⛩️💀 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
🔥 Hybrid (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
🧪 The Turbo (4 till 8-step) comes with latest and tuned distillation blend
💎 No style or lighting drift
💫 Optimized for better motion and faster actions
FAQ
Comments (170)
Quick tip for anyone struggling with ugly anatomy in H3: just generate your desired anatomy first in ANIMA Realistic or another model.
Make 2 samples, one front view and one side view (only part no face), use them as reference images in your workflow.
Prompt example:
'For reference, the man's bazooka resembles those shown in <Picture 1>, <Picture 2>'.
After doing this, you don't need a LoRA for specific anatomy anymore.
Going to try this and report back. Wasn't crazy about V1. But i love the workflows - thanks for the hard work
EDIT: The re-uploaded version is a significant upgrade over V1 in every respect. 🥰
I have no clue why it would do that.... my tests did not show that behavior. I'll look into that...
I have no clue, I just tested a simple woman walking prompt and nothing like that happened to me.
Use the re-upload.
@Darksidewalker Thank you DASIWA!!!! 🥰
Is better hibryd v2 or turbo v2 in terms of Quality?
turbo is always quality loss atm for h3
Comment edited. If you're experiencing blurry/quality issues, DISABLE YOUR LoRAs and test each one for compatibility with V2.
I also only do ref2va and did not encounter that issue. Can you give me a proper example what you did, so I can elaborate?
@selfishkingduga167 I came here after noticing the exact same thing.
With V1 you could see what you are about to get even at 1st or 2nd step.
But V2 you see more fog/blurriness. then the final it also doesn't output the way V1 does. its bit more jagged and stripee
@Darksidewalker Sorry, I appear to be an idiot, I forgot to test without my usual LoRAs first. Disabling all LoRA's produced excellent results, so one or more of my LoRAs must not be compatible with V2 (I'm guessing Mystic V4 but still testing each one). On the bright side, V2 seems to have much better native anatomy and motion, so I probably won't need as many, any LoRAs. I'll edit my original comment to avoid any misinformation.
Thank you for your awesome work!!
@selfishkingduga167 good call, I also had 2 lora's preloaded testing this myself now
@Megasherru It does seem to be a a bit of a tradeoff. After now testing without LoRAs like I should have to begin with, I agree the visual quality is much better with V1 (even without LoRAs), and in my experience V2 has better motion and anatomy without LoRAs.
Something I've noticed in my continued testing is that V2 seems to be VERY sensitive to LoRAs overall. Basically every LoRA added degrades visual quality of the thing that LoRA is meant for. In my case, I had a penis LoRA, which when used caused slight blurriness around the penis. When I used Mystic (overall NSFW and motion, so affects pretty much everything), it causes everything to become blurry.
Even without any additional LoRAs and only an 8 step turbo LoRA, there is some slight fuzziness, but not nearly as much. Still need to test with 0 LoRAs normal 20 step generation.
@Darksidewalker As others described as well, my results with v2 in general are a mixed bag. Looking very good in general with many improvements, but at the same time being quite pixellated in areas with fine details and movements. Especially visible in the face and hair, e.g. POV of a woman moving in the cowgirl position.
MysticXXX v4 piles on, degrading the visual quality, pixellating detailed areas with movements even more. Which is a shame, as this LoRA is quite useful in many aspects.
(Results from DASIWAv18 workflow, REF2VA using pictures only. Happens with v2turbo, as well as v2+turboloRA512 at 1., MysticXXX at .3 or .4 strength / euler/simple, 8steps, 12/3)
What's really awesome with V2turbo and your turbo-LoRA v1 is the specific improvement in audio: Excessive knocking sounds during the act and carrot cracking bj sounds are gone or very well controlled. Didn't see another turbo-LoRA yet that achieved this. And this alone is a game changer IMHO.
idk if something was actually changed or if it's a placebo, but after re-downloading the re-upload of v2, it does seem better. Still a bit fuzzy, and that fuzziness is still increased by LoRAs, but it's better.
Conveniently enough, I just finished updating my workflow to use the Upscaler + Refine node from the "ComfyUI Minimax H3 Latent Upscaler" custom node pack by xmarre, so I decided to use Dasiwa v2 w/ 8-step turbo for the initial generation and Dasiwa v2 w/ 4-step turbo LoRA for the refinement (4 steps at 0.4 starting denoise), and it's the best I've ever gotten out of MMH3.
Where did V2 go to ? i just got home and it's gone
V2 re-uploaded.
w4a8 ver plz
i used to pray for times like these 🙏 ty for all your hard work
Int4 is w4a8
Why re-upload? Was there any problems?
@HugMeIntoFace the files seems demaged somehow.
Nice. Thanks. Is there a possible to have a GGUF so it can run easily on AMD gpus on 16gb of vram? Like I dont know, Q5?
@lolmao500 Int8 CR easily runs on AMD 16GB cards. In fact much better than Q5.
It's very good that the style is not broken now. Speed is a crucial factor. You've conquered slow-mo, but too much. In V2, this was completely broken. Speed tags are needed, like: Slow, gradually accelerating, normal, rhythmic. Currently, only supersonic, machine-like ultra-fucking with a jackhammer is available. Some people might need it, but it should be optional, not mandatory 100% of the time. I think this is the biggest problem with V2
@Kitsune_Yukino I did not tune anything in that way. So this must be a side effect of some motion enhancements or maybe the turbo?
But that said, precise prompting was always needed for H3.
@Darksidewalker Just pay attention to the generations that people do in V2 Most videos look like they are sped up 1.5x2x especially in H-scenes.
I saw this here and I saw it myself, and I tested it with all lore completely disabled.
I only tested FL2VA
I know MiniMax requires more detailed prompting. But this is precisely where the model stops responding.
Of course, it's my little job to warn you about this.
@Darksidewalker
can you suggest for Hybrid V2 what settings i need to use
thx you, really!!! you are the best
May i ask if we're getting a 4 step ver of v2?
Pretty sure the Turbo V2 is both 4 and 8 step
@blackpawnx should be both
Holy shit, the v2 is so fucking awesome! I got the version with built-in turbo lora and was blown away by the quality. Currently the best custom model on Civitai. Thanks for the model, bro.
I’ve noticed that with low step counts, noticeable noise still appears when objects or characters move; it only improves at 10 steps or more. I am using the Turbo version.
V1 turbo 4 performs better than v2.
On what?
@Darksidewalker Yep I noticed this while was doing generation in v1 and downloaded v2 then did generation on same settings, v1 follows prompt better and generating details better.
@velanteg I would argue it is the other way around. So more detail what you are doing would be helpful.
完全错误的评价
@Darksidewalker I am using 4+1 generation. I didnt tests of full 8 steps.
What a am doing for test you can see in gallery.
@Darksidewalker Just look in gallery, I compared 8 steps generation of v1 4Turbo and v2 Turbo. Everything same, just v1 vs v2.
😸happy happy happy
I was very confused by the first upload of the non-distilled V2 because it's outputs were like a Q1 quant of a 2-step turbo model. Definitely wasn't up to the DaSiWa standard.
With the reupload I've done a little A/B testing against V1 with I2VA and Ref2VA, non-distilled, and I'd definitely say V2 is much better for both SFW and NSFW. Where V1 had moments that were reminiscent of WAN meltiness, character/style/quality consistency felt spot on for V2. Not to mention the audio is improved. I find any weaknesses in NSFW stuff that I can't prompt around are usually bridged by adding MysticXXX at only 0.2-0.3 strength (vs V1 running 0.5-1.0). Not in every case, but in general LoRA strengths seem to need to be dialed down compared to what I was using on V1.
Really glad V2 got sorted out. Thank you for blessing us with another gift.
I wanted to ask, which one would be better: the Dasiwa Hybrid V2 or the Dasiwa Hybrid Turbo V2?
Also, does the non-Turbo version still need to use 20 steps, like the standard version?
Sorry, I’m still a beginner, so there’s a lot about this that I don’t fully understand yet.
Non turbo = full steps 20-30 needed, turbo = 4-8 steps; Turbo always sacrifices quality for speed-up.
@Darksidewalker Alright, thank you very much, Dark
and i thought v1 was a game-changer. holy shit. this is just incredible. gonna be looking back at this comment in a year and laughing at myself i'll bet
It is like everything, a while ago WAN22 was the hot take, ... here we are iterating on H3... till the next big thing.
@Darksidewalker #Ling 100
v2进步很大，提示词遵守度更准确。感谢分享。
Hello,
what's the difference between "DaSiWa Hybrid v2" and "DaSiWa Hybrid Turbo v2"? I thought that the Turbo LoRA was baked in, but in "about this version" both have the "🧪 The Turbo (4 till 8-step) comes with latest and tuned distillation blend" option.
It is not that deep, just a copy & paste thing. Turbo is turbo and the other in non-turbo.
@Darksidewalker I see. thanks
An amazing version! Does the "turbo" version not require adding any additional turbo LoRA? Is the BF16 version publicly available, or is there another way to obtain it?An amazing version! Does the "turbo" version not require adding any additional turbo LoRA? Is the BF16 version publicly available, or is there another way to obtain it?
hahaha thank ya! shall I double answer it? 😸
No bf16 available. Yes, the turbo version is turbo baked in.
Did some tests with same prompt and same video duration 6s and 16s. Hybrid 4Turbo V1 is still better in absolutely all aspects vs Hybrid V2.
If you compare 4-steps, might be true for some visuals, but sound actually not.
V2 was my take to try a multistep lora for 4-8 steps, while v1 4turbo uses a dedicated 4-step turbo.
I would not recommend 4-step usage at all, regardless of the version, tbh.
4-step always sacrifices alot. It is more or less just for prototyping/preview.
But you can add a dedicated 4-step lora to the non-turbo and you will get better results, I guess.
@Darksidewalker i don't need sound at all. And i have no idea to turn it off. Mute doesn't do it's work in your node where video combines.
@Gordet you cannot generate without audio with MiniMax H3, mute just mutes the video preview.
@Darksidewalker really? but i can. Bypassed some nodes in your WF (Load VAE Audio,VAE Decode AudioIf/Else Switch) and i'm getting clean muted vids. Check it yourself if you don't beleive me.
@Gordet okay, I did not know that this would not throw an error. Normal a model needs VAE loaded, interesting 🤔 I'll try that
Does the v2-turbo = v2 + your h3 turbo lora 2929860?
V1 and your workflows works great. Love your amazing work.
This my first time I use your MiniMaxH3 I2VA workflow together with new turbo v2 model & truly impressive the results. Love it.
When processing anime characters, v2 seems to automatically add lips, and the eye stylization is also very pronounced, whereas v1 didn't have this issue.
I doubt that,... Did you redownload the upload version?
@Darksidewalker Maybe it's a coincidence? I'll try it a few more times later.
@Darksidewalker It really is a hit-or-miss issue. I ran another test comparing v1 and v2 using three identical sets of seeds. v1 consistently stays true to the face in my reference image, while v2 added an extra lower lip line in one of the runs—something that never happened with v1. Still, it's really no big deal. Overall, the model performs amazingly well, so thank you for your hard work and for sharing! I've already uploaded the comparison video using the same seed sets to the gallery
Anyone else getting this with V2? Whenever there are breasts in the prompt, they merge into the clothes — even when the character is fully clothed, it looks like there's no fabric at all, just a water balloon shaking. Is this a known issue, or is it LoRA/settings related?
How is it that the first version (v1) is better than v2? Am I mistaken?
v1 is better
Since you both posted literally nothing it's hard to believe. What in details is better or worse?
@Darksidewalker They are both talking complete nonsense.
@Darksidewalker + Animations have improved
- Video quality degraded (looks like it lacks inference steps); at 0.2–0.4 MP it's quite likely the model just gets pushed into higher resolutions where the quality scales up
- Model LoRAs stopped working (judging by the single one that was connected)
However, the sample size is very small (around 10 clips tested). Tested on the base ref2vid workflow using the Turbo model at 4 steps, across different steps with euler / simple .For low resolutions, v1 is better.
@Suri14d something must be off with your setup, maybe your download got corrupted. All loras are working and I did 0.3MP tests running fine and visually stable.
Also 4-step is not a good comparison to start with.
The new multistep is not tuned specifically for 4-steps.
I tried "DaSiWa Hybrid v2" and the quality is very good. But it just didn't follows the prompt at all.
When I use it for T2V I tried this Prompt:
---
integrated_multimodal_description: [Shot 1] A five-second cinematic single take inside a rain-lashed glass observatory at midnight. A radio astronomer presses an antique transmitter and says clearly (S1) <d>[English] "If you can hear me, follow this signal."</d> One brilliant blue pulse races across a translucent star map as the camera slowly pushes closer.
overall_soundscape: Rain on glass, a low electrical hum, three receiver tones, the synchronized voice, and a short stereo transmission pulse.
---
Out came a video with two arabic dudes in a modern living room with two arm chairs and a drum set, talking in a language I don't know.
Another seed gave me one arabic looking dude talking in something... I would say was danish.
What the heck?
Even other FL2VA models can follow that prompt in pure T2V.
Through, I tested the prompt and it exactly did render that scene. Whatever you do, is a fault on your side, not from the checkpoint.
What the heck...
@Darksidewalker Ok... that's weird. I did nothing differend as with the other checkpoints. Strange. Same settings everything.
@NRubric I did only use basic settings, 8-steps, the prompt, random seed, native resolution, no loras, t2va mode.
@Darksidewalker Same here. Nothing fancy. Beside that I used 20 steps.
Every single generation, in temp folder, over 10gb sized file is being created, and it doesn't delete itself, and I assume it should. After few generations I'm out of space. I have to delete it manually. Anyone having this problem too?
same
If you're using the RTX SR 2x upscale function in the DaSiWa workflow you could try this:
Open the backend, locate the DaSiWa RTX upscaler and refiner node.
Make sure, empty_cache is set to true and use_mmap is set to false.
I don't know why use_mmap is set to true as the standard. Maybe I dont't get right, but the descriptions sounds to me like it should be turned off by default?
However: activating empty_cache and deactivating use_mmap stops windows from filling up my pagefile. Which is growing in tens of GB with every run otherwise, even if everything fits in VRAM/DRAM. But this specific variation seems to by my very own special bug.
Try it, if it doesn't help, switch them back. :)
To free up space and keep everthing in VRAM/DRAM for upscaling & interpolation, I also added a "purge_vram"-node connected to the VAE Decode node output, purging cache and models. Use at your own risk - or pleasure. ;)
I've made a lot of AI images, but never videos. Could someone please tell me: I'll download this model, for example, but where do I get VAE, Text Encoder, and Clip Vision for it? I've found and downloaded some, but they don't work. I don't even know how to search for all this stuff and distinguish what's needed for which model... If anyone has any advice, thanks. I'd like to try creating a video from my images
You could dl my workflow (suggestion section), there are links in the description notes for everything.
This is probably the best NSFW fine-tune out there right now.
Hey now this retains style and likeness just like your Wan merges. Which could not be said about V1. Looks like I won't be shopping around any longer. ...and there I was, starting to get concerned that we might need to wait far longer for anything decent.
Keep up the good work!
P.S. FFS finally no more guessing whether wan will follow the prompt or not.
New Workflow, New Model. You're like Santa bro. Never stop. Lol.
Really love your model! I was wondering if there is an unpruned INT8 version available?
No, not available, will not be available.
how to use? please provide installation guide for newbies.
1. DL a workflow
2. DL Model files
3. Install all dependencies and custom nodes
4. Read the notes
I'm new here , could you please answer my question? , if I use the Turbo version I don't need to use the turbo LORA and just change the steps to 8? and if I want to use the non turbo version I have to change the steps to 20-25 and still no turbo LORA , right ?
yes，But the non-turbo version can also use Lora for acceleration, though I think the turbo version has a lot fewer hassles.
After using it for another two days and comparing it with other fine-tuned models, DASIWA remains the most powerful; the fluidity of its audio and motion is unmatched. However, I noticed that the "turbo" version performs better than the non-turbo one—though I’m not sure if that’s due to a settings issue. In any case, I’m grateful to the author for providing this model, and I look forward to seeing even better, larger new versions in the future.
Thanks for the kind feedback
Somehow Skin is very soft/flexible when touched, more like a dough. Did anyone notice this? Does anyone know how to work around this?
Better prompting, better references.
I'm using your C-MMH3 workflow from github. The notes say I should use: minimax_h3_fl2va_pruned_int8_convrot.safetensors
minimax_h3_ref2va_pruned_int8_convrot.safetensors
But what to use if I want to use this model? Set both to DasiwaMinimaxH3_dasiwaHybridV2_int8?
Oh really, I forgot to replace the links. Yeah, set what madel you want.
@Darksidewalker Do you have any recommendation for an upscaling model? I get a result, but it doesn't look so clean like your examples. Or is this a settings issues?
I am trying to use this model together with your workflow. Trying to add any nsfw lora, like the breast physics, or mysticxxx, results in the target video looking as if I'm looking at it through a cross-reeded glass. What does that mean? You mentioned "No tuning for ToS-violating concepts", does that mean the model is actively sabotaging nsfw concepts when a lora tries to add them?
😂 lol no ... The model is nor preventing any Lora from working, must be something with your settings or pc
A big thanks to Dasiwa"
in the turbo v2 version, i have found bit color change and very very mild blurry nature.
so for testing, i have added the turbo lora "minimax_h3_turbo_v4_step600_ema" non-pruned at 0.5 strength with enabled interpolation in dasiwa workflow and tested it ,
u wont believe, outcome is fantastic with good audio (with clear voice and background sound) also with fluid video motion ,
the best part is the rendered video is in excellent quality even at 576p in detailed without color change and brightness.
i run it in rtx 4060ti 16gb card + 32gb ram.
for 20 seconds video it took 12 mins (don't know the iteration is fast or not, may some will get in less than this )
for me it works perfectly...just want to share , try it and reply whether this setup is good or not.
Testing V2. I have a persistent issue where in the anime style the model constantly mistakes male characters for female ones—assigning them female voices and drawing female breasts ignoring the prompt and ref images with flat chest. Drawing femboys is practically unreal - it turns them into breasted girls all the time.
What are you saying, that's literally impossible 😂
I'm finding the V2 turbo model is giving me very pixelated results even at 12 steps, swapped back to V1 (8 step) and it works great at 8 steps. Using your latest workflow, though only working with an RTX 2080 8GB.
I Got contrast and color shift in the middle to end . :(
Hybrid V2 Turbo works fine for Anime in Comfy. Just load the base model first, then apply LoRA. This model understands prompts deeply, even if you leave out the soundscape details, it figures it out on its own. Don't overthink and stack soundscape descriptions, it will only cause overlapping audio.
i am using your recent workflow and model hybrid v2 but whenever i put lora it starts making pixcel like and hair is really foggy. I cannot use lora with this model?
i changed to v1 and it works perfectly
I found some fix, there are some loras will make this model completely broken such as mystic xxx and nsfw anime lora in my case. this fixed the video detail issue, being foggy.
then I changed latent upscaler steps to 4 and denoise to 0.05. then it fixed the issue first and last frame being pixelated.
hope it was helpfull.
Using the v2 turbo model generates very pixely videos but switching back to v1 works just fine
v1 works better to me. The v1 4 step have better motion and lora compatibility while v2 can collapse when loras are added. I also notice with v2 motion on 720p is a lot more stiff than 480p. The advantage compared to v1 is that it maintains art style better. Looking forward to eventual version that have the best of both!
I use V2 in WanGP and it is fantastic, much better than V1. I have been using your wan2.2 finetunes for a long time. Always high quality work. Thanks Dasiwa! Btw, I found that the combination of non turbo V2 with Taomate 3-step lora worked surprisingly well. You may consider to incorporate this accelaration lora into the turbo version in the furture.
Thank ya! I got reports that the 3-step taomate lora produces bad content... I did not try it myself for now. So I wonder. I'll have a try.
No se cómo hacer para poder usarlo en wan2gp me podrías indicar cómo ? O pasarme el archivo correcto ???
@Darksidewalker I couldn't tell any visual loss using non-turbo v2 + 3-step taomate vs your v2 turbo version in my limited tests, nor any difference between 3-step taomate lora vs 8-step PDD. Only in rare cases, the 3-step lora failed to produce decent video. Give it a try.
I haven't been able to get a good result with V2 yet, yeah the burn in from V1 is gone, but the quality is pixelated and the motions are usually too fast with or without turbo LoRa.
I haven't got a clue what I'm doing wrong here.
Turbo version, normal version, low steps high steps doesn't seem to matter.
In V1 normal version I was able to get realistic motions and high quality on even as low as 8 steps (and I never had to use LoRa's).
Yeah I can't get any good results with V2. Only thing that gets anything close to ok is when I use zero LoRAs. Turbo version seems to not be working at all with only 8 steps. Doing side-by-side comparisons to V1 with all other settings that same V1 is always better.
any good wf to use this chkpnts?
I did one, see suggested resources
@Darksidewalker WHERE DO GET THE DASIWA NODES?
@gigcamellos2023710 github or comfy-manager or my comfyui-installer
@Darksidewalker @Darksidewalker you mind sharing the links ? please
What are the settings for the turbo model?
4 or 8 steps which sampler and sheduler?
See page description
@Darksidewalker oh sorry i was blind!
I see v2 turbo have int8 merged int4 (aka w4a8)
But v2 normal just have a raw int4
W4a8 vs int4 which is better?
All int4 should be w4a8, why do you think it's normal Int4?
Again requesting a bf16 version to download. My use case: I'd like merge the 3-step turbo lora into the bf16, then convert the model to int8 myself (which should save a good chunk of VRAM/speed over running the LoRA).
But this is just one example. BF16 is just more efficient to experiment with, as H3's rather large loras can be merged in pre-quantization.
Declined, I have reasons not to. Sorry.
You can make this with the base H3. 👍
That's fine, I still appreciate the merge!
Confused on formats:
What is the difference between the INT8 and the TURBO. Strange my output is better and adheres more when I use: DasiwaMinimaxH3_dasiwaHybrid8turboV1.safetensors that I downloaded sometime ago. Is this INT8? I used 8 steps so I assume it is turbo.
Turbo models are with turbo baked in, the others without
Ok thank you for the clarifications. The original DasiwaMinimaxH3_dasiwaHybrid8turboV1.safetensors that I downloaded says turbo but no indication if it is INT8 as compared to the available downlaods. Is this 20GB model INT8 as well?
I also wish to reprt, on the Dasiwa Minimax H3 workflow, the version 1.8 works great even at 20 seconds. But the latest version 1.9 gives me OOM passed 15 seconds...
The 8 step setting is very good, but at 4 steps, the image and motion become blurry.
I noticed that too, so I've left it on 8 steps. (^_~)
Having issues finding this model in the resource allocator when I upload a video. It's just not appearing...so I can't tag it on my videos. Anyone else have the same issues?
I was struggling with V2 in Comfy, but I think I've figured it out.
My suggestion is to drag and drop the author's images made specifically with the V2 updates and build a brand new workflow. Don't just plug V2 into the model fields like I did. There are some significant changes it seems in the workflow/settings. test w/o loras, then add them in once you have some good gens.
Also, update ComfyUI.
V2 is clearly better at adherence and does some crazy good stuff.
But jeez, its quality is not that of V1. (on his own workflow)
Wish DarksideWalker could update his workflow to accommodate V2 and it not look like ass.
His model looks great on other workflows but not his own.
- So yes, I agree with you. new workflow is a must for V2
@Megasherru The workflow is fully compatible with v2 and works with high quality. If not, its the settings you choose.
@Darksidewalker Thanks. I was using the Turbo Checkpoint alone but noticed that I did not get the quality as I did with V1.
BUT - I just downloaded your full V2 Checkpoint along with your Turbo lora Alone.
And Holy WOW, that fixed the issue I have been getting.
Maybe it was my settings yeah. either way. Im super impressed! Big Step up!
@Megasherru Are you running the Turbo Lora at the same time as the Turbo v2 Model - or are you running the Turbo Lora with the None-V2 ??
@FollowTheWhiteRabbit Im now running the full DaSiWa Hybrid V2 Int8 Model not the Turbo. Then with the Turbo Model loaded in the Lora Loader using 8 steps.
It works great for me. The model takes a bit to load, after every 1 generation I unload models and free model and node cache. But works great for me now.
@Megasherru Thanks for the response, would be good to know if the full checkpoint v2 + 8 step lora is better than the turbo checkpoint v2 :P so many combinations and not sure which is best
@FollowTheWhiteRabbit best is what works best for your use cases
@Darksidewalker need to run some tests, V1 solo, V1+TurboLora, V2 solo, V2+TurboLora, V2Turbo solo :)
7500f, 32gb 6000mhz, 5060ti 16gb. This w4a8 turbo model, Qwen 3 VL 32B NVFP4, Video VAE int8_convrot, Audio VAE fp32, 8 step, 5 sec, 1440p (720p+RTX Upscaler/Refiner) = Prompt executed in 253.34 seconds. I’m speechless. Pure art. Perfect video, perfect movement, perfect voice.😭
Yaaaay! 😸🫶
can yo share your workflow :)
It's certainly very impressive, although I've yet to get to grips with all that it can do. For sure, the best I've managed thus far is 1536x1152 (double that RCAN upscaled), but only 5 seconds. Any more than that and even my overclocked 5090 bombs out and saves garbaged data. (^_^;)
@enjidifussion20820 (GOAT) DaSiWa workflow! Look at the suggestion resources for this model. There are already a lot of features: ComfyKitchen (Sage Attention without a headache), H3 Cache, Chunking and excellent work has been done on memory optimization.
@sherriterri320 I recommend trying the workflow I described above.😌 And yes, this model handles the anime style! Previously, only Wan 2.2 was on par, but it only generated 5 sec video and couldn't produce audio. There were some workarounds, but they significantly increased the time required to achieve a good result. Anyway... Good luck mastering it!
@Asuu_Maruu Yeah, I've giving it a go right now in fact -- have eliminated all but one error, which I'm going to ask about in another post. (^_^)
Where should I put this FFmpeg?
Can be anywhere so long as Windows/Linux knows where it is via a path statement. (^_^)
Liking this one a lot so far since it seems a lot more tolerant of anime/toon reference imagery (it doesn't try to superimpose realistic body parts where they don't belong basically). It's fast as well, on par with PixAI's V4.0L, and about the same quality overall, which is great since it's no longer costing me 100,000+ credits a pop. (#^_^#)
Still a novice with Comfy, however, and I'm sure I have much to learn, not least how to manage speech better (seem to be getting a lot of gibberish from the characters currently, even when I specify a line in a specific language). Also not sure what the NSFW limits of the model are yet, but I've already run into scenarios that the AI seems to have little clue about. (^w^)
Trying to run the workflow embedded in the first video, gives me an error message. Error
No output node found for id [1512] slot [4] MODEL
All requirements have been installed at their latest version, all models downloaded and placed on the directed paths. I can't see any red node or other error message apart from that one .
Comfyui frontend issue, make sure to update to 1.53.6 and dl the latest workflow and custom node
@Darksidewalker ComfyUI: 0.36.0 ComfyUI_frontend: v1.53.6, I installed the nodes initially from ComfyUI (latest version) and I also tried the version straight from github (nightly, plus requirements). rgthree: v1.0.2608312342, KJNodes: v1.5.2 Comfyui-MMH3-UltimateUpscale: v0.0.6, ComfyUI-GGUF: v1.1.10. The model itself works with the standard minimax H3 workflow with no issues, it is the specific template and other similar ones that follow the same structure that gives me the same error message. I'm not sure what I'm missing..
Okay, I'm taking your workflow for a spin now, Dark, which makes me feel like I'm looking at the wiring diagram for our battery charger from my apprenticeship days. Heh, seriously though, I've eliminated all the errors save one: the DaSiWa_SeedControl. I've installed the DaSiWa Custom Nodes Collection and it has no conflicts, plus I copied over the nodes_seed_control.py from Github, but the error persists. I noticed I can simply bypass it to remove the error, except I have no idea what that will do. (^u^)
Probably missing something obvious, so please forgive my lack of experience since I've only been at this for a few days. Definitely want to get this working as your model is easily the most stable for anime & cartoon imagery (blowing my mind some of the stuff that's coming out, even with the basic Comfy workflow, now I understand better how the prompting works). At any rate, if you could point me in the right direction, I'd be more than grateful. (=^_^=)
ETA: No worries! Used Copilot to narrow it down to the repo not being installed properly. Not sure why, but when you clone it directly from Github, it ends up twice the size with a lot more files than when you use the Nodes Manager in Comfy. (^_^;)
This means you are not using the latest, probably using comfy manager and it installed you an old version of the nodepack