beta 4:
rebuilt on beta3 config with some loras changed out for newer versions. Turbo was consensus merged to construct a ref/t2va hybrid turbo lora. That's the key element to the merge, there is no non-turbo version. The non-turbo version is bad. That custom turbo merge will need further improvement. This one preforms video, reference, and motion well in 6-8 steps without the issues from beta3, but audio needs shift configuration.
Use sampling like Euler/simple 8 steps with sampling shift - 12 video/ 7+ audio. LCM/simple or beta with 6-8 steps and no shift can also be better for audio and drawn styles.
Audio is lackluster and it's becoming somewhat apparent that H3's integrated audio is not good, like terrible actually. Any future multimodal models should avoid integrated audio if they intend to open source. I have almost no control over how the audio works inside the model. Don't post about it. I focused on motion and prompt response and of course when I get those working well, the audio ends up bad, go figure. I'll look at what kind of different turbo configs can enable more audio crispness to come back or likely will have to wait for Sulphur to replace MysticXXX which is contributing to the audio quality drop.
beta 3:
Rebuilt on the delta1024 reference hybrid model. Does t2va and reference. Treat i2v as a single image reference, don't do i2v prompting. Silver's merged turbo is integrated and tuned for 6 steps with simple schedule with the extra _emb layers tacked on to the model, not sure if they're needed.
Samplers like er_sde or multires or other turbo sampling setups work. Best imo is just er_sde/simple 6 steps, no shift, no spectrum, no cache, only a comfy_kitchen backend selection node. 3 video references in a 15+ second outputs can be done in under 10-15 minutes now on larger cards with no extra cache or quality hit needed.
Many (like a lot, all the good ones) on-site loras were fully combined into a consensus weighted merge with ranked drop-out to form an initial part. That merge is put against new, more powerful Wan and LTX grafts as a blend/reshape that uses the loras to consensus shape the grafts, but it also allowed some of the better loras clean pass-through. This is not a linear list of loras just merged. The main element that can present most is probably MysticXXX which was given the most pass-through weight since it's just good--and all 3 release steps of that are inside it. However, they're all-combined with a ton of other loras with agreement and consensus of shape and then it's only reshaping the grafts parts. The results of that are the actual weighted loras that are used to make the model.
Due to drop-out and consensus merge, pretty much all of the loras can all still be used easily on top if needed, and might work better even. All of this was only done to create a large rank dummy lora similar to what sulphur data will look like as a lora or extracted lora so I can start looking at how to apply it cleanly.
It's definitely not a few on-site loras that are linear merged, uncredited, and then renamed with some emojis. I'll only do this until sulphur tuning steps are in my hands and I can work with more targeted and shifted stuff, plus I was tired of waiting and I wanted fast easy i2v.
There is one quirk of the hybrid h3 usage: don't use it i2v. It should either be always used in reference prompting mode, or t2va prompting mode. Even if there is just one single image input it needs to be used as reference and prompted in the ref2va format. If you run an underdeveloped or manually written prompt you will get odd outputs, random camera changes, and blue lighting color shifts when you use the i2v prompt style.
Full credits to these lora makers for being involved somewhat in beta3 version:
alcaitiff, MisticRain69, diogod, FourBunny, HearmemanAI, tazmannner, simonishere, QualityControl, blo01, ComfyTinker, kermitfrog1202
Beta2 and previous:
This started a finetune-by-graft. Or maybe a GST - grafted shift of transformer (cross-architecture). I made both up, because there aren't any projects that have done it that I know, except one reddit post that made me look into it. I experimented with Wan and LTX on the side which led to the initial LTX Eros scripts that became what powered this, all before H3 ever came out. It seems like unified unbiased models like MMH3 can technically take attention influence from any other DiT without breaking if done correctly. Anima, Krea2, LTX, Wan2.2, Flux1 were all tried out, configs tested, about ~40 hours maybe of working in the dark without any paper or technical documents from Minimax. Eventually I developed linear-magnitude blend application and specific block and head gate targets allowing for a smoother graft on an attn-triplet-unfused version of H3 output as a patch file. That sent to lora extraction, then merged to checkpoint at taste. This is a merge but a merge of LoRas I extracted that interact to produce this current shift. I saved 5 ponds of water by recycling data in a few minutes on a single card instead of toasting a server up.
Turbo not recommended yet for i2v, especially when used with other LoRas. T2V use with turbo is better. Use 20-25 steps normal sampling with no dialogue, 25 steps with dialogue along with cache nodes and attn modes. More steps over 25 are not neccessarily better, and can be worse. Use full int8: int8 model, int8 VAE (if it doesn't crash comfy), int8 qwen3vl along with current cache or attn mode nodes. For smaller cards: quants, macOS ports, and Wan2gp support will likely appear on huggingface but not from me.
Known quirks:
Audio difference v.s. Base - This model's audio changes come from attention shifts seeking alternate audio pairing. Attn triplets were unfused before graft, both standard and triplet q_attn was grafted holding about maybe 10-15% audio influence, attn_k was frozen and MLP fc2 layers were untouched resulting in minimal audio interference. This was the main issue with the entire transformer graft and protecting audio. However this version is slightly louder overall than the base model.
Low resolution detail smearing - Some finger digits and fine motion will smear more at low resolution, also a problem in base model. As memory use gets more efficient increase resolution or work on the composition to get around it.
Odd outputs - This can attempt certain concepts more liberally than base model, but that can lead to some undesirable outputs in bad prompting and certain contexts. Data shift comes from completely different transformers and architecture. This shouldn't even work, so it is what it is.
This model is not dedicated to NSFW as that would violate community license agreement. Sure it can do it, just like base. Any NSFW generations are purely the result of advanced reasoning and tokenization resulting from experimental changes. All terms from the H3 community license also still apply to the users of this version. Don't be a dumbass.
H3 usage still requires very intense prompting for maximum effect. Every motion, every interaction, every sound plainly and fully described. Not with slang terms; with proper actionable words that can be tokenized. Refer to the h3 developer prompting guide, hand that .md file to an LLM or Chat agent and have them enhance or refine prompts along the released H3 developer prompt guide styles using the model's tag system. Certain concepts can be made from pure token reasoning. Consult the prompts in my previews to see certain physical descriptions that I use for some things. When using enhancement give the agent feedback about any issues in the generation and get them to describe motions in alternate fashion, or manually edit it yourself adding a negative like "no X, no Y". Still requires prompt refinement and trial/error for best outcomes.
Sulphur Project has 10k banked to attempt actual tuning. Right now training pipelines are sub-optimal. As always Eros is my personal side project, and this beta was also essentially a speed-run of finetuning, figuring out exactly in what configurations and target areas do you get helpful/harmful changes in the model. This is also a proof-of-concept of what and where to target while leaving the reinforcement quality of base unharmed by being additive.
Description
Hybrid delta base, turbo inject, graft+loras drop-out rescale, and consensus TIES merge.
FAQ
Comments (139)
This legend used my noob LoRa, thanks, downloading to test :)
It added it's piece on top of the soup yeah.
I am on a journey, nay, a crusade, to take the word slop into the positive realm and disarm all the AI haters and make their AI slop remarks meaningless.
So, @tenstrip , I think you mean the slop.
Bad meaning good etc
okey this is good, even without turbo lora....
TURBO is right in the name man, it is TURBO model and burns as such
which one is best to use? the INT8 or BF16? one is double the size than the other but not sure if it's necessary?
int8 since it's faster. The prompt accuracy can suffer sometimes with video reference being mixed up but that's about it.
@tenstrip Yeah was just testing and BF16 was taking twice the time, makes sense since that's more for the much beefier set ups, I'm working with a 5070TI so INT8 makes more sense. Anyway great job as always.
Could you include workflow in your examples next time, so people could just drag and drop the file into comfy?
I dropped a bunch of the comfyui saved outputs with workflow embedded. Civit doesn't read my workflows anymore for some reason. It's just the default comfy H3 templates with some image size sliders and image center crop instead of their MP resizing for i2v, and then ref is the same but no resizing of any videos or images. I don't think you want any resizing when you use reference. Then just the comfy_kitchen on the backend selector node added to those next to the model load.
Why you are not asking for a million dollars for the beta? are you ok?
Why not "skip edges" variant?
Far away to middle range human face in super low quality is still a big problem which make the model useless.
Are you talking about ErosMax, or MiniMax? This is a common issue with MiniMax, confirmed by the original developers. If you make a higher resolution video, it reduces the issue. Fixing it will require a new model, according to the developers. They can't just train it out.
@Jellai while it is doing I2V, for pic use in the first frame is a very sharp and clear photo with a clear middle range face, but using the the beta 3, it make the first frame quality more worse. while using the ref beta 2 (not this beta 2) for i2v, which do not happen. and of course this may be caused by my set up fault.
@takomli2013979 Oh, well, tthis particular issue may not be your fault, since the beta3 description says not to use it for I2V, but to use references instead. Still, it sounds like a regular Minimax issue to me.
Yeah if the person's face is only 80 pixels and the shot is zoomed out it won't have detail on any video model. Need to work around that it's not like that's only an h3 issue. Zoom in.
"Honey wake up! Tenstrip dropped a new model again!"
The fused Turbo is color-shifting/burning the image, which is expected. Not visible on short clips, but adds on top with longer clips / mutlishot.
Visible on short clips too right at the beginning and weird artifacts might appear on shot changes. (7steps, er_sde/simple, custom sigmas)
It also eats 6gb more vram without external turbo lora, I think ComfyUI struggles to create lowvram patches.
Use 6 step er_sde on basic simple scheduler, no shifts or spectrum or anything. And use the full constraint prompt strings like: For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced. Plus the entire reference prompting format for subjects and the reference summary: keep lighting, scene, color palette, hues, contrast, visual style, etc.
@tenstrip Will try, thank you.
Also that slight visual shifting sometimes isn't the turbo it's the sheer strength of the out_proj coming from krea2 and Wan2.2 as well. It's just the trade-off of putting t2v/i2v anatomy into it in this current method.
@tenstrip , @vladulidlo is right generating 10 or 15 sec clips always drift or fries the colors somehow either completely warm contrasted colors or opposite. Definitely turbo/distill behavior, I tested both my workflow and yours tenstrip, same output. Anything higher than 5 seconds gets fried colors or atleast drifts to the point where it's visually impacting. Hope this helps! other than that model seems to behave pretty good, great work!
@tenstrip a version without turbo lora baked in would be appreciated by a lot of people I'm sure but I think you would have released it along the turbo one if you could, let's see what the sulphur team come up with
@MiraStudio I've done 15-19 second generations video reference and single image and that doesn't happen when the models used reference with a prompt that locks down visual style and colors explicitly. I've said multiple times and on the model info the issue is doing i2v's with substandard prompting control and using the FL node. You need to run a normal default ref workflow, remove all the cache and spectrum crap, and do reference node for i2v with the proper prompting. Constrain visual style, lighting, lighting hue, and color palette inside the prompt explicitly.
@tenstrip Already did all the stuff you mention... I work on PRO6000 and never use caching and as I said I used your workflow embedded in the preview videos, depending on images provided as reference the model drifts the color heavily and it definitely looks like turbo lora behavior. with further testing it looks like it happens more on 3D images and images with more colorful palette than a realistic image. There's 100% something going on with colors and it's not due to prompting I'll do more tests and let you know. also might be worth mentioning it's not happening on all gen depending on images, I tested both int8 and bf16 and same behavior
@tenstrip just tested 10 more different images as well as running your exact workflow and input images. your preview run works fine and 5/10 of other random images I used all in 3D, half get color drifts and prompting is based off your prompting on previews with explicit color and reference image color constraint all locked in still get color drifts depending on images. realistic never fails pretty much, you can notice a slight color difference on the first seconds sometimes but on 3D images half of the time it color drifts. only speaking about long runs 15 sec. worth mentioning this does happen with turbo lora and base model as well as yours so looks related to me, I never have this kind of issue with base model int8 both ref2v and fl2v or img2v
@MiraStudio You can try to substract the weights of the exact baked Turbo lora. Set negative strength with the exact baked turbo lora, say -0.2 or more.
I don't think you should build on a hybrid model, fl2va is better quality and reference works great with it already.
Video and audio reference doesn't work accurately on fl2va. This does the exact same i2v, t2v, and ref outputs compared to the separate models. My i2v is an over-exaggeration of a rare issue on some seeds. I can drag any of my older i2v prompts in and the work fine.
Solid model, only thing Ive noticed thats jarring so far (been like an hour) is it always changes the reference penis quite a bit.
Check the reference prompting and <picture x> usage. You can always make the penis it's own <subject> and specifically target it and call it to lock down.
I did - but yeah it tends to shrink it if youre going for that extra long penis
To be fair, this is just a test, most people dont need a penis thats like 15 inches long lol
@Sassyphrassy Extreme anatomy is always gonna be on fringe stuff that needs it's own support. But that's the point of reference if you have the image(s) or videos of it referenced and prompted correctly the output isn't going to overpower it's own referenced conditioning. There isn't that much lora interaction going on you can still output 1:1 with the base model in most cases.
@tenstrip Yeah of course - I did prompt it correctly, and it works in the normal BF16, but I imagine something in this merge is overriding it a tad. I did get it closer the the reference image through some trickery though.
Other than that - great model. Everything comes out much more natural than lora stacking.
UPDATE: yeah was a skill issue. My b
@Sassyphrassy yeah that's usually the root cause of most feedback I get. I'm used to it.
Fast, accurate, amazing! thank you for your work! :)
Which version of MysticXXX was used? v3 was recently released and seems like a big improvement.
All 3 released steps of it with v3 having the highest weight and pass-through, drop-out and consensus merged with all the other loras.
@tenstrip Cool. Have another question I forgot to ask. If I am using i2v with ref2va prompting, does that mean I should use the ref2va nodes?
@d_A_v_I_d to my understanding you can use the i2v node as long as you use ref2v prompting as it was trained that way. I might be wrong tho
@KiraNugget it's something with the hybrid base used, which is a more accurate reference hybrid than the block merged hybrid models people usually use. But there's some issue with i2v being confused as a loose reference and it can have prompt misinterpretation. Just seems to work better doing i2v with a reference node instead.
@tenstrip I only 2 REF2V anyway now as it is extremely accurate in character preserrvation and general knowledge and understanding. Best thing since sliced bread.
Minimax is a very expensive model to train on large data. The time per step increase I got from increasing from 256x256 to 512x512 was insane. Went from 8s/step to 34s/step with no block swapping and compile fully enabled without silent graph breaks. So godspeed to Sulphur because minimax training makes high end hardware feel like a raspberry pi.
Well, that is 4 times the pixel count and just about 4 times the training time, so it doesn't seem too unusual, though it's true training on other models usually didn't scale with resolution the same way.
There were some promising test runs. The full run is gonna be on h100 or h200 stacks.
@wewewew LTX the time scaling was much lower but LTX was also much more compressed in terms of its latent space. LTX the time scaling going from 256x256 to 512x512 was a doubling in time instead of quadrupling. No free lunch I guess but minimax is worth the compute.
@tenstrip Spicy amounts of compute
I'm getting broken, complete noise outputs with beta3, and error "[WARNING] unet unexpected: ['adaln_basis', 'adaln_mean', 'silu_t_emb_grid']", probably related to the dareties Lora which also does that if I don't use the "H3 AdaLN LoRA Fix" node, though I don't know why nobody else seems to have this issue with your model. Lora fix node doesn't fix it either.
the unexpected are unrelated errors, those unet flags don't mean anything. It doesn't need that patch node either. Run it in the default comfyui template and don't use anything else. If that doesn't work you have a backend issue on your comfyui and need to update for fix some packages.
playing around at low resolution to build a llm system prompt and this is nice even for t2v ;)
The best by far.
This is so great. Thank you.
When I make sex videos, scenes of anal insertion keep appearing.
The model seems to be very much close to original ref2va model based. comparing to hybrid model that i use (b15-49), reference power is much stronger with lower output quality, just like using the original model.
How to use it? Replace the main diffusion model node? When I did it I get only colorful squares as a result, not video
Just use default ref workflow built into comfy and set steps to 6
Outstanding work. Best H3 checkpoint I've used yet, and its so fast. Bravo!
Really Great Model and speed! Thank You :D
gguf Q4 anyway plz ?
Ótimo modelo, muito superior em aderência ao prompt em comparação com outros por aqui. Parabéns!
Só uma dúvida: esse modelo já está com lora turbo integrado, é isso?
Se sim, então podemos executar com quantos passos, 4 ou 6?
6 steps with simple, and doesn't need spectrum or cache they don't speed much up at low steps but will take a lot of prompt accuracy away especially when doing reference.
To clarify here: do LoRAs need to be tailored for REF2V specifically? Lot of prompting motions that worked in I2V and such have results still come out weird.
I wonder too but in ref2v I tend to use as little loras as possible. one, maybe two at slightly lower strenght than reccomended. Except reasoning lora, which seems to be a good all rounder for everything, that is always in and at 1 str. Anyway in REF2V you can simply input an image of whatever you want so no need really for anatomy or character loras.
Ref loras don't make sense. They work better when they're merged into the model which gives it more insight into different motions that it can pull on to reference, but the way they work at inference doesn't give that affect it just interrupts the model. Better to just train them on fl2va to begin with and then let people make merges with them.
@tenstrip good to know, after a lot of testing this is what i've figured since everytime I stacked a couple loras or one at high strenght it would mess up the entire video. Would you still recommend the reasoning lora with this checkpoint? As it is in suggested resources.
@tenstrip Thanks, that's good to know. And when I said "come out weird", I meant in any other REF2VID model, sorry. Works great in yours.
@DarkEngine2024 I figured out the issue already and remade the turbo it works in every mode correctly in the next version.
Oh sick, looking forward. I can't believe how great Eros Max has worked in REF2V, it's bonkers.
On that note, do your examples also use the Heretic version of the Qwen 3vl text-encoder?
Identified the beta3 drift and blue lighting issue to the turbo lora. Making my own hybridized merged turbo and an improved mix for the next one.
My generations often tend to go green on skin tones throughout the clip, is that what you're talking about? Is there a way to fix it in the meantime?
Yeah knew there was something going on with this! hope it helped you find out what went wrong!
Yeah all of that is fixed I have the next version getting prepped. I made my own hybrid turbo merge it also has much better visual consistency and doesn't layer on any kind of saturation or contrast on conditioned i2v and ref.
Working well in DaSiWa's MiniMax H3 Workflow. Using 6 steps, the FP16 video vae, and fp32 audio vae, and the nvfp4 vae.
i had to modify the workflow for 0 shift. but it was easy to do. Currently testing without loras. I expect adding loras at low values will only improve output.
please report back on how you go! i'm currently using dasiwas checkpoint with dasiwa workflow but struggling to find the right balance between speed, output quality, and lora effectiveness.
I've heard some people using 12-15 video shift, but it's not balanced or made around using any.
"i had to modify the workflow for 0 shift" How did you do that? And do you run video and audio on 0 or only video?
@QuietSpark Enter the Settings subgroup, find he nodes named "ModelSamplingMiniMaxH3", there are two of them. Add two reroute nodes, and then connect the model string going into the "ModelSamplingMiniMaxH3" node to the reroute node, then connect it to the where the model output of the "ModelSamplingMiniMaxH3" nodes are going to. Do this for each "ModelSamplingMiniMaxH3" shift is now automatically zero. This can be further customized to add a fancy switch, but I didn't care enough to do it.
@hardwire666 Awesome, worked like a charm, cheers! And the results are way better!
Why use this over picking our own loras with the base Minimax H3? What are the tradeoffs?
Loras run un-normalized on top at inference and usually stack problematically causing so many people end up having to use shift to reshape the model so it doesn't look bad. The ranks also stack up taking 1-4 gigs of memory with just a few. I use them change all attention blocks based on a full model graft with magnitude based merges this isn't just model + loras. They're all also consensus merged forming an agreement that behaves more like a tune.
Impressive work! I've been using the FP8 ConvRot mostly, but this turbo model cut rendering time in half with no appreciable loss in fidelity
I would love just a smidge of info on the following
'It should either be always used in reference prompting mode, or t2va prompting mode. Even if there is just one single image input it needs to be used as reference and prompted in the ref2va format.'
So, I understand that the best use case is like how you genned, with 2 images, ref video, and good prompting. Are you also saying that I2V will work as long as we prompt as you did (Using the Ref2va format) and that T2V will work as long as we do the same?
If this is beyond your scope, or better explained on the main page, just tell me, I won't hold it against you for not reanswering something.
We need quantization! gguf! gguf! gguf!
@SolidCode Thanks a lot, I'll try. But I usually use Q8.
So if I get you correctly we are supposed to use this and do a 6 steps directly without any of the usual Turo LORAs?
GGUFs would be really nice, as it works much faster when there's not enough VRAM.
@SolidCode edit: nevermind, it's not GGUF. These smaller in8 versions give lesser quality, unfortunately.
@dobomex761604 Mine is a mixed quant similar to a gguf. try it.
prompt adherance is weird, on beta3,. characters barely move. i tries a fight scene for exemple .
Interesting and strange. It just seems to me that this model listen very well to the prompt.
even if it is written simply, not formatted according to the instructions.
Yeah once beta4 is out I might rephrase beta3 and just say to ignore it. I'd remove it but it has some okay gens. Beta3 isn't a real ref model because of it's turbo and i2v doesn't work for the same reason. The new one I merged for the next version is much better. This model is entirely prompting make sure your prompt even works on the base model with turbo and same settings first before making a claim.
@tenstrip But this version WORKS. Late i can upload more videos from I2V. For me is very good version of model.
@herkus_baronas631 The next one is overall improved though. Audio effect is there but not like base audio, it needs some audio shift to sound better. I had the blue/green lighting bug on 3 plenty of times and at least that issue is gone.
@tenstrip Would be interesting and nice. But as sayd - i would share some small videos with version3, for examples.
@herkus_baronas631 let's say motion is good on close up but when i use full body shot, it's less reponsive
By far the best model. The Ref2VA Turbo LORAs weren't working, they were causing motion blur. With this model, the reference mode works perfectly without using the Turbo LORA. Congratulations.
Incredible checkpoint. I rarely comment here- but by far a must have for H3 users. Excellent work OP.
unfortunately the skin looks waxy compared to the original H3 model.
Beat3 is an amazing model with excellent reference capabilities, significantly improved clarity, and eliminated some dynamic noise. Plus, it's much faster than the official model combined with accelerated LoRa! I achieved great results in just 4+2 steps; I think other models can be discarded! Fantastic!!
After entering the close-up shots of the face where the proportion of human faces is relatively large, many moles that did not exist before will appear on the human face. This issue exists in both int8 and BF16.
Yeah the turbo is too strong. I already made one that keeps style without overdrawing anything on 6-8 steps.
@tenstrip Thank you for noticing this matter and taking action to resolve it. Thank you for your efforts. If possible, would you consider providing a non-turbo version and a turbo version separately in the future?
Best model without a doubt. But the cum could use some work. However I prompt it just produces excessive amounts, and very often you get a weird popping sound...
Thank you for v4. Love it!
What is v4? You mean beta3?
@Eliz99 beta 4 is out on his hugginface.
Great model, getting some really fun results. Although I have two questions:
1. Oftentimes movements are too fast or exaggerated. I can sometimes get more normal movements by using words like "slow" or "deliberate", etc. But is there a better way to ease up on the motion?
2. I frequently get "blue" results. A blue filter, blue smoke, etc. I've even had two gens where the background "turned off" like it was on a blue screen, and another that went from a couch into a blue bathtub for no reason. Any thoughts or ideas on that?
Not sure about your 1st issue, but the 2nd one has already been discused here in the comments 3 days ago. The author has identified the issue and it should be fixed in the upcoming update.
Yeah both are something fixed. But the new version trades that for needing 6-8 steps and some audio shift set higher to have even okay audio, and audio is overall worse. Already working on future versions past the unreleased one to fix that too. Main issue is creating a hybrid turbo lora merge with the underserved reference ones available since they all focus on t2va.
@PseudoGrafx , @tenstrip thanks to both of you.
Oh my !! We are getting there!
IT's So Fxxking Great! I WANT TO GIVE U MY ASS,IF U WANT.
ew, it's turbo again
@zanebe The non-turbo mix doesn't even work. Even with other turbo loras used with it which aren't hybrid or ref enabled. It's a waste of a 40g upload and download. The model is a hybrid model and needs the hybrid turbo, then the whole recipe is grounded in the deltas from the turbo I merged and only really works well with it because it normalizes the merge's behavior through it's fixed target strength. Last thing I'm gonna do is constantly tell people what turbos to use and how to sample it or deal with people commenting who can't read, especially with how bad the mix actually is without a turbo. Also it doesn't have any issues from the last one or other turbo issues except audio issues.
I would love to try this, but it seems like I have be granted access?
@tenstrip i've noticed the audio seems to break at the very end of a video if it's person's voice. But its only like a jump in sound level like thing for me.
@zanebe even with a 5090 I couldn't imagine using minimax without turbo.
@delta45424155 save the video at 24 fps and use the h3 flow shift node
No clue why but training audio into h3 seems impossible. Want sex plopping noises? Too bad you get the sound of a piece of wood being hit by a rock.
This can actually do some of those sounds, "dull flesh thudding" will do it. But without shift and some luck it sounds like the metal talk tube things on playgrounds. It could also be invasive bad audio from one of the 23 loras. I know that I cut HM's breast lora because it has bad audio, maybe all of his do.
Beta 3 had been my go to model for the past day or 2. Can't wait to try beta 4. Things change so fast with this stuff.
This was just a rushed patch version for beta 3 because that one had serious reference and i2v issues. I know that there is a sulphur preview lora coming that will improve this a lot and that with some better turbo will make a lot of gains.
Way to go my man! Best checkpoint available for H3. Looks like you’re pretty much out of beta, excited for V1 to officially drop.
If you haven’t already definitely incorporate these Lloras:
https://civitai.red/models/2856467/mmh3-mystic-xxx?modelVersionId=3266628
https://civitai.red/models/2834417/hmnsfw-aio-sex-lora?modelVersionId=3268303
https://civitai.red/models/2837571/minimax-h3-turbo-loras?modelVersionId=3275758
https://civitai.red/models/2856004/breast-play-and-jiggle-for-minimax-h3?modelVersionId=3278283
Keep up the great work. You’re about to have the number 1 NSFW H3 model
How did you managed to get that great quality with only 8 steps? Mine comes out like the camera lens just surfaced out of a the jolly green giants used toilet.
You have some sampling issues or problems. Load it into the default comfy H3 template and just change steps to 8 and sampler to euler and try that before adding any other model nodes on it.
@tenstrip Turns out.... I'm just an idiot.
I had a wrong image used for I2VA that was vastly green and was making everything look like a wet vegetable pudding. So it was just user error on my part.
Keep up the fantastic work by the way! Everything looks solid!