## PK Parasyte Turbo - For H3 Minimax
help me get a 5090
https://throne.com/plaguekind/item/073bace1-2e1a-4492-95e6-024b4da6459b
- This is an experimental purposeful mutation of Dareties in an attempt to remove "turbo" behavior and clean up the models visuals and audio.
Reccomended starting point for settings are
- 1.5 strength, 8 steps, 12/3 shift, er_sde beta57
- some weird prompts might want 3.5-5.0 strength, 6-8 steps, 8/3 shift, Euler Simple.
- Requires my Adaln lora fix node.
- It is designed to be used with my SLA node.
- It can be used at 4 to 8 steps depending on the scene.
- I would appreciate it if you would mess with the settings when testing it and let me know, I'm sure we can find a more appropriate setting.
- Tested under all modes, T2V, I2V, R2V. They all seem to function well.
- 5 audio shift might be useful for some situations. please mess with both shift settings.
- Make sure your timestamping and prompting are accurate before counting this one out.
### Support
If you find this project useful and want to support development:
Monero (XMR):
865BrcfWLdwELwuq5faV1uVTbh93zVK6AUYLY2c3mX6sFfAGRfS6axe1kBTYYKuM7ccN7zBZDAZvnT7E4NKmUazySdbpc7pDescription
## PK Parasyte Turbo - For H3 Minimax
- This is an experimental purposeful mutation of Dareties in an attempt to remove "turbo" behavior and clean up the models visuals and audio.
Reccomended starting point for settings are
- 3.5-5.0 strength, 6-8 steps, 8/3 shift, Euler Simple.
- Requires my Adaln lora fix node.
- It is designed to be used with my SLA node.
- It can be used at 4 to 8 steps depending on the scene.
- I would appreciate it if you would mess with the settings when testing it and let me know, I'm sure we can find a more appropriate setting.
- Tested under all modes, T2V, I2V, R2V. They all seem to function well.
- 5 audio shift might be useful for some situations. please mess with both shift settings.
- Make sure your timestamping and prompting are accurate before counting this one out.
### Support
If you find this project useful and want to support development:
Monero (XMR):
865BrcfWLdwELwuq5faV1uVTbh93zVK6AUYLY2c3mX6sFfAGRfS6axe1kBTYYKuM7ccN7zBZDAZvnT7E4NKmUazySdbpc7pFAQ
Comments (65)
Read the description before commenting.
i love the name you gave to it
Gotta keep it in theme, it's what it is afterall hah
I'm seeing very good results here. I'll post 2 I've done here no turbo 20steps with spectrum and 8 steps no spectrum with this lora at 4 str 8/5 for now very good results. I'll send it soon
cheers
Hey man, do you ever sleep ? XD you deliver like crazy.
Hah, it's been a hectic few weeks.
A good initiative, but at the moment, with a very precise prompt in line with the manual, it has multiplied the number of characters and is disrupting the audio input. I compared it to minimax_h3_ref2v_turbo_4step_v0.1_comfyui_resized_avg_rank_21_bf16.safetensors + spectrum, for 15 Euler steps.
GOAT
Since I'm not from an English-speaking country, the differences I personally notice might be a bit different from others — but the er_sde / simple setting feels like it gives better intonation and a more native-sounding result. The audio is so much improved!! Amazing. I'm also using it together with the digicam LoRA 1, adjusting the turbo LoRA strength around 4 – 5, to reduce the degradation a bit for realistic/live-action work.
Yep told understood the message, I will test it out. I did do a few with er_sde as well.
Excellent work. Audio performance is perfect.
But in my case, with FL2VA INT8 model and 1 anime ref image, I tested 5 LoRA strength values ranging from 1.0 to 5.0, and get obvious color shift when LoRA strength > 2.0.
5.0 might work better for realistic images.
Additionally, when turbo LoRA is enabled, other stacked LoRAs may require lower strength values.
Using shift 8/5 seems to resolve this issue. It’s now working fine in my workflow.
Just wanted to drop in and say that your workflow for minimax H3 is out of this world and now with the introduction of this lora, you're really doing a lot to shake things up! You're not a plague, but you are kind, lol. Many thanks!
Lol! thanks. I try to make it easier for people to avoid the problems that I've encounter by releasing what I can.
Trying at 5 str, 6 steps, 8/3 shift, euler/simple. With a few terrible prompts. Initial impression is that it's faster, better quality and somehow nvfp4_awq text encoder is almost working as well as int8, but most importantly it doesn't mess up my lora's and their weights. Mystic v4ref2va can be kept at 0.9 fx, where i got poor mixed results even as low as 0.2 using other turbo loras.
Yeah i just straight up deleted the dareties and 768 turbo's. This is massively superior, even in experimental.
Boy I'm really hooked on the PlagueKind eco-system, from WF's to loras.
Glad to her it's working well for you!
str5 shift 8/3 8 steps , seems to be on par in terms of audio video quality with combo 0.4 0.6 of dareties loras u added in your latest wf. str 4.5 and lower produces metalic voices, 4 produces softer image. Prompt following is good with any str. EDIT: complex prompts 15 seconds it breaks prompt following, it does not complete prompt which I already completed easily earlier with dareties loras, tried with multiple seeds and it did not help...
I'm doing a 25 second prompt right now which is working fine. I'll be posting the video later.
@Plaguekind I have isolated audio problem to a Lora that was active in loader and was not needed. However prompt following is hit and miss, unlike with dareties in current v7 and before that v6 workflow that just stick to what is written, here str range 4 to 5 it cannot do these parts properly that other dareties loras chew through first run :
[Shot 2] At 00:08.000
The camera stays on <Subject 1> as she reaches a diving board. She quickly kicks off her high heels and takes a short run-up before performing a skillful, graceful dive into the water of <Subject 2>.
[Shot 3] At 00:10.500
A sudden cut to an underwater shot shows <Subject 1>'s body slicing through the turquoise water, bubbles trailing her form as she dives deep before beginning her ascent.
(she does not kick off her high heels, she jumps into water with her feet first, then underwater it shows her diving in head first. and then randomly diving board will be shown as underwater object, also she starts jump from the middle of the board with old trusty dareties she walks to the very end of diving board and then dives into swimming pool )
just did test with pruned v4 rev2. silveroxides lora str 1, 6 steps and it completed prompt fully at str 1 and 6 steps.
@Plaguekind I found what caused issues, timeline was too short, adding extra 2 secs while using 4.8 str on the lora and the prompt was executed correctly. It gave breathing space to the model.
No more melting distant faces!
Bout time right?
teach us bro
I'm sadly way to dumb to understand how this works it seems.
Figured it out :D
whats the issue?
@Plaguekind It was mainly me being dumb tbh :D But i figured it out now. I checked your workflows and the one from the guy who uploaded a sample and now i made it work with my own workflow. Maybe is should sleep more... anyway, thank you <3
@Lora_Addict I think I will be able to now finally. Maybe?
Replaced the 2 loras with yours and created some kinky clips. Looks good to me.
Shift Video: 8
Shift Audio: 5
Steps: 8
Glad you like it.
I agree. Shift Video: 8 Shift Audio: 5 seem to be the best.
This is really good! thank you!
Cheers
hmm, so I'm not sure what I've done wrong. I tested with your v7 wf with default settings. Ran the warmup gen with the two original dareties and got great results. Then I just disabled those two and enabled this new one at .5 str, 8/3 shift, euler simple at 6 steps an the results look bugged. Smeared, low motion, metallic audio. Obviously other people here have gotten good results. Not sure what I'm missing.
It's 3.0-5.0 str not 0.5 💀
@Plaguekind yeah that did it
@tehorhay24 mind editing your original comment? Thanks
Edit: I did one last test with a more normal lora strength (1.5) before dismissing this lora and got outstanding results! Like, better than my 20 steps no accel baseline in all aspects, video, audio, motion, adherence. I need to test some more, but you should take another look at your 5.0 str recommendation, it seems wack.
I got alright results with euler simple 8/3 and er_sde beta57 12/3, but overall this lora is way too sensitive, randomly gives awful results, and the best ones don't come close to turbo_v4_step600_ema anyway. I don't see any better compatibility with your SLA node either, turning that on still destroys prompt adherence.
Glad to hear you got it working.
Which lora loader are you using? Because with a "normal" lora loader, the functional strengths are between 3.5-5.
With adaln lora fix, and sla
@Plaguekind With normal loader, yeah. I've been testing 1.5 str more and it's performing consistently amazing in both live-action and anime, T2VA and I2VA. 8 steps er_sde beta57 12/3 shift, 0.5mp, adaln fix, comfy kitchen attention, nothing else.
I'd say someone should verify these settings if it isn't just something weird on my setup, because otherwise they're really outputting great results on really difficult testing prompts that even vanilla 20 steps struggles a lot with. Like I2VA anime dancing with rapid spinning and complex unusual moves, that usually destroys any recipe, but this lora is the first time I see decent quality and legs actually remaining anatomically coherent.
@Plaguekind Hello Sir. Is Adaln lora fix node entirely necessary? What benefits would I gain from it? Versus rgthree Power Lora loader, which is my favorite. As you can from this image: https://i.imgur.com/Gp48AYV.png
I like to test loras but The Adaln Lora fix comes up short in lora capacity. You can only fit so many. Any intention of increasing the lora count? Perhaps as soon as possible. No rush. Altho' i still don't know the significance it lol
@wewewew you may have just found the sauce bro. going to keep testing.
@ProvenFlawless you need the adaln lora fix and it's supporting file for it to function correctly. without it will still work but be gimped. it's not a loader, its just a node that brings weights from unpruned.
@Plaguekind I've found that 2.0 and below creates jarring camera jitters for me so I do not know what he's doing different. 4.75 is where I landed.
@InAbsentia alright, give his settings a shot. i've updated the description with them, they work great.
I came to a similar conclusion a day ago, but I still don’t know when 3.5~5.0 is required. Perhaps it depends on the prompt structure?
@Plaguekind Could you give some examples?
@a91921722914 some highly chaotic scenes with a lot of noises and music going on, and some style scenes might require it. If you don't encounter them, great.
Version 2 will be named...drumroll...FaceHugger Turbo lora!
Lol
Returned with a bunch of Error lora invalid error nodes adaln.proj.linear...
Did You use the adaln lora fix node and it's supporting file?
@Plaguekind No, I don't know how to set that up.
@sarahbonestone lol. Description clearly says "Works with my Nodes"
@sarahbonestone download my wf it's got everything there, and description of what you need to do. only difference is the lora isn't in it. I will add a version with it today
@Plaguekind Ok, will try.
Checkout my huge H3 PK Parasyte turbo and fasth3_6step turbo lora test.
https://huggingface.co/datasets/Turpo/Minimax_H3_Parasyte_fasth3_6step_test
Feel free to give feedback.
Thank you for that.
what did you conclude?
@kknightowl At the moment I think 2.0_1.3, 2.3_1.2, are good for low action scenes. 3.5_0.5 works the best for low res in action scenes. If you go higher resolutions you may need more juice (0.2-0.3) for fasth3 to compensate Parasyte "weakness" in combo. However, you can't go higher than 1.3 with fasth3 if you run it in combo with Parasyte. If you run fasth3 alone, then it works for me up to 1.5 weight good. Parasyte alone 4.0-4.5. Overall 3.5_0.5, 3.5_0.7 works best, fast action, background people, etc. looks the best.
@Plaguekind I've made up my mind. 3.5/0.7 works best with my dance videos. I get the best motion without artifacts and the least back tiddies and weird anatomy spins and things like that. And with er_sde/beta57. I haven't found better sampler/scheduler combo that beats er_sde/beta57. I'll even ditch sa_solver/sgm_uniform combo for my dance videos.
Has anyone managed to get whispering to work?
I tried both
1.5 strength, 8 steps, 12/3 shift, er_sde beta57
and
3.5 strength, 8 steps, 8/5 shift, er_sde beta57
the character still yells at me lol
<Subject 1>(S1) says: <d>[English] "<whisper> hello"</d>.I just uploaded my tiled latent upscale tests with V8 workflow with H3 PK Parasyte - fasth3_6step lora combo. All 12s. airport dance clips were done with my Lenovo laptop with 4070 mobile 8Gb and 32Gb system RAM.
I'm totally mind blowed by the results. This upscale solution is not only much faster than I thought, but the quality of the upscales is great too. This is basically a "local free Topaz AI upscale" solution. As I said, this is absolutely mind blowing stuff.
I tested with three different resolutions. 512x960, 416x768 and 320x576. And I made great discovery. More on that later. My laptop actually can go as high as 1.6M with both native + latent upscale in 18 minutes. But I think 1.2M is a sweet spot for quality vs. speed.
You can find my test clips below. I put all generation details in the negative prompt field of each clip.
First clip was made with native resolution of 512x960 - 0.5M and upscaled to 0.9M - 704x1248 in native + latent upscale in 14m12s. It's looks awesome. Best of the test, probably because native gen resolution was highest, even if it's only 0.9M
Second clip was made with resolution of 416x768 - 0.33M and upscaled to 1.2M - 800x1440 in 12m 21s. This looks great too. Not much to complain about. However, in comparison to the first clip, the first clip looks "richer" in details because of higher native resolution source. But If you haven't seen the first one, I'd say you wouldn't care too much. It's still good enough.
Third clip goes to category "it's probably shit, but I test it anyway, because that's what I do. Before this I've considered that 768 vertical resolution is Minimax H3's absolute bottom that you should use. Everything below 768 vertical looks like crap. I might have found solution here for the users of 16Gb system RAM and maybe with only 6Gb GPU. I would say it's worth a try at least. It's 320x576. I honestly think it looks decent enough if you're on "bottom cutoff" system with low VRAM and RAM. I absolutely want to hear if any of you guys with 16Gb system ram can do this. Maybe with shorter clip like 6s. at least, I would think. Anyway, here are the details of the third clip: 320x576 0.19M to 1.2M 800x1440 in 11m.20s. Pretty damn impressive, right?
What do you guys think? Aren't these impressive results, or what? If Plague you could pin my comment and airport dance test videos post, it would be great.
hi pirate, at what strengths did you use both loras?
@necroryona Parasyte - 3.5 fasth3_6step - 0.7. But 2.0/1.3, and 2.3/1.2 works good too for me. It seems that 3.5/0.7 is always the setting that I return to after testing. It just works for me. However, I recently changed fasth3_6step to fasth3_v1_dense_datafree_comfyui because it works better with bunny Motion Repair lora.
I keep getting a Windows fatal exception: access violation error when this specific Lora is enabled in your v9 workflow (which is awesome).
I've got the adaln lora fix node. I have the extra file h3_silu_temb_grid.safetensors in the h3_adaln directory, all my stuff for the rest of the workflow is up to date, I have the latest Nvidia studio drivers, and I've tried a variety of .bat launch arguments from chasing my tail with talking with LLMs. Hate to ask, but I've exhausted my troubleshooting capability, do you have any insight? The v9 workflow is so good I'd hate to not use it the way you intended.
WORKAROUND: Change the AdaLN node value from 'port' to 'strip'
Loving the results