⛩️💀 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:
🔥 REF2VA + FL2VA compatible
🌟 Enhanced Quality and Reasoning
🎶 Stable Audio with full reference support
🪡 Finetuned
👘 Strengthened visual consistency/understanding of various concepts
💎 Preserved details
🗝️ No BIAS
Minimax H3 has 2 Versions FL2VA and REF2VA. They are somewhow interchangeable.
You can try to use the model as FL2VA.
There are also pruned and non-pruned versions. In almost all cases pruned is same quality, but less demanding on VRAM.
My Hybrid is fully compatible to both use-cases on my testing and can replace FL2VA and REF2VA.
The 4-Step Turbo (distilled) version is tuned to preserve details as much as possible, work with reference audio, and prevent frame-burn on first-frames.
⚠️ 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
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
8 steps
FAQ
Comments (77)
Have there been any updates that would require me to re-download the Non-Distilled version?
The REF2VA quality for 4 steps is insane. How did you do this?
the turbo model？ still downloading the model, looking forward to the results!
@azsaber Yeah, this was my first test with just 4 steps euler/basic:
https://civitai.red/posts/30675099
There is some smudging, but I did it at .52 MP and 4 steps which is crazy. Bumping up the MP or steps to 6-8 will be my next test to see the generation time vs value add. I am poor and on a 3060 12g and the linked 10 second clip took around 8 minutes to generate.
Black magic!🧙♂️
@Darksidewalker CLEARLY
@FirstPrinciples Actually it is a blend of distillation, I tuned to fit from the ltx turbo loras.
Most of the time it works flawless, but depending on extra loras or latent-upscale it can overbake. But yeah, nothing is perfect. But hell yeah, 4-step high fidelity outputs!
is there a chance for a gguf models ?
INT8 CR is already AMD and NV 20/30/40/50xx compatible and much better, is there a reason for gguf?
i have only 16 gb vram
@bartoszpaprotny1997806 The model works with 12+ GB VRAM, like the normal int8 base model
"I shot this with a 0.7-megapixel setup — 5-second clip, 20 steps total, which took about 250 seconds to render. Then I tried the 4-step acceleration LoRA, and it finished in just 87 seconds. Honestly, you can barely tell the difference in visual quality between the two — the 4-step version looks just as good as the full 20-step render. The only catch? The audio from the 4-step LoRA is basically nothing but harsh static. But visually? Absolutely killer."
The released of the v1.1 turbo 4 step 768 was a gae changer for turbo lora FL2VA, 8 days ago.
If you are still not using it yet for all your T2V, I2V, FLF2V, you are missing an amazing change in term of quality and speed. and just like to waste your time :p
An personnaly i found the way to get the exact same video /audio quality for R2V at 4 steps, even at the lowest resolution 0.20 mp it's amazing quality.
But as it was a long testing phase for few days i will keep it secret for now.
@Pat3dx Thanks, I'm gonna go give it a shot right now.
@Zh_an https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main
Don't forget to use 1.3 strengh for the v1.1 turbo 4 step lora, this is very important.
Shift video/audio 6/3, and Euler/Simple.
don't work for R2V.
@Pat3dx this lora is still giving me smeary visuals at higher res and garbage metallic audio.
for the checkpoint (turbo) you need other settings than for normal, or you will have issues.
@SweetAyanna Strange, i have no issue and perfect quality, even at 0.20 mp. perfet video and audio. something must be wrong in your workflow.
@Pat3dx We have extensively tested most turbo loras in the official Comfy Org discord server. And not a single individual that I've seen has gotten any good results with that v1.1 lora. And no there is nothing wrong with my workflow lol. You're the first case that claims to get good results with this. If you want to prove me wrong however you can link a generation with that exact lora, with SFX + dialogue in it. meta data attached so I can verify.
But to actually add to this discussion. When talking 4 steps loras, they are always inferior when it comes to visual and audio (atleast in H3 so far). But the absolute best 4 step lora right now in my opinion is still KJ's: https://huggingface.co/Kijai/MiniMax-H3_comfy/blob/main/loras/minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors at 0.75 strength with er_sde/beta. And if you want the absolute best results, almost indistinguishable from base, you use this 8 step lora at 1.0 strength and 8 steps: https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors
no video/audio shift needed since those work with the base shift values in comfyui.
@SweetAyanna This is the Lora i was using before i saw the results of the enw v1.1, of course. i tested all turbo lora with all checkpoint official int8, nvfp4, etc...or fine tuned. Eros, remix, and pinkcherry.
but the v1.1 just destroy it 10x better.
check my last video posted for the Penislora this afternoon.
https://civitai.red/posts/30670172
Resolution 384x544 - 4 steps, using the new turbo 4 steps v1.1 768 lora at strengh 1.3 - 40s video generated in 387s with my 4090.
And this is just an example. As i'm making like 100+ tests everyday while i'm developping and enhancing my custom workflow/custom nodes.
giving you one of my video will not help you, draging n dropping my video will just load a full red workflow, as i never use native comfyui node, and always develop my own. So, all my workflow are fully custom, no native comfyui node in any of them.
@Pat3dx That example shows the exact issues with that lora. Smeary/Ghosty motion. Bad/Metallic audio. And 387s for that res is extremely long even for 40 second gen on a 4090.
@SweetAyanna i don't used my dual sampler with latent upscaler. Just my normal H3 workflow single pass.
With my new dual latent upscale, it's 3-4x faster of course.
And trust me i tested all lora with same seed, prompt, with all lora. The difference between the new v1.1 FL2VA turbo lora and older turbo lora, is huge.
You will never get this quality at this low res, NEVER.
My goal posting this video this afternoon, was not to proove and show you can get production ready/excellent quality with the new v1.1 lora, of course not, but to show how good is the quality compared to previous lora at this extreme low res.
and all the comunity on the Lightx2v ntocied exactly the same amazing quality this new lora provide. You may be one of the only one not getting this quality. So maybe you should make another tests with different workflow/parematers.
And what you surely not noticed, and that make this video even amazing, is that it's not an I2V, but R2V.
I used my technic to take advantadge of the new fl2va v1.1 turbo lora 4 steps, for R2V. what is sown on this video is fl2va turbo lora used for R2V video.
And trust me, with the R2V turbo lora, good luck to get this amazing quality at this extreme low res and only 4 steps.
Everyone know the FL2VA is more powerful in term of quality, motion, consitency than the R2V model. So getting this quality with R2V and previous lora, is just IMPOSSIBLE.
for a correct video quality, not production ready, consistency, you need at least 10 steps with R2V model+previous lora.
you can male all tests you want, at only 4 steps with all sampler, you will get horrible video.
大丝袜的模型，除了油腻，堪称完美。希望能去去油。
我也发现会些微改变女性的皮肤颜色变得更偏黄色和油腻。期待后续版本
is there a chance for a int4 version?
1) Acc PDD gives arror. It says that model is FL2VA
2) As Ref2VA this model bad
1) 💀
2) 💀
хз все работает идеально
@Darksidewalker This man is using a potato and blames you. Lol.
I am shocked. Why you are not asking one million dollars to download your model?
Free stuff for great art ✌️
@Darksidewalker we all appreciate it
I was wondering what happened to this model... Why did it suddenly disappear?
Looks like a licensing issue. Description suggests this was resolved.
this model without turbo lora is better and faster ,its the fastest model minimax for my workflow with ok quality good job (this model strong in image reference not in the prompt)
record time (latent upscale) 0,2 to 0,5 , 10 second =200 sec with rtx 3070 8gb, ram 16 gb.
😳
what settings are you using? oh and which version?
@mugibot72 i using the 8step turbo hybrid with latent upscaler workflow..
Your model and workflows are working very well (even though they seem quite complex to me) and are doing a great job—thank you and congratulations. I have a request: please add an on/off function to your System Monitor node. I can’t turn off the components I don’t want, and there are a lot of indicators that are completely unnecessary for me. Every time you update the node (I’m sorry to say this, but), I have to delete your System Monitor node. :)
There is a turn off inside the menu and there is a ENV you can set to completely disable it. Description is on github, as always.
is it the quality of 8 steps is better than 4 steps but lots slower? And how about the not turbo one?
I've downloaded all three but they seems are the same quality in ref2VA...? My setting 8 steps res_multistep simple shift_video12 shift_audio3. Am I confused something?
for turbo you may use euler+simple, not res_multistep.
Naturally it should be: non-turbo > 8-step > 4-step ; In quality.
@Darksidewalker Boss, one thing had to be ensured first, should I add turbo lora in those 8 steps/4 steps model? As your precious wan2.2 is lora free. And could I add turbo lora for the non-turbo version to speed up?
@biosal Turbo Model's do not need extra turbo lora's. The Non-turbo is for adding whatever you want.
@Darksidewalker Thankssssss!!!!!!!!!!!
Turbo 4 step has atrocious audio lol
Depends. But I agree sometimes the audio is not the same quality like non-turbo and 8-step. If there is any method to make 4-step distillation better, I would include that next.
use 8 to 10 steps to fix audio, its all h3 tunes
8step model is amazing, audio is great too! Did not expect it to be so much better than the original. This is brilliant work, and very much appreciated! Will you update the original (non turbo) with whatever you learned from making the turbo models?
I agree, the quality is GREAT, better than base (int8 pruned) model with 8step lora.
They all share the same enhancement, only the baked turbo and how the turbo is baked is different.
@Darksidewalker I see! it's strange then that the 8step delivers such a noticible higher output, the audio in particular took me by surprise. Once again, thanks for everything you've done, this is fantastic!
4step turbo is actually 10/10
Somehow this model makes everything darker, at last in rev2va. the previous model didnt had this problem.
Same here — all the new models show a color shift. On 10s videos the first and last frames differ quite a lot. if you extend the video to 20s, the semi‑real animation turns into a full anime style. This wasn’t present in the previous version.
+1. Despite explicit prompting of brightly lit room, the output just makes the windows brighter but the characters are very dark and lots of shadows.
I did a full 25s realistic video test with the 8turbo and no loras. There was no style shift at all.
There was no prompt given to not influence the model, that's why I posted it into my discord on the fails channel.
@Darksidewalker I testet it with a a 15 sec anime style scene, with lots of references and while vanilla h3 and your old model could match the colors, your new models made the character way darker. I testet it with the 4 step and no turbo version.
I was doing some tests with the hybrid and 8-step hybrid models, and I'm not getting the results I was looking for.
EDIT: Sorry, I hadn't read the "What NOT to expect" part. It does not have NSFW capabilities beyond the original model. My mistake. Since I use your other models, I just assumed incorrectly.
Great work! (I'll leave it at that!)
I tested all three models, and it seems that using an external Turbo_LoRA—rather than the built-in Turbo—is the best approach for now. Of course, that depends on the use case...
i do agree, after some test, using non turbo hybrid+ external turbo lora follow my prompt better.
When using the four-level model, a phenomenon occurs intermittently where there are two main subjects. It is like a doppelganger.
This is a feature! (Doopelganger powerrrrrrr!)
Damn, this works so well with 4 steps.
I have to say, you've always been the best!
Don't know what's the secret sauce, but the recipe is good !
thanks
I just want to say thank you for all your contributions. Please don’t let the words of a small number of disrespectful and ill-mannered people get to you. They don’t deserve your attention.
Thanks :)
@Darksidewalker weird tangent here, but if you go to ANY review on Consumer Reports, every single large appliance has ~1 star ratings because the annoyed and angry people are ALWAYS louder and more willing to voice their opinions than the satisfied... don't forget that for every 1 critic you have, there are probably 100 people using your work to see their fantasies come true... we appreciate you
@rkyha639 I try to remember that :) Thanks
WTB slave slime girls, in rl, dm...
@zugzug69 Poor sweet slimegirls! 💦
Hi is there a workflow instead create 30 sec long vid but also in 1 run make 3 video but seamless in 1 prompt?it Will help with low vram People like me
look up the h3 multishot node from joeygambino
Pls help fam this is a good model i'm having a blast with it but a 10s video takes like 30 mins to gen while a 5s vid only takes 200~s ToT i don't know what's wrong with my setup
Are you ever going to release again a update of your pulled version. I'm doing single frame images with it, and whatever you had going on with that thing is f'n amazing for still images.