Note: rgthree's group bypasser (the toggle nodes used to turn options on and off) is currently bugged due to comfyUI's most recent update. The workflow functions fine but it may not remember your toggled ON/OFF settings upon restarting. I will remove this message when they've fixed it. But for now just be diligent about making sure the options you want are toggled on or off correctly. (I hate you comfyUI for doing this right as I released my workflow.)
v2.0 Adds Ref Mods, a highly organized Ref/Media Manager, and makes everything smoother and easier to use. Make sure to update the latent upscale file to the new one.
This advanced filmmaker's workflow cleanly and quickly generates 3 low resolution sample preview gens, allowing you to select your favorite to then take to upscale finalization. The workflow has every feature an AI filmmaker could need, and easily produce the highest quality and resolution Minimax is capable of. It comes equipped with all quality of life features, the top known speed ups, and many options, from seamless video continuation to custom audio, to frame interpolation to go from 24 FPS to 48-60. Everything is here!
v2.0 Usage Guide:
Shortfilm entirely made using this workflow:
Some nodes in this workflow are too new to be in the Comfy Registry to be installed automatically with Comfy-Manager! Therefore, they need to be installed manually.
.: Latent Upscaler
- Node: https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler
- NEW Latent Upscale Model: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/blob/main/minimax_h3_latent_upscaler_3d_conv_v1/minimax_h3_latent_upscaler_3d_conv_v1_fp16.safetensors
(put into /latent_upscale_models/ folder and, since 9/17, this needs to be reselected in the WF's selection for latent upscaler choice due to the author renaming the file)
.: Prompt IDE
https://github.com/ethanfel/ComfyUI-H3-Prompt-IDE
.: Resources Loader (MiniMaxH3-PromptBuilder)
https://github.com/Adudeguyman/ComfyUI-Fantastic-MiniMaxH3-PromptBuilder
.: RefMods
https://github.com/Luisacaotica/ComfyUI-MiniMaxH3Mod
.: Turbo LoRA
This particular Turbo LoRA is my favorite. I recommend 6-8 steps if at 1.0 strength, or 8-10 steps if using at 0.6-0.8 strength. If you're upscaling, 6 first pass + 3 upscale pass works well.
.: Kijai's int8 Video VAE (much faster than default!)
Everything else should be easily available from the ComfyUI-manager Install Missing Nodes screen!
.: TaeH3 (for Kijai's Model Preview Override):
https://huggingface.co/Kijai/MiniMax-H3-TAE/blob/main/vae_approx/taeh3.safetensors
(note, using Kijai's Model Previewer will slow gens down by 6-16% thereabouts, I have it disabled by default now, but you can turn it on to test for yourself).
If you support what I'm doing, please consider following or subscribing to my Patreon! It's free (if you want it to be) and all my content and updates go there!
Description
Adds Ref Mods (by popular demand)
New Resource/Media Manager
Cleaner, easier to use, more organized
Removes several custom nodes that are no longer needed
FAQ
Comments (70)
Great update as always, question - I see with the new seamless method we need to use the scissors - but there's no way to just write in the cut # of frames and we gotta use seconds - do you know how many seconds I have to shave off if I have the overlap set to 5 frames?
You don't need to think about that kind of thing at all. Just select a chunk of your video making sure the rightmost blue trim handle bar is where you want your new gen to start, and set your overlap frames accordingly. Note that the overlap frames come from the LEFT side of the video's end trim bar selection, not after it. So mathematically, if your gen should start at frame 250, and you put your end selection there, with an overlap frame value of 5, it grabs frames 245-250 as the overlap, not 250-255.
Yes! Thank you!
I'm so excited to test 2.0, this looks so well put together!
It's a good one!
Seems like 15 second upscaling is not possible on my hardware. 16gb vram (4080 super) and 64GB RAM. It just stays stuck on model initialization and never continues. Are there any workflow options I can do?
I would love to increase the chunks from 3 to like 9, but there doesn't seem to be an option.
[INFO] SELECTED: input1
[MinimaxH3-3D] Latent 30x54 -> 48x86 | Pixels 768x1376 | scale=1.590
[MinimaxH3-3D] Loaded upscale model: minimax_h3_latent_upscaler_3d_conv_v1_fp16.safetensors
Params: 345,280,216 | Attn: forced off | Temporal: on (every=2, kernel=5) | Backend: CUDA 12.6 | Precision: fp16
[MinimaxH3-3D] temporal chunking: T=87 chunks=3 overlap=5
[MinimaxH3-3D] ✅ Model offloaded to CPU. VRAM released.
[INFO] Requested to load MiniMaxH3
[INFO] 0 models unloaded.
[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19995MB Staged. 0 patches attached. Force pre-loaded 210 weights: 1175 KB.
0%| | 0/6 [00:00<?, ?it/s, Model Initializing ... ]
Yeah, of course it depends on how much you're upscaling. You'd likely have no problem at 1.0-1.2 MP, but once you start going up to 1.5-2.0 (1080p) the lack of available VRAM can spike and cause comfy to lock up. Some people said they have success with the "force unload" option in the upscaler settings. But for me, I prefer to just run 1.5 MP as a single pass, the quality is better than any 2.0 MP upscale.
You shouldn't have any issues with that. I have a 4080 and a 5090 and on the 4080 I've upscaled 1 minute videos in under two minutes while using my system without any latency.
I've got a 16GB card and am trying to upscale to only 1.2 megapixels just to test, and it is really bogging down. Actually it just finished at 8:30 (127.53s/it). I think that's just how it is. I had 30 GB of system RAM free when it was generating. Maybe someday they'll update the upscaler or comfyui will fix it.
I see the new 2.0 version has the option to apply ref mods.. but is theres a option to create new refmods like packing all those references in the media loader to a new refmod safetensor?
Yeah but that's all different workflows, I'm not going to make my gen workflow have options to create ref mods. this is just where you gen once you create them.
@foxydits ok thanks... btw I think i encountered a problem.. i did a seedhunting , selected a preview and hit the upscale and took a break and came back and the final video is one of the cropped 2 second refrernce videos, but clearly the modal preview sampler override node shows the whole 8 sec upscaled video, I dont know where the upscaled video went.. not in temp , not in outputs.. console shows that the prompt executed successfully like following...
[MiniMaxH3 media_io] crop 2560x1440 -> 565x326 at (1387,875)
[MiniMaxH3 Loader] 6 item(s) in state -> 4 picture(s), 2 video(s), 1 soundtrack(s), 0 standalone audio
[MiniMaxH3 Loader] video_audio_1: [1, 2, 100800]@48000Hz rms=0.0651
[INFO] Prompt executed in 00:18:14
[Update]
I managed to extract the upscaled video by plugging in a video combine node directly into the audio and video decode node behind the 2nd pass node.. i thinks theres might be some problem with the video selection part at the bottom most part of the workflow's canvas.. I only have the comy kitchen and model preview selected.. But i did tried to check native output spliter in the new media loader node, may be that caused the problem...
Hello. Thanks for the version 2.0. Unfortunately, this version doesn't keep chosen main options - they reset back to default after any change of ComfyUI open tab (or restarting Comfy or reloading page). :(
same here
same here
Is it this version or did rgthree just shit the bed? Because I'm noticing some toggle issues as well. I assumed it was comfy's most recent update not playing well with rgthree.
Amazing. Firts WF, where is perfect lipsync. Maybe you can make WF with only single pass, without Seed hunt and uspscales? For users, who want to make videos faster? CS - perfect, audio - perfect. but seed hunt and upscaling making work too complicated... But anyway - TY for your work. Lipsync work.
This workflow can be used as single pass. It's an option with clear details in the workflow. Toggle Single Pass on, disable previews #2 and #3, and run the workflow. Boom, single pass & no more seed hunting.
@foxydits Yes, i see it, just mention that for 'maybe you make simply wf' only with single pass. And for me second pass dont work - shows errors...
excellent work. thatnk you for this. wish there was an option to auto save last frame of a video so i can just put it in for the next scenes first frame.
Yes, that would be a useful addition. DaSiWa’s Enhanced Video Combine already includes that. I’m not very familiar with ComfyUI, so in V1.6 I was only able to use the node on the 3 low-resolution sample previews, not on the Final Video.
I can see how that would be a useful option for people who don't always have a video editor running. I will add that to the list of things to toss in since it's a fairly simple addition.
thanks for the update
Thank you for the update. your workflow is the only workflow i have been using for a while now. no offense to others, but none can even come close. well organized and fast.
I'm using a local LLM to generate my prompts. up until this version, i used to be able to hook up the images, audio files and video files directly into the LLM. i use set nodes for them, and then set the get node to feed the LLM. but now with version 2, I'm unable to do that, since the files are inside the media loader. is there anyway i can get that functionality back? anyway to let the LLM node access the media files?
Which ComfyUI node are you using to feed images, audio, and video directly into your local LLM? I’m currently using ComfyUI-LLM-text-processor, but it only lets me feed images.
Also, do you have a workflow you’d recommend for generating MiniMax H3 prompts with a local LLM?
Hi @latensaseeds896 -- so that's an interesting use case I didn't consider. If you move the Media Loader and Prompt IDE nodes out of the way, you'll see the reference connections node. Click the little left-hand window expand button to see all of the connections. You can absolutely still wire your LLM to receive the media data. I just hid all of those connections underneath the nodes to prevent so much wiring crap on peoples' screens.
@foxydits ah, that works. thanks! the media loader is convenient, but it's less flexible than before. for example, i vibe coded a feature into the load audio UI node that allows me to save the duration and seek position of a voice file. then later, i can just load it from a drop down list in the node itself. the file would load at that exact seek point and duration. i don't think the media loader allows that. there is no way to load a saved file. is it possible to have V2 but without the media loader? i can see how the media loader is more streamlined, less clutter. but, the old system with load audio\video UI nodes and the load crop image was simple and intuitive. just drag and drop your files like any other workflow.
@bamib1971533 i use OllamaGemini. it has a node called Ollama API. i use it with ollama running locally. the node allows for 5 images, audio, and video input. just download the largest gemma4 model your pc can run. feed it the minimax prompting guide as the system prompt. connect your images, audio or video file. keep in mind that the model can see one media type at a time. so no image+video or audio at the same time. also, as far as i know, the only modern LLM that can listen to audio is the gamma 12b.
@latensaseeds896 I get you. Well, why don't you just simply add the nodes you like to the workflow and plug their outputs into the h3 reference slots, replacing those specific ones from the media loader? That way you can use the media loader for things like pictures, but any audio like how you would have it set up would come from your own nodes. It's not like the media loader forces you to use everything or nothing at all. In the end, each link ends up connecting to the same place; the main h3 node, and you can mix and match as you please.
@foxydits oh, i can do that? i assumed there is some kind of underlying connections under the hood. i'll try that. thank you for your awesome work! EDIT: i just checked, i think it's more than just connecting the nodes. there are other connections for other things like audio latent and ref video and first\last frame etc.. i''l have to deal with the media loader. let me just say thank you for your great work. just looking at those lines and connections gave me a headache.
@latensaseeds896 Yea there's a few more things going on, sorry to not say so earlier, but if you don't use that functionality it shouldn't be harmful to disconnect it. I just never know how comfortable people are in comfyUI to be able to make changes like that themselves.
@foxydits really appreciate the engagement. i just want to leave my feedback, for whatever it's worth. i do not like the media loader. let alone the usability of the node itself, like the seek bar behavior for example. i think the principal is limiting. for example, for video extension, i use a custom LLM system prompt that automatically looks at the video, it can look at the exact second and duration from the load video UI node and generate the correct prompt for me. unable to do that with the media loader. that's just an example. there are many other scenarios where the media loader would be an obstacle. like, say i wanted to run an image through an edit? say sharpen or black and white then send as reference? another practical example, i run my audio file through the MelBandRoFormer filter to remove background noise and music and isolate the vocals. i used to do it on the fly after the load UI node. you are limited to doing whatever you need inside the media loader node. it just cuts the flow of the process. sorry for blabbering. i hope you accept the feedback. much love.
This workflow was great while testing, but I've had some problems this morning when trying to actually use it. There is something going on with the media loader node. I'll try to load images into it and it turns them into just solid black frames. They will not delete. I have to restart comfyui. Videos are even worse. But if I manage to get that to work and I go to upscale, it takes 10 minutes per step (606.20s/it) just to do a 1MP upscale. I went back and tried the old workflow. Same settings, same image, same prompt, same everything, whole upscale completed in less than three minutes.
Twice in a row I have generated three .3MP seedhunter videos. Their previews show up. I select the best one, turn on the upscaler section, and send it through. Over six minutes later it spits out the reference video for some reason. Last night and earlier today I successfully upscaled a few videos and it was painfully slow (much slower than your old workflows), but they only had images for reference. No video.
I don't really have a good suggestion for you on this one, I'll be honest. The idea that the final upscale pass spits out the reference video instead of upscaling a latent you select suggests something is terribly wrong or misconfigured but I'm not sure what that could be, since the symptoms you're experiencing are so outside of the realm of possibility as far as how the workflow is wired. It's not like I recreated the workflow from scratch-- the 1st pass sampling to 2nd pass upscaling is all untouched and exactly the same as it was in v1.6. Just not sure why the addition of the media loader is creating such a problem.
Could you try, for the sake of testing, seeing if the media loader doesn't have problems in a different browser?
@foxydits I actually just tried it twice again with a fresh workflow and this time had no problems loading a reference image and a reference video, but the upscaler just passed through the reference video again. The file it spat out was slightly larger in terms of MB and 24 fps instead of 25. Same resolution (1088x704). I don't know. I must have done something to cause this if nobody else is having problems.
I have some stuff to do but will try a different browser later. I'm using Firefox.
@potatometer350 I have the same problem.. For now i've plugged in a video combine node to directly combine the decoded video and audio behind the 2nd pass subgraph and it does the job for now since im not doing the video continutaion and the stitching parts that happens after the 2nd pass are ones that messing up the upscaled video.
@potatometer350 are you trying to do V2V editing? If you're instructing for V2V and it's not doing anything, that's typically just a prompting problem tbh. I've tried to do character replacements (replace girl A in video with girl B or whatever) and the model didn't do anything to the video at all. But then I prompted it correctly and described the differences between the characters and suddenly the model actually lifted a finger. But it was all prompt related, not that the model isn't getting a proper video to work with.
@foxydits It's happening to me and I'm definetly not trying a v2v editing.. I tried following the wires down and I think the 'final video selection' rgthree switch is always getting 'any 1' input which is the raw video reference 1. this make the final upsceled video to never pass through this 'final video selection' node . im talking about th regular ref2v workflow where all extra options are diabaled except the 2 which is 'comfy kitchen' and 'modal preview override' options.
@sivanandravi90347 Are you on the most recent comfy front end release by any chance? ComfyUI did break rgthree nodes on it and I'm dealing with issues caused by their fuckup in other areas in the workflow too (such as toggles on/off not staying selected). It's a nightmare. I'm wondering if this is also affected.
My biggest confusion is, if you aren't doing any V2V editing and you're not doing video continuation, why do you have a video loaded into video 1 slot? To be honest, I never really considered that anyone would do that. I'm just curious what your use case for it is.
@foxydits 2 sec of video 1 as one of the regulr video reference for how the character runs. not only that ... i have two more videos to guide the modal on how some motion works... I have been doing like this since minimax released. Isnt this how the minimax reference works.. pictures and video reference to guide the generation.? I'm confused right now.. Btw, yes Im on the latest comfy version.. I'm not good in this... will switching the inputs wires any_1 and any_2 for 'final video selection' node, solve this problem for me?
@sivanandravi90347 That's how I was using the video reference, too. I'm trying a baseball scene, referencing the throwing mechanics of a particular pitcher.
@foxydits so this is a comfyui issue? I've only been into this AI stuff a few months and have never had an update break a workflow, believe it or not. How long does it usually take them to fix it?
@sivanandravi90347 I have confirmed your case that if you try to use video references without it being a video continuation (using the option in the workflow), you do indeed get your reference spit out in finalization. But yes, I was mostly curious about your use-case because I was trying to ascertain that you weren't just doing a round-about way of video continuation. Sadly, the way this workflow is currently wired, use of video references for anything but the Video Continuation feature isn't functional. As mentioned, I didn't really think to test video references since I never use them myself, so I just figured "hey, vid continuation is working, regular genning works, so everything works right?"... I can fix it but it will take time that I do not have today, unfortunately.
Temporarily, removing 'any 1' from in the Final Video Selection anyswitch node should solve the issue, as that's where the ref is coming from. You won't be able to use Video Continuation without plugging it back in though, but it sounds like you need the other functionality more.
@potatometer350 This particular issue is a wiring issue. A lack of testing on my part to realize that people wouldn't be able to use video references outside of Video Continuation. I will fix it, and for now you can just disconnect the 'any 1' link that connects from the Vid Continuation group to the Final Video Selection switch to fix it (I believe). However I am busy all day today and won't be able to fix it for everyone.
@foxydits Ah that worked. And it did the 1MP upscale in only 176 seconds.
And you were right about the media loader. That seems to be a Firefox issue because it works a lot better in Brave.
@potatometer350 Sadly, comfyUI has many such problems with firefox. I've also had to adopt using chromium-based browsers for comfy.
Heya, reporting 2 issues. 1: The Set Guide node is missing the video vae connection (Edit: My bad for not having Everything Everywhere installed!) 2. The toggle for the Force Audio option is mislabeled (I see its duplicated from one of its neighbors)
Set Guide gets Video VAE from 'Anything Everywhere' -- a node that transmits certain connections wirelessly across the WF to applicable input slots. This is intentional and not a problem.
As for your Force Audio option being mislabeled... let me check with a fresh copy of the WF and report back.
Edit: Ahh, okay, it's mislabeled on the node but not in a way that affects usability. Thank you for mentioning it, I'll update the WF.
@foxydits Ahhhh yeah, crap, I didn't think of that - Sorry!
@altoiddealer All good, I appreciate you reporting the mislabeled node. That would have confused someone for sure.
Thanks for the workflow:
Just one minor issue, your notes suggest using Latent Upscale Model: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/blob/main/minimax_h3_latent_upscaler_3d_bf16.safetensors
But you have a different model name as default in the latent upscaler node. Which one do you recommend?
Apologies, it appears I only updated that link here in the civitAI page. It's supposed to be:
Thank you for bringing this up. I will update the WF.
Thanks for the workflow and update, it has become my go-to workflow now.
I'd like to mention a small change I made: I noticed that the "[Finalization] LoRA (on Upscale Pass Only)" node overlapped with the LoRA node used in the First Pass, so I connected the "[Finalization] LoRA (on Upscale Pass Only)" model input directly to the "[Attention] Comfy-Kitchen" node instead. Is this modification feasible?
Hi there, here's the logic currently programmed into the WF:
Anything in the main Power LoRA loader gets applied to both 1st stage and 2nd stage upscale passes. Anything in the "Upscale Only LoRA" only applies to 2nd stage. You are right in mentioning that if you have a turbo lora loaded into both, the upscale stage will get it double applied. There is no current way to only apply a LoRA to stage 1 but not stage 2, since the LoRA loader on the main part of the workflow applies to every stage.
What is the difference between the new minimax_h3_latent_upscaler_3d_conv_v1 vs the old one? Faster and/or less VRAM usage?
I don't know the under-the-hood details, but I have noticed it made upscaling look a lot better. 3-4 steps with turbo lora upscale now is a lot sharper than it was with the old version.
@foxydits Thanks. I use 12 steps + 8 upscale (no turbo) with the previous workflow. Is that good or do you recommend something else (no turbo)?
@teiji25762 That should be good. Personally though I'd rather just do 16-20 steps as a single pass at a higher MP to start. You'll always get the best quality from using the Single Pass option.
Total Commander's file compare says that files are identical, so I believe it really is just "Organize checkpoints into versioned release folder" as it says @ HF.
@nonemo13 I see. I guess it was probably just the node's update itself that fixed some of the issues, not the upscale model itself. Good to know, thank you.
they are same as far as i can tell.
Hey gingerbeard :) was happy to see this. Let's be frank, when you make kick ass stuff, there'll always be haters. Keep rockin', F them. I have checked out a shit load and I go back to seed hunter, why? Because it produces quality. Hands down. Slainte!
Really excited you added Ref Mods and the Fantastic mods, choice move. So thank you!
Curious, will you be making a quick video for this as well?
Hi hi! I do indeed have a gingery beard!
Anyway, not sure what you mean by haters, the reception has been pretty good so far as far as youtube comments / reviews here. Outside of the 'X isn't working' types of things, of course.
Anyway, regarding a 'quick' video... I dunno if that's possible. But I did make a long ass tutorial video, which is here on the workflow's civit page and on my youtube.
@foxydits from sd. The point is, your work IS stellar and credit is due when due. Just genuinely cheering ya on man. About your vid, oh I've watched them, was wondering if you made a new one that shows usage for the Ref Mods and the Fantastic mod.
@foxydits Nevermind, I see it! Watching it now. Thank you! :)
I'm stumped, i am getting a Cannot read properties of undefined (reading 'workflow') error whenever i try to generate. I "fixed" this in earlier version by simply using an existing video on this site with metadata to drop it into comfy and it worked from then on... what is this error?
So 5 seconds into the future: there is a "This checkpoint includes a config file, download and place it along side the checkpoint." right under the big blue download button... download from there and no issue. I don't understand since its the file with the same name i get by pressing the big download button. Anyway, i'll leave this for anyone with same issue (noob ;_;)
Tested this with a few images and worked wonders! I'll try some video extensions in the future but I mainly wanted it for multiple images and worked like a charm.
Defn recommend this person's works. Top notch and always worked for me!
Thanks for the update!
After some tests with your recommended settings, I'm unfortunately getting pretty bad results from the upscale pass with Turbo LoRA enabled. It looks like the LoRA isn't being applied correctly even though I've set the right path on my end and made sure to watch out for the rgthree bug.
Upscale pass without Turbo LoRA works fine at 2 MP. With Turbo LoRA it's a mess, even at 2 MP so something is definitely off there. Could be my fault, not sure. Either way, the workflow works fine without Turbo LoRA on the upscale pass, so I'll just use it like that for now.
Single pass mode is looking pretty good at 1.5–2 MP with Turbo LoRA on the first pass though.
One thing I had to change on my end (might just be my setup - PyTorch 2.12.1+cu130, same issue in v1.5):
Upscale node: device CUDA + precision bf16 - otherwise it throws an error.
EDIT: If someone else got that error. simply serach for that input and change that to max. Done.
[solved by myself, sorry.]
Hey there. So I just installed everything needed. Double checked. I am trying to generate I2V but id instantly gives the following error:
The value C for MiniMax H3 Reference to Video's ref_image_size is not available.
The image is loaded via the managerinside the workflow.
What am I doing wrong?
Cheers and thanks for the hard work you put in.
insanely good WF! You can leave nearly everything on standard and get awesome results! Thank you so much for this!