STUUBZZZ
MiniMax H3 long videos in ComfyUI: retake one chunk, keep the rest (Ultimate H3 0.13)
Ultimate H3 0.13 renders long MiniMax H3 videos in ComfyUI in chunks. On an RTX 5090, retaking chunk 2 of a 12 s Speed video took 54 s, both chunks 83 to 121 s.
Self-hosted AI video · built and taught in ComfyUI
Your brand, rendered a thousand times — identically.
Custom AI image and video pipelines that lock a character, product or style, then reproduce it at production scale. Built on your infrastructure. Owned by you, not rented from a SaaS.
Neon street · LTX 2.5
● every shot rendered locally, in ComfyUIone RTX 5090
- ComfyUI
- LTX 2.5
- MiniMax H3
- Stubelius nodes
- LoRA training
- Character consistency
- Self-hosted
Two ways to work with the studio
For creators and solo builders
Learn it, or get un-stuck.
The same pipelines, explained from the first node. Start free and go as deep as you like.
For brands and studios
Have it built. Own it.
A character, product or style locked into a pipeline your team runs without me.
New · the ComfyUI course
ComfyUI From Zero
A three-part course that explains the machine instead of handing you another workflow to copy. Part 1 is the whole beginner course and it is free.
Free
Part 1 · Read it
Understand the machine
Install ComfyUI, build the seven-node text-to-image workflow from an empty canvas, and fix the eight errors beginners actually hit.
- What every node and every KSampler setting does
- Which model file fits your card, and where it goes
- The eight errors beginners hit, and the fix for each
Paid bundle
Part 2 · Change it
Open the hood
Change a downloaded workflow and know what will happen before you press Run.
- Fit tables for 8, 12, 16 and 24 GB cards
- LoRAs, the four ways to edit a picture, and video explained from the VAE up
- Structured prompting, chaining clips, renting a GPU
Paid bundle
Part 3 · Build it
Build the engine
Train your own LoRA, merge models, write a custom node and run your workflow from a script.
- A real LoRA trained on screen, from the first samples to the winning checkpoint
- LoRA formats, conversion and model merges
- A custom node in about twenty lines, and the API in twelve
See the full course
28 modules · just over 10 hours · slides and live ComfyUI demos
The problem
Two walls every brand hits with off-the-shelf AI
Wall 01 · Consistency
A different face on every render
Generic tools hand you a new character, product or style each time you press generate. Campaigns need the same one, across a hundred stills and every second of video.
The fixA trained consistency LoRA and reference system that locks your subject, so it stays itself from front, side and in motion.
Wall 02 · Ownership
Renting your own creative
Midjourney, Runway, Sora: per-seat fees forever, output caps, and your prompts, references and IP leaving your walls on someone else's servers.
The fixThe whole engine deployed on your hardware or private cloud. You run it, you own it, and the marginal cost of the next asset is near zero.
Flagship
The Brand Character System
The fastest route to a repeatable visual identity your team can run without me.
Fixed scope · fixed price
One subject. Infinite on-brand output.
from €4,500
- A character, mascot or spokesmodel, designed from scratch or adapted from your existing IP
- A trained consistency LoRA that locks the subject across stills and video
- A ready-to-run ComfyUI workflow, tuned to your look
- A brand prompt-assistant that turns plain briefs into correct prompts
- A handover session so your team ships without me
Outcome: unlimited on-brand stills and video, in-house, at near-zero marginal cost.
Same identity in every frame. How the dataset is built
Free tools
The workflows I build for myself, released for free
Patrons get them first, then they go public on GitHub with a walkthrough video. Everything in the reel above was rendered with these two workflows.
MiniMax H3Stubelius Ultimate H3
One MiniMax H3 workflow for ComfyUI. Speed, Hybrid and Quality modes, seed hunting with full videos and sound, LoRAs per chunk, 24 to 120 fps, and RTX VSR or DLSS 5 up to 4K.
Get it on GitHub
LTX 2.5Stubelius Ultimate LTX 2.5
The same idea for LTX 2.5: seed hunting with full videos, a full-size first pass, Licon MSR reference images in both passes, audio lock on the refine, and output up to 4K.
Get it on GitHub
LoRA · write-up2.5D style LoRA for MiniMax H3
A style LoRA that holds the 2.5D look MiniMax H3 tends to skip. The post covers how it was trained and which strength to use, and the newest version's training run is on screen in Part 3 of the course.
Read how it was made
Frames from the two workflows' release videos. Every one was rendered locally. Setup guides: Stubelius Ultimate H3 · Stubelius Ultimate LTX 2.5.
Who the builds are for
Teams that need the same thing, at scale
01
Game, anime & VTuber studios
Character sheets, turnarounds and real-time talking avatars, locked to your canon.
02
DTC & e-commerce brands
Endless on-brand product, lifestyle and ad creative without a photoshoot per campaign.
03
Performance & creative agencies
Ad variations at scale across clients. White-label the whole pipeline behind your name.
04
Virtual-influencer & talent agencies
The full character lifecycle, from bible and LoRA to launch and monetization.
05
EdTech & training
Multilingual avatar video and localized course content, produced at volume.
06
Localization & dubbing
Swap characters into reference footage with matched audio, across languages.
Ways to work together
From a single session to your whole pipeline
Start small and self-serve, commission a fixed-scope build, or bring me in ongoing.
Entry
Self-serve & 1:1
For creators and in-house teams getting unblocked fast.
ComfyUI From Zero new3-part course · Part 1 free see the course
Part 1 free
Stubelius workflow & prompt packscommercial license
from €149
1:1 working session60 min · live · ComfyUI / LTX / H3 book
€180
Commission a one-offLoRA · sheet · short movie · pack browse
from €50
Pipeline review90 min + written roadmap
€390
Cohort workshopper seat
from €450
Core
Productized builds
Fixed scope, fixed price, yours to keep and run.
Brand Character System flagshipcharacter · LoRA · workflow · handover
from €4,500
Content Engine Sprint2–3 wks · self-hosted stack + team trained
from €9,000
Custom node & tooling buildbespoke ComfyUI nodes / workflows
from €2,500
Deep
Ongoing
For volume, or when you'd rather it just run.
Production retainerdone-for-you monthly output
from €2,800/mo
Fractional Head of AI Mediaembedded build + train
from €5,500/mo
Agency white-labelyour clients, my pipeline
partner terms
Why me
I build pipelines, not one-off prompts.
Custom ComfyUI nodes, LoRA training and diff-LoRA merges, and self-hosted video generation on the current open-source frontier: LTX 2.5 and MiniMax H3. I've taken two AI characters from bible to launch, and I run a paying community of builders on the same tools I'd hand your team.
And after twelve years in business development, I care whether the pipeline actually ships and pays back, not just whether the demo looks clever.
2AI characters taken from bible to launch
12ybusiness development: it ships and it closes
3free ComfyUI node packs published on GitHub
Localruns on your hardware, nothing leaves your walls
Four steps from brief to owned engine
How does a build work?
A build runs in four steps: a scope call, the build itself, a handover that trains your team to run the engine, and scaling up when you want more volume.
- 01
Scope call
We define the subject to lock, the output you need, and where it runs.
- 02
Build
Character design, LoRA training, and the workflow and prompt-assistant around it.
- 03
Handover
Your team is trained to run and extend the engine on its own.
- 04
Scale
Move to a retainer or white-label when you want more volume.
From the blog
Notes from the pipeline
Settings, benchmarks and teardowns from the same LTX 2.5 and MiniMax H3 builds that go into client work.
FLUX 3 Klein vs Qwen-Image 2.1 in ComfyUI: 428 renders on an RTX 5090, round by round
In 428 renders on an RTX 5090, prompts as typed, Qwen-Image 2.1 won 7 of 23 image types to FLUX 3 Klein's 4 (12 ties). Klein won short text, Qwen editing.
Qwen-Image 2.1 Turbo in ComfyUI: Qwen's 8-step schedule, settings and speed in 300 renders
Qwen-Image 2.1 Turbo made a native 2K image in 11.8 s vs 36.9 s for the 30-step base model on an RTX 5090, and needs Qwen's 8-step sigma schedule in ComfyUI.
Start here
Let's lock your visual identity.
Tell me the subject you need to reproduce and where it should run. I'll come back with scope and a fixed price.