🎁 Fan Registration Bonus
Register to receive 1000 RH coins:
🔗 https://www.runninghub.ai/?inviteCode=rh-v1325
⚙️ Workflow
Ultimate Simplified MiniMax H3 Full-Feature Auto Prompt Workflow (Lenient moderation, NSFW content allowed)
🔗 Experience link: https://www.runninghub.ai/post/2088114005419282433/?inviteCode=rh-v1325
🔗 Civitai free download link: https://civarchive.com/models/2882357/
📦 Models
🔗 https://civarchive.com/models/2879272/minimax-h3-remix
⚡ Turbo LoRA
🔗 https://civarchive.com/api/download/models/3257181?fileId=3140963
🎬 Clip (uncensored)
🔗 https://civarchive.com/api/download/models/3254025?fileId=3137729
🛠️ Required Nodes
- Easy-H3: https://github.com/FX-FeiHou/ComfyUI-FeiHou-Easy-H3
- FeiHou-Toolbox: https://github.com/FX-FeiHou/ComfyUI-FeiHou-Toolbox

🌟 MiniMax-H3 Remix
Dual‑H3 Hierarchical‑Recombination Merge Dual‑H3 Hierarchical‑Recombination Merge is a merged checkpoint built with a custom hierarchical recombination technical pipeline based on two H3 base models. It adopts targeted layer‑rearrangement instead of naive weight averaging, preserving the solid stability of the base model while importing artistic bias, dynamic expression and fine‑grained detail features from the overlay model. LoRA weights are statically fused internally, requiring no external LoRA file during inference.
Two precision variants are released for different hardware conditions: full‑precision BF16 for quality‑oriented workflows, and ComfyUI‑native INT8 ConvRot for low‑VRAM deployment scenarios.
Recommended to use alongside the FeiHou Easy H3 custom nodes and workflow: https://github.com/FX‑FeiHou/ComfyUI‑FeiHou‑Easy‑H3
⚠️ Please follow local laws and exercise responsible usage when generating content.
📦 Model Variants
🔹 BF16 Variant (Full Precision)
Optimized for complete numerical precision and maximum detail retention. Best for high‑quality local generation, rendering and art creation.
Installation: Place the safetensors file under your
ComfyUI\models\diffusion_modelsfolderUsage: Load via FeiHou Easy H3 workflow
Recommend VRAM: ≥24 GB
🔹 INT8 ConvRot Variant (ComfyUI‑Native)
8‑bit quantized release with native ConvRot implementation for ComfyUI. Prioritizes VRAM reduction and practical deployment convenience.
Installation: Place the safetensors file under your
ComfyUI\models\diffusion_modelsfolderUsage: Load via FeiHou Easy H3 workflow; no extra quantization plugin required
Recommend VRAM: ≥14 GB
⚠️ Except for numerical precision and memory footprint, both variants share identical layer structure, merge logic and inherent model characteristics. ⚠️ LoRA effects are statically baked into both checkpoints; external LoRA files are not needed.
🎉 Merge Technical Notes
This checkpoint is produced through dual‑H3 hierarchical recombination workflow:
One H3 model acts as the stable base foundation
The second H3 model performs weight overlay within designated response blocks
LoRA effects are statically fused directly into checkpoint weights
Targeted re‑orchestration is applied to key modulation layers including AdaLN, instead of simple weight averaging
Retains base‑model stability while inheriting overlay‑model visual tendency, dynamic performance and fine details
✨ Core Capabilities
Stable Base Generation Reliable baseline output inherited from the primary H3 base model, avoiding common merge‑induced artifacts and model collapse.
Enhanced Visual & Dynamic Features Absorbs painting tendency, dynamic expressiveness and subtle detail features brought by the overlay H3 model via block‑level targeted recombination.
Zero Extra LoRA Dependency All LoRA contributions are statically fused into checkpoint weights. No additional LoRA needs to be loaded during inference.
Flexible Hardware Adaptation Choose BF16 for supreme quality or INT8 ConvRot to run on mid‑range GPUs with limited video memory.
Optimized for FeiHou Easy H3 Designed for the FeiHou Easy H3 node suite and workflow for best generation performance.
⚙️ Recommended Settings
Sampler:
res_multistepScheduler:
simpleSteps: 8‑12
Turbo LoRA weight:
0.75
This combination balances generation stability, fine‑detail preservation and visual consistency, suitable for most creative workflows.
🎯 Use Cases
General‑purpose text‑to‑image generation
Dynamic‑oriented artwork and illustration creation
Complex‑detail‑heavy scene rendering
ComfyUI local deployment on high‑end or mid‑range consumer GPUs
Workflows expecting both stable baseline and enhanced stylistic features
Iterative creation, concept art, character illustration
📌 Known Tips
Do not load additional matching LoRA together with this checkpoint, since LoRA effects are already statically fused. Applying extra LoRA may cause over‑saturation or distorted outputs.
INT8 ConvRot variant has minor precision loss compared to BF16. If you observe obvious detail degradation on key creative projects, switch over to the BF16 version.
This is a Checkpoint Merge, not a native trained H3 checkpoint. Some edge‑case rare concepts may behave differently from original base H3 models.
Description
Update Notice for MiniMax‑H3‑Remix (Version ID: 3254025) The int8 quantized version bug has been fixed. If you have downloaded the int8 file before, please delete your old file and re‑download from this page. Do not keep using the old broken file.
FAQ
Comments (63)
github link broken for me, found the working one:
https://github.com/FX-FeiHou/ComfyUI-FeiHou-Easy-H3
thanks op will test out!
question: does it really matter uncensored text encoder?
I download this model and text encode, using the FeiHou nodes but what version of turbo lora i need use? all my videos become like a pink filter and terrible image
yeah mine is completely distorted.
same
@BigSad11 same
same
+1
有bf16或int8的嗎
现在已经上传了int8,bf16还在传
太棒了,非常感謝
remix原来是b站的肥猴大佬的作品啊,果然是大神
que lora de 4 pasos podria utilizar en este modelo?
感谢!会有pruned_int8_convrot的版本么?
pruned版效果很差,不建议使用
@FX_FeiHou 显存内存低的人不在少数,效果也要先能跑起来再说。请大佬也考虑下我们这些低配人群吧。而且剪枝版也没差到不能用的地步。
很棒!使用上很直覺的節點,模型的遵循度也不錯!
Nice
那个2B的提示词能分享一下吗?
great job ! like everytime ! checkpoint have sex action ? (anal, double bj, deepthroat...) ?
从wan时代就开始用了, 可怜的12G显卡, 有 w4a8 或者 nvfp4 就好了, gguf 也可以, 但是好像慢些
which shift scale do you recommend with your turbo lora? 6 or 12?
EDIT: actually shift scale 6 works well. steps 12 is nice
Suitable for both ref2va and fl2va?
You need to put this model on a mirror like HF because this sites server to download from is slower than dialup these days.
I have a problem, it always sends this error:
sageattention is not new enough version or could not determine CUDA architecture, cannot apply MiniMax H3 Memory Efficient Sage Attention Patch.
but I have no idea why
30G,有点大,看来16G 显卡跑不动
你有64G内存就可以
确实跑不动,16g显存+32g内存没法搞
能否做一个gguf q6的版本呀
Great! No need of Lora for NSFW. But, if I want to use, I can't.
Usage: Load via FeiHou Easy H3 workflow ?? does not work with normal workflow ? just replacing the model in comfyui template workflow ?
模型的保真度和皮肤材质很强,作为图生视频角色一致性保持也相当不错,但作为武打动作类模型是真的不行,高动态动作几乎都会崩坏。模型似乎无法识别对打的情况,角色非常倾向躺在地上。
How does this model even run? It’s 61 GB! Much larger than consumer GPU. I have a 4090 but this is way bigger to run than 24 GB.
Many people seem to think they can't run it because it's too big. H3 has a special architecture, it doesn't all get loaded into your VRAM. It mostly gets loaded into your RAM and then your VRAM only takes chunks of it. I am using bf16 H3 models and encoders on my 12 gb vram laptop 4080 and it works great (my mind was blown when i decided to try it for the hell of it). You need like 96-128gb ram to make it work using bf16 base model plus bf16 encoder. Otherwise you can try to use int8 convrot encoder with bf16 model. Rtx 40 series has hardware bf16 acceleration so it runs surprisingly faster than things like fp8 or q4 quants at least in my case. All this assuming you set up everything properly.
Did you rent a data center to train this model? I generated a couple videos with my dgx spark (took days for each) and I can safely say even in sfw videos it is much better than the base model
Just as a feedback ... there are enough people who can or know how to run this. This seem to need an update though - i hope you got something in the pipes ;-)
gguf q6 ..please
Many people seem to think they can't run it because it's too big, even the uploader of this model thinks you need more than or equal to 24gb for the FB16 model. H3 has a special architecture, it doesn't all get loaded into your VRAM. It mostly gets loaded into your RAM and then your VRAM only takes chunks of it, the parts that it needs. I am using bf16 H3 models and encoders on my 12 gb vram laptop 4080 and it works great (my mind was blown when i decided to try it for the hell of it). You need like 96-128gb ram to make it work using bf16 base model plus bf16 encoder. Otherwise you can try to use int8 convrot encoder with bf16 model. Rtx 40 series has hardware bf16 acceleration so it runs surprisingly faster than things like fp8 or q4 quants at least in my case. All this assuming you set up everything properly. If you want a workflow that works for this, search for multishot seamless on civitai or something.
Hi! Does your laptop have more than 96GB of RAM? I have a 4090, but only 32GB of RAM, I had to make sure to buy more before the price hike 😅
@Eliz99 Yup I have 96 GB RAM which is just enough to run the BF16 H3 base model plus BF16 text encoder. My ram is at times almost completely full, in the 90-95 gb. But it works well. How much you need depends on your situation. 96 gb should be enough but if i had to buy new now, I'd probably go with 128 gb (if i used a desktop tower, id go with 256 gb at pre-price hike prices). With 32 gb RAM I doubt it'll work as well as with me. I got my RAM just before the insane price hike.
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
顯存不足,又要一次性生成長片,屬於左右腦互搏了,我可以跟你說,不可能
There are just now, released youtube videos, on how to seamlessly create up to 1 hour videos with minimax h3.
@jiasai001432 兄弟,你理解错了 😄 我不是想一次性生成一部长片。
我的意思是一段一段生成,但是每一段都能无缝衔接起来。这个问题其实已经有解决方案了,我现在已经找到方法了。
不过还是谢谢你的解释!我明白你说的情况,如果是一次性生成长片,那确实不现实。只是我想解决的是分段生成 + 保持连续性这个问题。
@juxflux Bro, thanks for the info! There’s definitely a lot of interesting content to learn from today, especially about the MiniMax workflow. 👍
Is turbo lora required?
I will answer myself: yes it is. Without it video becomes a mess. Also don't add any other loras or it results in a mess too. I like how clean model generates faces and body, but it needs to be compable with loras otherwise it's very limited.
Can you make a LOW ram version?
5080居然可以跑bf16的,占用的是内存,我96G内存跑通了ref10秒720P的。
我也是5080+96G内存,刚看到模型60多G大准备放弃,看到你这我又打算试试了。
good....
👍Fantastic work.
high-quality output with lora-free workflow, no need to pick or tweak loras anymore.
Magnificent work, my friend. Of the available options, I consider only two worthy of applause, and your model is in the top three.
Thank you for sharing these bf16 and int8 models; your contribution to the community is invaluable
👍💖👍
And what is the other one from top 2? Eros max or the one from dasiwa?
@TekeshiX Exactly,
1 - MiniMax H3 Remix (yours, because it yields the best results with cosplay and fantasy concepts, in bf16).
2 - Dasiwa (best style for 3D, 2.5D, 2D, anime, etc. / Older INT8 31GB version; currently available versions are inferior int8-19GB ones).
3 - Eros (realism and fantasy semi-realism, also in bf16).
Basically, each one holds the top spot where its particular strength shines; they are all fantastic, each with its own unique vision and approach.
Once again, thank you for your fantastic content. Keep it up, your model is perfect and offers something that others can't.💖😇💖
For some reason, the model name doesn't specify whether it's the reference model or the FLF version—it's completely unclear, or is it a two-in-one?
未删减版视频链接是错的,里面不是视频呀
I don't mean to be a doom sayer. But this was an absolute nightmare to get running. Code loader changes, all kinds of changes to file names, nothing matched. The recommended downloads needed file name changes and extra files to download. I have a 4090 gpu and it's still taking 5-10. minutes to load a 3 second i2v generation, 30 min for a 5 secs. And still throwing errors. Even then, the prompt adherence was not even close. Maybe it's just me? But this was NOT an easy set up at all. Beyond frustrating. That being said. I appreciate and commend you for all the work that goes into making something like this. but... if you're going to put this out there for people to use, for the love of all things holy, please be more descriptive on what is needed, and the extras to make it work.
try this https://youtu.be/ISL_J7E7EPU
Very complicated to get working correctly - but when it does it is an incredible model.
