(There is 2 variants, start with Default)
Takes a video input, results with bigger / skimpy attributes.
Trained carefully, but didn't had the time to test it further than these provided examples.
These references (left side) was generated with low quality settings, and turbo model, to show the worst kind of results it can produce (target on the right).
Doesn't seems to lose details, preserves composition and movements, as well as identity, but can change outfit colors (If it matters, prompting may help prevent that).
Workflow example is provided (into the Optional Files section "SecretSauce_MiniMax REF2V.json") that use my custom nodes, may become your next go-to for generating MinimaxH3 / LTX 2.5 videos (you will find the LTX ones in the custom node directory if you're interrested)
Note that the input video must match the Unified Sampler duration (frames) that is quantized to the model's requirements, if the node is set to 90 frames, and your video length is 89 frames, then you will need to use the previous quantized value on the Unified Sampler, wich is 73, and set your video "frames_load_cap" to 73 as well, both need to match if you want proper A/B comparison.
As for the Aspect ratio, if you want to preserve it, set matching choice in the R2V mode tab, all other settings like steps, and megapixels values, as well as the realtime preview / optimization stuff are shared across all modes in the "Settings" tab to avoid useless cluttering.
Two main datasets with different triggers that was used for the whole training.
-InvokeSS (clothed)
-InvokeSSX (nude)
Added secondary keywords mid training such as:
-Large
-Sagging
-Breasts
-Cleavage
-Clothed
-Nude
-Thicc
-Body
-Ass
-Hips
-Oily
-Nipples
You will find an alternative stronger version, that can add more motion physics, but also increase some attributes, barely tested, experimental)
Output embed Input audio, Identical (just cutted it from the examples)
If you want to support this model, you can do it on https://buymeacoffee.com/designedbycrt
Or https://runpod.io?ref=u7b2habt if you are using RUNPOD, so we both benefits with free credits, everything I earn from these is destined to train and share here and HF https://huggingface.co/UntMods anyway. Appreciated
Description
FAQ
Comments (46)
Morpheus: "No one can simply explain what SSX is, you have to experience it for yourself."
at starting to become concerend with the community knowlwedge of H3 because it seems like nobody understands that it can do this and scene transition nativly
I assumed that this forces them into a specific body type, since none of the body types match the original. But yeah, maybe H3 can do that too, if you're specific enough.
H3 an do a lot natively, you just need to be curious enough and prompt correctly, to get the result you want :p
A lot of H3 loras posted i tried with the lora and without lora and only my prompt, and get the exact same result.
Can you hook us up with a guide or something? I noticed I couldn't get it to match the original movement without the lora, but it also forces them to look like someone stapled the tits and vag of a dying 500 lb bitch onto the subject.
You need a really solid prompt for that, and the results will not be consistent over multiple runs, please share your prompt to achieve that kind of task without LoRA, I'm curious about the knowlege you're reffering to
does anyone know how to remaster a video with better lightening r2v i tried many different prompts but fails each time
I require sample prompts please :)
I also require a "already owned" models workflow,my brain is gonna explode toggling it by myself xD
"InvokeSSX. she is naked" works JUST fine. just doesn't put the audio back in for me. gotta figure that part out
@kronos1959777 In the workflow, the input audio from the video loader is connected to the output video combine (so it should be identical) verify that it is if you are using your own workflow
@kronos1959777 what about oily etc, i keep getting surprised lol
@pgc i am using wan2gp. i know not these nodes of which you speak.
@kronos1959777 but the breast is so big... how deal it
How should I write the prompt words? Should I follow the official format, or just use keywords with a simple natural language description?
Use few keywords only (5-6) see what works best for you. A single main trigger word (InvokeSSX) will have slightly less impact, but works too
holy crap!! that works. oh yea.
不管好不好 先给你点赞
Does this work well with normal Ref2va workflows? I really don't liked the unified thing. The S/IT seems a bit higher than a normal REF2VA workflow (no optimizations) and staging takes a long time. I set the MP, wired turbo and CK attention directly to both turbo and non-turbo inputs for the model pipe, so that it can't be a case of it choosing one. set MP to what I normally do, steps normal. frames = to unified sampler (73). no tiled or any VRAM saving options set, no spectrum.
even got memory leak warnings with the workflow.
I'll try with my usualy REF2VA workflow, with and without the lora, and report back
So, using my usual workflow, it's pretty good at "inpainting", subject awareness, so without the lora it will often replace the person with someone of similar likeness and not retaining body shape entirely. With this LoRA it retains likeness it pretty well and can judge body shape (under clothes) more accurately
it failed for me at r2va int8 , how many steps you used and just turbo lora or somthing else as well?
The unified sampler is mirroring the official pipelines merged in one single node, it have low vram mode that force unloading models after usage. Maybe try to set "Low VRAM" and "Unload VAE" off, if you don't OOM it could be faster, since models are reloaded from the system ram instead of disk. If your workflow doesn't have VRAM cleanup nodes, this could be the reason why it's slower on the provided workflow, try to untick these two and it should be similar
@pgc i am trying the default lora now, i didnt see that one before, i m generating only 5sec at 1mp , in normal r2v i have made 1mp 12sec videos easily with other loras.. lets se if default helps , thnx for quick replies
@chaiwala008215 I used various, still testing, but you can just grab one of the dareties turbos from silveroxides huggingface, and use the Adaln lora fix patcher node from PlageKinds node pack. I go 6-8 steps.
@pgc I usually don't need any VRAM saving because Comfyui is pretty good at managing memory. I have 8gb VRAM and 32GB RAM. any workflow i use for minimax doesn't have any memory saving besides KJ's Low VRAM attention(at 4) because it doesn't cost anything, it's fine without it too, though.
@AI_DK Yes dynamic VRAM is used anyway, shared GPU memory is always filled even with these options, but sometimes (for me) when the resolution / count is low, GPU memory at 23+ used VRAM, the dynamic VRAM doesn't kick in, so it doesn't OOM but also get stuck in between. But damn 8gb GPU with 32 RAM is tight, I didn't know it could run on that, I have a 3060ti in another PC but didn't even tried to run minimax on it
@pgc 3060 ti. using this lora 192 sec gen at 0.5mp 3sec. 6steps . I can do 3 sec .5mp fl2v at like 80 seconds (prompt+image to gen after model loads) with optimizations and quality is good.
its not working for me, how many steps to use on int8 ref2v with 4step lora at 6step it fails todo any changes with just prompt "InvokeSSX she is naked"
All shared examples are consecutive runs, it didn't failed to change the input over these runs. Are you using the provided workflow? Is "not working" means that the output is 1/1 identical to the input?
@pgc yes its coming as identical , the workflow you provided i cant seem to change the models unified workflows are so difficult, i tried to get my own r2v model int8 and lora but i cant seem to do it there, so i ran it on basic workflow at 1mp 5sec
@chaiwala008215 The model pipe node is a subgraph, you can open it and replace model loaders by your own, these ones in the provided workflows autoDL if missing, for ease of use, but you can use yours. If your results are identical over multiple runs, means that something is wrong with your lora, verify that the download didn't truncate the model, it would surprise me but it's possible.
For the provided examples, I used the 8steps turbo LoRA
@pgc please create a workflow where the users already own unet,lora's etc. :)
yes plz @pgc
you can keep the integrated_multimodal_description start.
I also used "The woman in <Video 1> is <Subject 1>" then you can describe what <Subject 1> should be wearing or not, I have also been able to make them do something, like raise their arms, but they might be forced back down to the position the video is guiding them for so i will put "at 00:00.100 <Subject 1> raises her arms and for the entire video they remain". worked.
gonna keep messing with it.
are you gonna train a t2v version?
I already did before this one, but I want to take the time filtering the blocks before sharing it
Thanks! I’d like to keep the same body type as the reference. But the breast keeps coming out unnaturally large though. Any way to fix that?
The body type change is the intended behavior, there is nothing to fix about that as it's the purpose of this model, you can try to use prompting to avoid what you don't want "small breasts", but then Idk if you even need this lora to achieve that
@gx_ground136 it's h3 model stuff, it makes tits bigger even without any lora.
@pgc but the breast output is too big ,can't adjust with prompt
Turn down the strength of the lora to 0.2-0.4. The lower you go, the less likely the effect will work at all, but when it does, it follows the reference image much better.
@tacocat any minimax lora is not working . to adjust the breast
having some success adding a picture reference to the workflow of just a pair of breasts with the size/nipples i want.
This is pretty amazing. You're a genius! This can actually be used to make even all that crap Hollywood released in recent years worth watching. Or well, it could be, if we had considerably more compute so FIXING a 90 minute movie with this wouldn't take... I don't know how many years.
Either way, thank you!
Bro. Can you make more choise in your workflow? Without auto downloading anything
WE AGREE ! XD
Replace the models autodownload nodes by your own, and save the workflow for later uses, so you don't need to do it again
@pgc just post a optional workflow please c: