One take, 40 seconds, no joins. This free ComfyUI workflow makes one unbroken MiniMax H3 take from one start picture and one prompt, with voice and sound. Long H3 videos are usually short clips joined together, and the face and light jump at every join. This one makes the whole 40 seconds in one pass.
What is in the zip
- The workflow, with notes in every group
- The start picture and the prompt from the video
- A QUICKSTART with every model file and where it goes
Settings that work
- 736 x 416: the size that kept the face steadiest when we made the same take at five sizes
- 957 frames for 40 seconds (the note has a table for 5 to 40 seconds)
- About 2.5 spoken words a second, so 40 seconds needs about 100 words. Fill the whole length.
Sharp at the end. An optional group upscales the take to 1080p with the free SeedVR2 nodes and keeps the same face. Switch it off if you are short on video memory.
Needs: ComfyUI with MiniMax H3 support. The take uses core nodes only. The 40 second take took about 26 minutes on our machine.
More MiniMax H3 workflows and the full write-up: free H3 long shot guide
Description
FAQ
Comments (13)
What I've observed so far with h3, is that often, unless you describe the reference very accurately down to minutiae, and are adamant about nothing changing, in quite a few cases (and sometimes despite all efforts to the contrary) it will simply jumpcut straight away from the initial image to whatever the model thinks fits the prompt, with a vague resemblance to the source. Seemingly even moreso, when any loras are introduced.
Have you been able to get better consistency without painstaking prompt refinement, like a lora or specific settings? Besides the size mentioned - that solution in particular is a bit involved as well. At a glance, I haven't exactly seen anything magical about this wf for that.
Give a prompt youve tried, and what resources you used (images, audio, video)
Sounds like you arent prompting correctly and or using ref2va mode which defaults to selecting the image as a reference not a start frame.
@pumaai487 I took the i2v template as basis, I have seen how the r2v template is connected and have definitely not ventured into that territory.
I think myself a fairly mediocre prompter, I'll give you that, but assuming you are prompting correctly here in this example: https://civitai.red/images/145114598 ,
Why do you have to specify "preserving her dark twintails, mini top hat, choker, and black-and-pink layered dress. The camera remains completely static with locked framing, maintaining a continuous single take without cuts." ? Why not just "preserve the same character, hair, clothes [and background]" generically which would fit any image? Or for that matter it being implied that the character you start with is the same character you end with. Contrast it with wan or ltx prompting, I don't think I ever had to do that.
That's the sort of thing I'm talking about, that some things need to be specified precisely or they're gonna jump out of existence / morph into something else (especially monstrosities). I've seen it just about everywhere here, not just in my own prompts.
And "she shifts into alternating arm extensions, tucking one hand near her collar while punching the other arm sideways in sync with snappy hip bounces.", "she brings both hands down to frame her chin in a "flower hand" pose, tilting her head playfully with a final hip bounce" etc - tell me that doesn't feel to you like taking on a choreographer's day job.
Me, I'm a lazy programmer, if I can find or build a program to automate my job away, I'll take it every time (as long as I'm still getting paid the same salary, that is). I'd like to believe there's some part of the nodes or weights that tells the model how much it is allowed to deviate from the starting image, and it is turned up too high (to allow more free movement / alteration at the expense of stability).
@firemanbrakeneck Try foxydits or promptmastery on yt for their workflows and their example prompts, might work better.
In that vid I just used the prompt my llm gave me, using the minimaxh3 prompting guide but there is like 10 other prompting guides from other people that i occasionally use. And all the video models prompt differently, minimax is probably the most specific, ltx below that, and wan, you can say the girl does this dance.
Haven't seen any of your vids or prompts so can't really give any further advice about "jumping out of existence / morph into something else (especially monstrosities)". I haven't really seen that.
Thanks for the detailed write-up, and thanks @pumaai487 for jumping in. You’re right that H3 wants more detail in the prompt than Wan or LTX. With a loose prompt it can drift from the start image or cut away, and in my tests LoRAs can make that worse. I haven’t found a single setting that controls how far it’s allowed to move away from the first frame. If I do, I’ll share it here. For keeping the same person, the thing that helped me most was giving H3 the face as a reference instead of only a start frame. That’s what this RefMod workflow does: same face in every clip, no LoRA. It won’t fix every drift, but it means less describing the character down to the last detail.
Nice WF thanks. On a 5090 took about 9:30 to gen the 40 video and was another 11 min in to upscaling when i stopped it (my prompt wasn't very good).
Thanks for trying it, and nice that the 5090 got a full 40 second take out in 9:30. The upscale pass is the slow part and I have not timed it on a 5090, so I can’t promise a number there. On the prompt: the shipped one describes the example picture, so if you change the start picture, rewrite the prompt to describe your own picture first, then the action. That usually helps most. If you try again, I’d like to hear how it goes.
@Stable_Yogi i think it need the preview panel (i'm not very good with Cumfyui) so we can abort if it doesn't look right. I added a turbo lora 8 step and got the gen down a lot. As a matter of interest, you got 40 sec one take, what's stopping 60 second.. Gen time?
It doesn't matter which starting image is chosen: the workflow only interprets the prompt. As a test, I chose a picture of a cat; the prompt read “a 23-year-old woman, an adult, with long, wavy, light-brown hair,” and the video showed a 23-year-old woman, an adult, with long, wavy, light-brown hair - no cat at all!
? what kind of prompt is that? lol
@pumaai487 The one this wf was published with. That's the sort of thing I'm talking about, the starting image is taken as a recommendation at best, if it doesn't fit the prompt as the model understands it is ignored.
Thanks for testing it! The start picture does go in as the first frame, it isn’t ignored. But the prompt that comes with the workflow describes the example picture (the woman by the pool). If you load a cat and keep that prompt, the picture and the prompt disagree, and H3 tends to follow the prompt and move to what it describes. Try writing a prompt that describes your own picture (the cat, where it is, what it does) and it should stay with it. If you still lose the cat with a matching prompt, please share the prompt and I’ll take a look. And if you check the very first frame of your video, you should see the cat there before it changes.
You need detailed promps for H3, this isn't Stable Diffusion era models anymore. 99% of people's generation issues has been how they're prompting.
Use the Qwen-VL nodes available for comfy UI (no idea why people are using convoluted LLM methods) and use its preset for H3 or whatever model you're using. It'll handle the hard stuff for you. You also have the option of adding more details and what not too.
But prompting, "Blonde, high detail, 24yo, girl,pool" Is about as useful to H3 as it is to any of us. You're not asking it to do anything other than to render some random shit.