Hey guys, LD here, Spent months working on this. otherwise it would be free.
you don't just buy this node you buy everything i make, For future video gen models too.
i'm also in sulphur discord. if you need a hand.
13/10/2026 - Ref swap BETA - Been refining ref swap, its mostly automated. theres a guide on the side of node to help. its in ref tab. add a video first.
manual + other links on the website. - H3 Studio 2 – interactive tour by Lora-Daddy - new website layout. - NEW CONTENT ADDED 22:45 GMT 10/10/2026 - FX - Cinema - Many more camera and
12/10/2026 New: H3 Studio CHAT - It can control the entire node. And even See your Ref bank
Tell the chat your idea and it writes an intent you can paste in with one click. It also gives you click-to-apply chips for Ref Bank cards, dial settings and clip length (up to 20 seconds). It doesn't write exact lines of dialogue. It says who speaks and roughly what, and leaves the real lines to GENERATE. If you type an exact line yourself, it keeps it word for word.
It knows every dial: 16 dials and 2,232 options by name, including:
1,052 characters (391 animated from 118 shows, plus 661 from films, TV and real life)
570 TV shows and films
129 places
110 cinema looks
75 camera moves
49 accents, each with 3 strength levels
49 music options
46 styles
43 ambience options
41 FX
36 talking styles
24 colour grades
Name Elsa or Frozen and it suggests the right dial straight away.
It knows your Ref Bank: every saved card by name and kind (person, object, motion, background and so on), plus whether it has a voice or clip. Cards marked Private stay hidden when the bank is locked. It only suggests a card when the full name or first name matches exactly, so "Elsa" never brings up "Elisa". One click puts the card on your rail, and anyone without a card goes to the character dials instead.
Also new: a CLEAR button on the intent box (UNDO brings the text back). The PEOPLE, TV/FILM and ANIMATION dropdowns now scroll all the way to Z. And multishot now stops and tells you why if it can't use your reference images, instead of quietly rendering without them.
H3 Studio 2 – interactive tour by Lora-Daddy - new website lay out
Thanks guys.
updated weekly.
workflow is a placeholder you do need the node pack unfortunately.
Description
Refinement to the instruct
new wardrobe
CHAT + more
FAQ
Comments (6)
This is exceptionally well done and even at the paid price a bargain. I'm still learning to use the tool. It would do to have a better instruction set. I'm using Claude to build an instruction set for my use.
apricated I'll try and make a use pdf for it.
Hello Lora Daddy, I'm getting frustrated and wonder if you might give advice on best practices.
I have an rtx5090 and 192gb of ram(Purchased over a year ago before prices went insane)
I have 2 issues.
The prompt writer keeps running out of tokens. I feel like it's being more verbose than it has to be. It only succeeds to write a complete prompt on the shortest of clips. I have increased the tokens, and it seems like it helps a little, but not as much as I would have thought, and too high seems to make it go really slowly. Using LM Studio with the same model you had in your example clip.
My other issue is that the people in the videos don't look like my reference images at all. They are reminiscent of them in body shape and overall vibe. Like a slender woman is a slender woman, a busty woman is a busty woman, but they are not the same person. This is true in ref2v, i2v, multishot and the music video.
I'm certain I'm just doing something wrong. Any advice on best practices would be most appreciated because I see the huge potential here.
This is the first feedback i've had like this, What model are you using, im using qwen3.8-27b-uncensored-hauhaucs-aggressive-mtp
is thinking off in studio? -
it sounds like its not using vison at all.
i too have a 5090 with only 64gb of ram so it wont be that. something seems off. you could always try ninfer?
I'm having a similar issue, but with multishot. In the Ref panel, references work fine.
[H3Master] motion_ref is unused and stays at 0 / 1.00
[H3Master] pack: multishot 896x512 @ 141f (~5.88s) x 3 shot(s)
[H3Master] T2V - no start image; the latent supplies geometry only
[H3Master] MV stills — none on pack or Studio (clip 2+ will not switch photos)
[H3Master] reference stills: 1 at 1024 (from Studio's rail)
[H3Master] staged refs — 1 still(s), 0 clip(s), 1 audio — R2V on every clip (Picture N / Video N / Audio N)
[H3Master] stitch: 3 clip(s) x 5.88s delivered (413f / 17.21s total), carry 39 via official Motion Context → 1 video
[H3Master] 3 clip(s), 896x512 @ 141f, carry 39
[H3Master] stills sized to a 1024 short side: 1024x928
[H3Master] reference conditioning unavailable (_C_flashattention unavailable); text encode only
I've got the old PromptMasterLD running fine in a separate Comfy instance.
I also noticed the sla, RAM optimize and cache nodes aren't wired into the h3 master node in the recent workflows.
Hi, Sorry. This guy drmattiske did a much better job of troubleshooting than me. I am also having the problem he is describing.
I don't know why the prompting wasn't working correctly initially. It was working a little, but terribly and erroring very soon. I'm ashamed to say that a simple reboot of the computer resolved this issue.
However, I'm also having the same problem as drmattiske reports. In the Conditioning subgraph there is the "Group Switch - LD" node. I'm guessing that when you go through the different studio modes that this is supposed to change the things the workflow is doing in the subgraph. On my machine this isn't working.
And I've seen this exact error as well. Which makes sense, if it's not switching between modes I guess it's defaulting to text.
[H3Master] reference conditioning unavailable (_C_flashattention unavailable); text encode only
If I go into the subgraph and manually switch the modes and also manually remove the bypass from the ref group, it works for 1 clip. But for multishot the next clip reverts back to text only.