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.
Reference bank sticks to Base model methods - Allows you to add 3 photos + and mp3 for voice. My website offers a free tool to make Reference sheets perfect for H3.
My H3 Encode node is all ref/i2v/t2v in one, all handled by the Main Studio,
20 reference images and 9 mp3 vs the 9/3 stock options.
Hover over anything in the UI to get Hints of what it does
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 styles. <3
NEW IN 2.1.0 -
- CHAT — talk to the writer. Show it a picture, ask for ideas or dialogue, and it gives you intents with a USE AS INTENT button. It sees your tab, your rail, your frame and length — and on STEP BY STEP, the real last frame of the clip you kept.
- SHOTS dial — auto, or 1 to 20 shots a clip. 1 is one held shot; more is more action and more camera. Every clip is one continuous take: no cuts unless you ask for one.
- SHOTS form — every shot is VISUAL / DIALOGUE / SOUND. Pick who speaks from the cast, type only the words; the speaker, voice and language tag are written for you.
- SPEECH — 36 ways of talking (roadman, pirate, news anchor…), on the main characters only.
- HINTS — small preferences the writer folds in where they fit.
Updated weekly.
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
FAQ
Comments (5)
Thank you for your hard work! The workflow looks simple to use, and the node you created is a masterpiece. Unfortunately, despite how polished everything looks, I’m not always getting the video I expect from the prompt.
Maybe the issue is with Bionic or the Qwen3.8 27B model I’m using, or perhaps I’m doing something wrong. I’m trying the REF mode to replace a character in a video while keeping the background, visuals, movements, and SFX unchanged, with everything continuing without interruptions. However, the generated video sometimes turns out differently from what I requested, even after I’ve tried adjusting the prompt. It still looks cool, just not quite what I asked for, hahaha.
Keep up the good work! I’m not saying it’s bad at all. I might just be doing something wrong.
yeah there's a big hurdle here some videos can transfer really well some don't - Try this, add the video via the bank instead. select it as a motion, then click describe. this turns the description into a breakdown of movements (the llm Vision scans every 0.5s) . - then save , add that to the rail with your person you want to swap this way makes it a lot more consistent. and faster as it doesn't call the llm if the descriptions are filled in for the reference bank items
this also allows for faster swapping - i will work on it more but the issue im see is a recipe for 1 video making it perfect fails on the next.
Question from a RTX 3090 user. I read your page and it mentioned a 12B class model on a 16GB card could work. Do you have a recommended 12B model and Q size? I assume there is no way to use LM Studio and ask it eject the model before moving on to H3? Thanks.
yes, its all automatic. when u click generate comfyui models are purged. lm studio or ninfer is called.. prompt is made. when u click run to make the video. the llm is purged so nothing is carried forwards.
for your card a 12b llm would work fine probably this or something a little extra. whatever is about 80% of your vram, + a little leeyway, - its never carried into video gen. HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced
@Cowgary JIT is the setting in LM you want to set for controlled load/unload.