Stable Diffusion @ CloudDock
Your waifu’s favorite render farm.
Sweet Japanese vibes, zero dependency headaches. You draw, we bully VRAM and watchdogs for you.
- CloudDock Universal Usagi 4.0
- Momonga 4.0 “night-shift” twin
- 3060 does 3090 jobs
- Built-in SD Training Center
- One-click Model Pull
- Token-gated, no shared shells
Usagi · Releases
Universal Usagi release line
One container line, many betas. Pick any guide — they all share the same core workflow.
- 4.1: Automatic queuing. Just rest, we'll take care of the rest..
- 4.0: SDXL training, App Store 3.0, faster A1111 boot.
Usagi 4.1 Beta 2 — Overview
4.1.1 lets you sleep. 4.1.2 lets you sleep better — one true queue, faster captioning, and AdamW8bit done right.
Usagi 4.1 Beta 1 — Overview
Queue it, sleep, wake up
Usagi 4.0 Beta 7 — Overview
A1111 in seconds. SDXL training built-in. App Store finally grows up.
Usagi 4.0 Beta 6 — Overview
Same tools, calmer workflow. The Dog thinks twice now.
Usagi 4.0 Beta 5 — Overview
One Launcher for everything. Click icons, not terminals.
Usagi 4.0 Beta 4 — Overview
Click and draw. If it crashes, it’s not on you.
Containers
Universal Usagi & Momonga
Two faces, one brain. Usagi runs your daily waifu workflow; Momonga takes the night shift and heavy jobs.
- Stable environment across A/B/C/D groups — same UI everywhere.
- Prewarmed venv, ControlNet, VAEs, extras ready out of the box.
- Per-container watchdog (CloudDock Intelligent Dog™) watching only for bad behavior, not your art.
Choosing Your Machine
Pick the right CloudDock GPU in 60 seconds — based on VRAM, model size, and what you’re actually doing (not your ego).
Your first image on CloudDock
From queue to first anime portrait in about a minute.
Switch between SD · Jupyter · Deepspeed
Launcher, not SSH: everything lives behind powered doors.
Training Center
CloudDock SD Training Center
LoRA & DreamBooth training baked into the container. SD 1.5 and SDXL scripts, loss curves, VRAM ladder — all on one board.
- Dual scripts: 1.5 & XL with shared UI, shared dataset layout.
- Smart Caption Beta: better captions than naked “auto caption”.
- Resume LoRA, .safetensors export, training logs one click away.
LoRA Training — Step-by-Step (Portrait)
Clean portrait LoRA from a single dataset. From folder layout to first test image.
DreamBooth Training — Full-body / Mascot
How to train a mascot or full-body waifu that actually repeats correctly.
Dataset & Captioning Guide
Folder structure, tag strategy, and caption rules that Training Center expects.
Tuning & Troubleshooting
VRAM ladder, batch size, learning-rate tweaks for A/B/C/D groups.
Control · Hires
ControlNet & Hires Fix, the non-mushy way
Pose, hands, face, depth — and then a two-stage upscale that doesn’t melt everything into soup.
- Layered pose / hand / face ControlNet presets tuned for anime.
- Hi-res ladder: 1× base → 1.3× detail → optional final polish.
- VRAM meter & “danger” hints inside the guide, not after you OOM.
Pose · Hand · Face presets
Hands strong, faces medium, body soft — avoid mannequin-fu.
Lineart + Depth
“Shape first, color later” pipeline that keeps composition clean.
Fixing Blurry Faces & Bad Hands
Three tweaks that clean up faces, fingers and edges without nuking style.
Two-stage hires ladder
1.3× first, low-noise repair second. VRAM-friendly for A-group.
Launcher · App Store
CloudDock Launcher Workflow
One glass panel to boot SD, JupyterLab, Deepspeed console and the Model Store. No SSH circus, no guessing which screen is “that” A100.
- Launcher buttons wired per container: SD, Training Center, App Store, Deepspeed.
- App Store 3.0: per-app page, screenshots, rating, download status that survives refresh.
- Hash-verified models only — no mystery checkpoints from random corners.
Model Store — Checkpoints & ControlNet
Install curated anime checkpoints, VAEs and ControlNet packs in one tap.
Using your own checkpoint & LoRAs
Where to drop files, how Launcher sees them, and how to keep paths tidy.
“From blank instance to finished set”
SD → ControlNet → hires fix → Training Center validation in one sitting.
Privacy · Safety
Your waifu stays in her own box
Single-tenant containers, token-gated reverse proxy, watchdog only cares about root tricks — not your pictures.
- Each instance = one user, one token, isolated filesystem.
- Normal shutdown wipes container and local logs by design.
- Abuse gets quarantined with logs; normal use is forgetful by default.
W.T.F. Lab · After-Sales Field Notes
Receipts-first stories & playbooks - by Usagi Team
W.T.F. stands for What The Frand (Warranty · Transit · Fine-print) — the three places where “support” quietly turns into a maze. We document what happens, how to respond, and how to keep your case defensible.
- No witch-hunts. No doxxing. Names get redacted.
- We publish what can be backed by timelines, screenshots, and policy text.
- Actionable templates: what to say, what not to say, and what to ask for in writing.
Want to contribute? Post on our X homepage (or DM the mods) with redacted receipts. If your case is selected, we’ll turn it into a clean, reusable playbook.
Why we publish W.T.F. Lab
Our rules, our purpose, and why SD creators deserve receipts-first after-sales transparency.
ASUS Fluxgate: when “residue” becomes a denial template
How a single phrase can be used to shut down RMAs — and how to respond with clean, written asks.
The 50% Mirage & the Label Swapt
a high-value GPU order that “delivers” to someone else.
Guides
Guides & playbooks
Short, opinionated guides instead of wiki soup. Start anywhere, they all assume you’re busy.
Back to top ↑
