⛩️💀 DaSiWa MiniMax H3 💀⛩️
My new MiniMax H3 model.
Version overview: https://civarchive.com/articles/23495/dasiwa-model-versions-and-timeline
⚠️ Make sure to open the DOWNLOAD dropdown to see all quants possible.
🔮 Key Features
See Announcements!
⚠️ What NOT to expect
No tuning for ToS-violating concepts
Not more unaligned/altering boundaries/breaking guardrails
🍒Workflow
Make sure to checkout my easy to use Workflows!
🍄LoRA's
But: This checkpoint is not meant to replace LoRAs, it is meant to:
Perform better overall at his own
As easy as possible to use
With LoRAs to be more awesome
🛠️ Recommended Settings
Non-Distilled
res_mulistep/simple or euler/simple
Shift Video 10-12
Shift Audio 3-5
20-25 Steps
Turbo (Distilled)
euler/simple or LCM/simple
Shift Video 6-12
Shift Audio 3-5
4 or 8 Steps
Dependencies
See Workfow notes.
🩻 Known issues
Tell me 🫵🫢
🩺 Fixes & Feedback
Update your ComfyUI ❗
🖤 Why I Made This
Pushing MiniMax H3 to its limits!
This checkpoint is also my personal playground.
Closing words
🤩 I want to thank all the fantastic other creators who made super nice LoRAs and concepts to play with! Support that awesome creators by using their LoRAs and post to their gallery and share the meta-data!
⚠️ I made all this with permissions or open-source resources (the time it is incorporated).
I share as much insights as I can without compromising my work. I'm doing this for fun as my hobby and just do not want my hobby to be destroyed.
More details can be obtained in the corresponding announcements!
If you would like to contribute in my awesome (😉) checkpoint or willing to share resources I'll gladly give credit! Just contact me!
✅ All credits / resources are mentioned inside the announcements! - Since different versions may have different resources.
YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.
🔐 Official Permission from MiniMax
In August 2026, MiniMax granted me written authorization to use MiniMax H3 and MiniMax H3 Works via my license request. This authorization is personal to me as the author of this model - it does not relieve you of your own responsibility under the MiniMax H3 community License Agreement and its Acceptable Use Policy.
⚖ Disclaimer & User Responsibility
This model is a community-quantized derivative of MiniMax H3, distributed under the MiniMax H3 Community License Agreement (see the license on this page). By downloading or using this file, you accept and are responsible for compliance with those terms, including the Acceptable Use Policy and the laws of your own jurisdiction.
All outputs, and all use or sharing of this file, are your sole responsibility. Outputs are machine-generated and must be disclosed as such when published publicly (AUP §12). I do not support or take responsibility for illegal, harmful, or harassing uses - and I will report content I become aware of that violates the license, the AUP, or platform terms of service.
THIS FILE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.
I am not liable for any damages arising from its use. I am not affiliated with MiniMax.
If this authorization is revoked, or MiniMax objects to this community redistribution, the files and this post will be removed promptly upon notice.
📜 DaSiWa Custom Addendum: Fine-Tune Integrity & Attribution
Base license: MiniMax H3 Community License Agreement -
https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
This is a general-purpose community derivative. It is not a specialized or task-specific variant; its capabilities and restrictions are those of the underlying model and its license.
1. Verification & Integrity
This checkpoint is a quantized / merged / finetuned derivative of MiniMax H3. Notice of Non-Support: copies hosted on third-party platforms (mirrors) are considered Unverified. I provide no warranty, support, or safety guarantees for unverified files. Per license §III, redistributors must keep the embedded metadata, recipe files, and license notice intact.
2. Trademark & Branding
The name "DaSiWa" and associated logos or promotional imagery are the creator's i intellectual property.
3. Commercial Use
Per license §IV.1, commercial products or services generating more than USD 20 million in yearly revenue require a separate prior written authorization from MiniMax; commercial products must prominently display "MiniMax H3" (§IV.2). Using this model's official branding to market a paid or ad-supported generation service requires a separate agreement with the creator.
Description
FAQ
Comments (121)
Hell yeah, huge fan of your workflows and checkpoints - can't wait to try this one out over the weekend
# omg
Wow, that's fast!
Hybrid? Should we use this model for ref2v and fl2va? Cant wait to try it
I had no idea how to use Minimax. Then I saw your WF. I still have no idea how it works but it does 10/10!
Downloading to test... maybe a few recommended settings would help, just saying. going back when i have some real feedback
Oh I'm ready to see what we have cooking here! Lets see if my Hazel scenes can be 100% H3 yet!
I gave it a try right away.
It’s excellent work.
It looks like I can finally say goodbye to Wan2.2.
Any chance to get a bf16 vision?
Oh look its Christmas again
Did you build the Turbo lora into this or do you recommend the 20 step res_multi?
20-25 steps, no turbo in. The turbo's are not ready yet
Not only it's Darksidewalker's masterpiece, but also superior ref2va?
I need this 😶
Thanks for your work, you're making so much for the community
Have fun with it :)
Are you planning to make an I2V version?
this is a hybrid model so it should work with t2v/i2v/fflf/ref2v
Yeah should work for both
works with t2v and i2v as well. turbo loras destroys the quality though
both the 19gb and the 31gb are labeled int8 convrot, accidental mislabeling of bf16?
No, the smaller one is Pruned, the other isn't pruned.
@AI_DK makes sense, ty!
great job, if there are nvfp4 or w4a8 format, it would be awsome.
check the variants, there is a 4bit
@AI_DK 4bit destroys quality the only format ideal for minimax w4a8 everthing else sucks (h3 is not good dealing with 4bit quants)
@brahianvalles good chance it's w4a8 if it's listing as int 4, nvfp4 would be nf4
The int 4 is the w4a8, there is no selector for this on civitai. Hovering the filename reveals it.
Thank you very much
@Darksidewalker Perhaps you could add some annotations(int4 is w4a8) in the page, I found this answer by clicking through each discussion >_<
how well does it do feet stuff?
The turbo lora isn't included, right?
no, turbo loras aren't ready yet to be baked into Minimax H3.
Did not include, all turbo loras are unstable atm
Incredible. The new go-to. Really fast model too.
普天同庆
什么时候上存到huggingface
can it do realistic style?
Thank you ! ! !
Wow... Very High Quality
Wow! Its awesome!
Cool thank you very much, especially it is ref2va.
Could you please do a turbo version with 8 step support ? Since the 4-step Loras significantly lower the quality, 8 steps would be great ! :)
Not until turbo getting stable
I need the download channel for Hugging Face.
I’m new to H3 and I’m trying to understand the difference between I2V and Ref2VA. Are they basically doing the same thing in the sense that they both animate the input image into a video? Or are there important differences in how they work?
I2V use the pic as first frame same as the pic, Ref2VA use the pic as reference but not the first frame, for example: only use the woman in the pic but do not use the background，do not use the pose (u can define the outcome result in the prompt, like use background and pose but the out come will not 100% be the same as the pic), Ref2VA model also have ability to handle audio and video.
Another amazing work as usual. It's my fav so far !! XD
there is an issue with your DaSiWa node pack there is some invisable overlay on the screen proventing clicking on anything in the grid and zooming its somthing off to the right side i allways just have to remove your node pack since somthign is broken and makes comfyui unusable and my comfyui is up to date
It only helps to reduce the size of the node window if you move the right border of the window to the left so that it becomes to the left of the graph border to enter a parameter inside the window.
ps: sorry for the vagueness of the answer, I use a translator.
I would love to know why, but the issues you describe are not reproduce able for me that way.
The only thing I can think of that you may have incompatible other node-packs installed.
@Darksidewalker if you have any overlays of any kind or buttons or anything ui related id say remove that
@PastelPastelPastel I'll not remove the UX I made for the nodes
@Darksidewalker can you make it a toggle at least in the settings so im able to use your node pack
@PastelPastelPastel That's not possible, that is not how comfyui nodes work.
You may have to look into your setup, the nodepack UX is working.
目前用下来最棒的微调模型,配合4step加速效果也很不错,God bless you.
请问能否推荐一下你所使用的 4-Steps Lora？目前市面上的 Turbo Lora 似乎都是基于 fl2va 做的，这个模型目前似乎只有 ref2va，难道用 fl2va 的 Lora 效果也不错吗？
@Aetherlyn minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors 直接用的这个,目前看效果不错
@krezip123 Do you feel that after using this model, using 4-step LORA is slower than the original model
@oxo1314 Everything is good at present
This is incredible work! It's got just enough base NSFW knowledge for reference images to fill in the blanks, no LoRAs even needed so far. It's amazing what it can do.
As always, great work!
The REF2VA and FL2VA mixing comes with some degradation of the accuracy the models following references. especially when fully copy music or continue videos. I didn't test your model with video continuation yet but can confirm the audio issue is in your model too. Not sure if this is fixable i just want to let you know.
Thanks for the amazing work as always.
Any reference or proof to this?
@Darksidewalker any soundtrack should do it. just need to use audio as reference, set the exact length of the audio to the length of the generated clip. It will have some artefacts or other distortion, even when used "fully_copy" in the retention_analysis. Standard Minimax h3 ref2va model can do it without the distortion.
If Comfyui crashes, try this: --disable-mmap --disable-dynamic-vram --disable-async-offload --disable-pinned-memory
I tested it using only 6 steps and 0.2 megapixels, and it still turned out to be a pretty good video! I'm going to try it with more resolution and steps; I'm sure it'll be a masterpiece! Thanks!
disabling dynamic vram is not recommended since it's what makes most people be able to run h3 with shitty hardware.
@torikoko My Comfyui crashed without it, maybe it's because I have a Triton.
i love your moddels, i'll be happy for a nvfp4 one
Minimax Dasiwa is out! The best news of the last two weeks. But I'll try it again; the joy is still in the air. How do you make a video longer than 1 minute?
check the save last frame option then in your next ref2v generation define it as the first frame of the new sequence; you can chain videos together end to end
@FirstPrinciples in my first attempts at doing that at the release of minimax, the clips kept getting darker and darker. Is this still an issue with fl2va?
Also with ref2v you can apparently input a video and prompt for a video continuation based on the reference but I haven't tried this yet.
@KiraNugget Hmmm I didnt have that issue with it getting darker, you could always try to reinforce the lighting/brightness/contrast in the prompt. The video reference sounds interesting approach, but video references are just sooooo heavy
@FirstPrinciples Yeah I might try again to check. It was on the first few days when I was still testing and figuring out the settings. My guess is that with my settings now it would probably work fine but i'd have to try it :P You can see what I'm talking about in this video. https://civitai.red/images/139534069.
@KiraNugget Yeah I see it - reminds me of what Wan used to do. I haven't encountered it in H3 yet, and the only thing I can think of is if you are using an upscaler so when you take the upscaled end frame and plug it in as the new first frame it gets compressed
@FirstPrinciples ahh thats good to know. But im only using ref2v now anyway. Its easier to cut to a New angle at the start rather than trying to match the frame exactly.
@KiraNugget for sure, love them hard cut cheats
25 steps. Well nice joke 30 minutes for 10 seconds video... Then this model just useless.
4 steps can also work
@krezip123 even with 6 steps quality awful (with turbo lora obviously). 4 steps just forget.
@velanteg Maybe there's a problem with your workflow. check it, for example, don't use the NSFW clip, use the original clip instead,and int8 base model
Maybe something went wrong with the int4 (w4a8). I removed the file for now.
I'll run some test and re-up.
Thanks for trying! I'm writing this as someone why have 8GB of vram.
I did not find any problem with the int4 (w4a8) - it works, just massively lower quality depending on the use case, what is expected for a int4 quant compared to int8.
Re-up done.
@Darksidewalker The quality of minimax h3 is extremely poor at 4-bit,I have tried other 4-bit model and concluded that minimax is not suitable for 4-bit
@krezip123 That's right, but I got some ppl on my discord doing great stuff with that one here. Depending on the scene and motion. At the end it is only for extremely constrained hardware.
Pretty nice results, but it runs quite a lot slower than any other int8_covrot models I've tried, including other hybrid models. Are the files mislabeled?
I'm not aware of this, I run a 5s test video today on 177s, native resolution and on CK attention and H3 cache.
Did you compare ref2va?
The Ref2VA model typically gens 5 seconds in about 10 minutes on my hardware. The hybrid models off Hugging Face ran about the same. This one has been taking 15 - 20 minutes for 5 seconds.
I am using Sage Attention, the Ref Turbo LORA and Spectrum for speed.
@Gemini3443 I wonder what should cause that...
@Darksidewalker If CUDA is not updated to version 130, you may update it, as it delivers significant speed improvements.
@malysdean702 My cuda is up2date, the OP had problems
@Darksidewalker Okay, I missed that.
@Darksidewalker What H3 Cache settings do you use?
@Haarfus default
W4A8 Quantization PLZ
Already available (download dropdown)
Visual quality and adherence is very good, but it seems that it will not understand to use an audio track as is (music and vocals), it will only use it as a reference to create an audio track that vaguely sounds similar with giberish. :( will have to use an other model for now for my music video. But a very nice first version, keep it up! :)
Depending on ppl on my discord it works well and is seed dependent. Mind that H3 does not clone audio exactly, that's why there are nodes to do it if needed.
@Darksidewalker I'll try again but 3 different seed and they all did not work, when the same seed with eros ref2v model all worked
@Darksidewalker Yeah I tried it again with a seed that worked with the eros model, and it still gave me trash audio, I can confirm that in 3 hours of testing yesterday it did not output the same audio I gave it once. even with "audio reuse" tag in sumary, "<Audio 1>: fully_copy" in retention, <Audio 1> is directly reused as the performance soundtrack in overall soundscape, and all the proper prompting in detailled description.
By chance will you be doing an NVFP4 variant for the Blackwell owners?
This has no advantage over w4a8 (int4 CR). Also nvfp4 and int4 quants are compareable low quality to int8 CR.
But I may look into it.
Why no speed ups like your Wan models?
They are unstable atm if you refer to the turbo lora's. There are massive speed-ups in the workflow.
People are clueless there no speed here because its the base model improved and dasiwa decided there is not need to bake a speed lora (Which i agree) until a clearly better one comes out.
What point of this model if native w4a8 model works with turbo lora and give much better quality while dasiva just dont work even with lora on low steps?
@velanteg This is nonsense. w4a8 will not surpass INT8 CR.
Punctuation exists.
32G's friend!
Well, for now it's not a clear winner over minimax h3 base, but rather another tool you can use, depending on your task. It has better camera and character motion, but probably lacks a little adherence (not confirmed enough). Also it's not as strict to follow reference media, which actually can be a good thing, because it helps to blend styles. Also it means less chances, that it will leak your reference image exactly as is, which in most times is a bad thing for ref2va tasks
From my community on discord there is another opinion on better adherence and better visual stability. Maybe it depends.🤷
"pruned is same quality, but less demanding" what less demanding means ? and would you upload fp8 version of the diffusion model only for comfy ?
The pruned model eats less vram. In most cases it's ok to use.
FP8 makes not much sense here. Stay with int8 convrot.
I found a bit of trouble with the workflow for some reason it was broken with unconnected nodes. Thankyou for your work. Hopefully a turbo tuned model is on the way soon. I used it with the turbo lora & results were decent for the first release.
Do you know what was not connected? I did not notice any disconnected nodes.
@Darksidewalker a few nodes from the director, may be it was being patched when I downloaded it. I'll give it another try tonight & share if I'm able to get through it. Btw I also made a workflow for Minimax. If you can share your review & output there , it would mean a lot. https://civitai.com/models/2883915/minimax-x-ltxv-low-vram-8gb-cinematic-video-pipeline-or-4k-and-2x-fps
Am I the only one who finds that ref2v doesn't preserve the reference character?
I've been using Dasiwa's old workflow, and it worked great. I switched to this one to try it out, but no matter how hard I try, it doesn't preserve the reference model, and it also takes noticeably longer. Does this happen to anyone else?
P.S.: I can't use comfykitchen because even though I installed it, it can't find it, and h3 cache causes too many artifacts.
Translated with DeepL.com (free version)
Use this LORA: https://huggingface.co/fal/MiniMax-H3-Realism-People-LoRA
the refrence model i think is holding the model back refrence seems to make everything lower quality
Another great model from the grandmaster himself. Thank you so much for your work <3
will there be a w4a8 turbo ?
I'll not do a w4a8 soon. The quality is sometimes abysmal on that quant.
@Darksidewalker i tried your newest int8 turbo for ref2va, it worked very well on input image anime style , but i try to put input Realistic person to create realistic style video, it auto changed the input to anime and make half anime half realistic output lol
The deleted w4a8 worked on realistic
The checkpoint works on realistic, check your prompts ;)
There is no BIAS.