Lightx2v version of the Turbo 4-step, which is my current go-to turbo LoRA. Worth grabbing if you haven't already done so and seems to work with SLA too!

The Video Extend section of the workflow has had a complete overhaul. First up, it now uses the Reference To Video nodes from ComfyUI. As it happens, the FL model can handle references to some degree as well. This means you can now use either model for extending, and you can use reference images / audio / video as well!

Because you may opt to use the FL model and start with First/Last (or custom) images, I've also added an option to no longer pass those images in to the Reference To Video nodes. That way you can still extend a First/Last image generation, but no longer have the images as references... if you don't want them. Or keep them if you like.

Using the MiniMaxH3StreamLiveExtensionAVToVHS node helps to clean things up a bit - but at the cost of being able to easily use the interpolation group because the concatenation happen "in-node". To this end, I've set the defaults for the interpolation group to pull from Reference Video 0. That way, if you do ever want to interpolate a video, you can just load whatever you want in reference video 0.

As MiniMaxH3StreamLiveExtensionAVToVHS does make things a but less cluttered, I've gone to 3 extensions instead of 2. The node supports another 3 if you fancy.

For the 3rd extension I've also added an example of using a re-reference. This can be handy in many cases. It could be that perhaps you want to introduce a brand new reference image, or maybe you want to include a previously generated video frame as a reference. The latter case is how I've set up the example. This way, you can generate a video the FL model, zoom in (losing character details such as lower body), and then when you zoom out again it can restore the generated character you had no references for - prior to them being generated!

Extending existing videos still works OK, but there is now an input limit of 124 frames for this type of extension. If you're cloning voices in videos, I think the Ref model works best, but FL will work as well. There is also now a switch to choose between extending something you generated, or something loaded via the video reference 0 group.

Though a lot slower than Euler, Seeds3/beta is now also my favourite sampler/scheduler combo for turbo models. Pairing that with a Turbo LoRA strength of around 0.6 and 6+ steps seems to give a good level of image clarity and sound quality when compared to Euler. Do let me know if you have your own favourite combos!

Also, as a number of people have expressed confusion over switches, I've also updated the notes to explain them in more detail. Do let me know if you need more assistance there!

I've also been working on more examples to help you get the most out of this amazing model, so do check out the video for some extra creative ideas. Blade Runner style "image enhance" is super fun!

Workflow Information

To open the workflow, either drag the .json file into ComfyUI, or use File -> Open (ctrl-o).

Update ComfyUI

As always, an up-to-date version of ComfyUI is required to use current features. As with all software, older versions will not have support for new features.

Personally, I update both ComfyUI and all my custom nodes each time I start because updating is part of my startup script.

Model Downloads

There are a wide variety of model sizes available, but I've been using the INT8 / NVFP4 versions because they are smaller (and faster)!

Links are provided in-workflow:

Note the NEW Turbo LoRA (V1.1), pdd_acc and the SLA LoRA!

Custom Nodes

Only a few custom nodes are required:

The PlagueKind nodes are pretty cool as they do have some images nodes worth exploring too if you feel so inclined! PDD-Acc is very optional, so if you're not into using those LoRAs then it is just a single node. I've included a version without PDD-Acc as well.

Model Loaders

There are 2 MiniMax H3 models - one for References (REF) and one for First/Last frame (FL):

Be sure to select the correct unet_name for each group! The fl2va model should be the one selected in the "Load Model - FL" group:

The ref2va model should be selected in the "Load Model - Ref" group:

Note: I'm using the pruned_int8 versions, but you can use whichever model is best for you!

I've also added an extra note with details about which models you should select in each group:

For the LoRA loader groups, I've switched to the PlagueKind nodes:

As mentioned, I've not seen any errors relating to dense LoRAs myself, but the H3 AdaLN LoRA fix is there in case you have. Do let me know if its something you've had issues with.

The LoRA loader node also features an easy on/off tick-switch so you don't need to disable the LoRA loader. If you don't want to use the LoRA - simply untick!

If you do enable the H3 SLA Attention node (Spare Linear Attention Attention?), there are a number of things to be aware of. The main thing being the dense_backend option which mirrors the settings available in the Model Options group (pytorch, comfy_kitchen, sage, etc). If you do enable H3 SLA, then the options for the dense_backend may get overridden in Model Options.

For example, with the SLA node disabled, you pick the attention you want (comfy kitchen, for example):

With SLA Enabled, the dense_backend attention is selected in that node:

The same options for sage, etc are available:

You may want also wish to experiment with the sparsity ratio. Perhaps 0.85 is better, for example.

Model Options

There are various model options, the first of which is a Fast Group Bypass switch to let you switch been the First/Last model and the Reference model:

The model loader group you aren't using will have the nodes in that group turn purple, which means they are bypassed.

I've included two Sage Attention options - the original Patch Sage Attention KJ & MiniMax H3 Mem Eff Sage Attention Patch. Either options give a nice speed boost, but the original seems to change outputs less. If you haven't installed sage attention, simply flip the switch to bypass the Mem Eff one, or set the original one to disabled.

If you don't have sage attention installed (or even if you do!) then the Comfy Kitchen Attention is another alternative to do much the same thing:

ModelSamplingMiniMaxH3 gives shift values like we're used to with image generations. The defaults are just fine, but if you like experimenting, that node is for you!

MiniMax H3 Low VRAM Attention also does what the name suggests, and combined with the MiniMax H3 Chunk FeedForward node can offer considerable VRAM savings!

Token Counter

Note that you will get issues if you go over the 160K token budget. This can happen with extreme settings such as 2mp, 16s videos with audio and image inputs. The H3 Token Counter group will give you an idea of token usage:

Video Settings

I've moved the Video Settings group (shortcut key 4) next to the Ksampler for ease. Now you can see the steps, length, sampler, scheduler, etc all together.

You may wish to start off with a 0.2 megapixel resolution to see how things go (this is quite fast and a great way to test prompts). The float node for megapixels means you can now use more floating point accuracy, such as 0.98 (rather than just 0.9).

Note that video rendering can take a while depending on the length & resolution you use, as well as your own computer hardware.

The KSamplerSelect and Steps settings will apply to all Ksamplers.

Prompting

Only a positive prompt is required (no negative prompting):

The Qwen3-VL model has decent understanding, so prompts can be quite complex! The prompt guide is also included if you like a long read, but you don't have to follow that at all (and in some cases it may be better not to).

Input Options

There are a lot of input options!

As usual, the switches will toggle between "yes" and "no". If a group is set to "yes" then it will be used. In the example screenshot above, the group for IMG_0 has been enabled ("yes").

Various notes also provide additional information on switches:

Say, for example, you instead wanted only a Last Frame image (assuming you're using the FL model), then you could say "yes" to the Last Frame group only:

You'll see the nodes in the group(s) you said "yes" to will no longer be coloured purple (aka bypassed) - note how Load IMG_0 is bypassed! Don't forget to pick the image you want to use as well.

You can change the image size using a number of methods:

The default resize_type is scale total pixels, but there are other options.

Note that the Ref model will scale images anyway, but with First/Last you may wish to change images to match your video resolution. It is a good idea to have image aspect ratios which match the resolution you select for First/Last images. A square image, for example, will get stretched out if you use a non-square aspect ratio for the video.

Custom First/Last Frames

When using the First/Last model, it is now possible to pick a keyframe index for your images. Maybe you want a particular image part-way in, or perhaps a mid-position? Pick your own custom First/Last images and positions, then enable the option to use them:

As you can see it is wired up for 3 images, but you could have 1, 2, 3, 4 or whatever!

Inputs

You can use more images if you like, but I've wired up 4 to start with:

There are also 2 x custom video inputs, and 2 x custom audio options. I found on occasion that certain videos without sound could confuse the VHS nodes - but the addition of the Any Switch seemed to prevent such errors. If you do get an error saying that you video doesn't have sound (and so the Video node doesn't want to play ball), simply disconnect the audio output noodle.

The voice cloning works great with 5 second audio clips, but may also work fine with shorter durations. Personally I think it seems to work slightly better when you use a video with audio over just a plain audio clip, but plain audio does the job too.

Input Image Masking

Input Images now have options manual/auto-masking:

I found that sometimes it was a bit tricky and cumbersome to write prompts telling the model to change certain features - like if the image is wearing a hat or a particular costume and you don't want that part. Masks get rid of this by allowing you to send in only the data you're interest in, such as only the face, hat, a particular bunch of flowers or whatever.

By default the whole image will be sent in as usual, but you can opt to use your own manual mask or an automatic one if using a mask at all. Flip the switches to match your requirements.

KSampler

Unless you're using the "NoPDD" workflow, the initial KSampler now also has an option to use the PDD LoRAs:

Personally I'm not that keen, but your mileage may vary!

Base Output

The output selection uses the Video Combine node, so you can adjust the various options there, such as for generating PNG or GIF outputs:

In the example video face-swap above, you may find the generated audio isn't as good as the raw audio going in. If the audio is meant to be exactly the same as one particular input, then the "Use Raw Audio?" switch is for you! Make sure to select the audio you want to use (in the example screenshot above it is set to REF_AUDIO_0, but it could be any audio input you have enabled), and flip the switch. The generated video will then use the audio exactly as it has been input - no processing by MiniMax H3 at all.

Context Extend

If you want to make your MiniMax H3 videos longer than 15s, or extend existing videos, you can do so in various chunks:

Do note that if using an existing video, only up to 124 frames are allowed by the H3-Motion-Context-Multiref nodes, so set the frame loading cap as required.

Three extensions are included, but you can just copy & paste the various groups to add 4, 5 or 6 extensions.

Though it uses the Ref model conditioning nodes, you can actually use either model when extending - and the FL model can even handle references to some degree!

Each extension has a new prompt; Prompt 2, Prompt 3 and Prompt 4:

There is also a final video combine group which puts all the chunks together:

The active_extensions controls how many extensions will actually be used/combined:

NOTE: If you updated the custom nodes after 30/Aug/2026 use v18 as the Final AV Extension node now needs an output sink node attached.

If previews are not enabled, only the required extensions will be generated:

If you do enable previews, the extensions will be generated even if you aren't using them as active_extensions.

Frame Interpolation

Should you wish to go up to 48fps (or more), then enable the Frame Interpolation group and download your choice of RIFE/FILM model. Change the value in the FPS Multiplier for even more FPS goodness:

By default, it is set to interpolate from reference video 0, where you can load whatever you like. By disabling the initial generations and making sure video combine isn't active, the only processing that will happen is the interpolation.

The Engine

If you do want to play with the engine, it is up and outside of the usual map:

The MiniMax H3 Reference to Video node can take a few more inputs - so if you want more than 4 images, for example, then connect away!

Enjoy! :)