amer21/blender-kaggle-hub

🎨 Blender Desktop on Kaggle — from your Android phone

Run the full Blender GUI on a Kaggle Linux notebook (Kaggle CPU + GPU do the work) and control it remotely from any Android phone — no PC needed.

┌──────────────────┐   public https URL    ┌────────────────────────────────────────────┐
│  Android phone   │ ────────────────────▶ │  Kaggle notebook (Linux x86_64 + T4/P100)  │
│  Chrome / Edge   │   (Cloudflare quick   │                                            │
│  → noVNC web UI  │    tunnel, cloudflared)│  Blender GUI ──▶ Xvfb :99 (virtual X)     │
└──────────────────┘                       │        ▲                        │         │
                                           │        │ renders to             ▼         │
                                           │  Cycles GPU (OptiX/CUDA, N GPUs) x11vnc   │
                                           │                              (VNC :5901)  │
                                           │                                             │
                                           │  noVNC/websockify (:6080) ◀────────────────┘
                                           │  /kaggle/working  ← your .blend files + backups
                                           └────────────────────────────────────────────┘

How it works (and the honest trade-offs)

Part	Runs on	Notes
Blender viewport (what you see)	Xvfb + Mesa llvmpipe (CPU software GL)	Xvfb has no GPU access, so the viewport renders on CPU. Fine for modeling at 1600×900. Use Solid shading for best phone responsiveness.
Cycles rendering	GPU (OptiX → CUDA) — all detected GPUs	T4 (7.5) & P100 (6.0) both supported by official builds. 2× T4 → both enabled (Cycles multi-GPU). No SLI/vulkan-style combining for the viewport — only for Cycles/EEVEE compute.
Remote view	x11vnc → noVNC → cloudflared quick tunnel	Public random *.trycloudflare.com URL, no account, no exposed Kaggle ports. Session-lifetime only.
Persistence	/kaggle/working + optional Hugging Face mirror	Kaggle sessions are temporary; /kaggle/working becomes an output dataset, and the autosave loop can push every .blend to HF.

Security note: the tunnel URL is unguessable but anyone who has it can see your Blender. It dies with the session. If you need auth, put an nginx auth_basic or a WireGuard tunnel in front (out of scope here).

Quick start (Kaggle)

1. New Notebook → set GPU: T4 (optional but recommended) and Internet: On (required).
2. Paste notebook.ipynb → Cell 1 (or clone this repo into Kaggle and run).
3. First run takes ~3–6 min (apt packages + ~350 MB Blender download; cached for the session).
4. It prints a https://xxxx.trycloudflare.com URL → open it on your phone.
5. Run Cell 2 (component tests) and Cell 3 (GPU render smoke test).

The 3-line copy-paste (if you don't use the .ipynb)

!rm -rf /kaggle/working/blender-kaggle && git clone --depth 1 https://github.com/amerameryou1-blip/blender-kaggle-desktop.git /kaggle/working/blender-kaggle
!cd /kaggle/working/blender-kaggle && WORKDIR=/kaggle/working bash setup.sh && WORKDIR=/kaggle/working bash launch.sh
!cat /kaggle/working/blender-env/tunnel.url

Connecting from Android (Xiaomi Redmi Note 14 Pro 4G / any Android)

1. Chrome/Edge → paste the tunnel URL → VNC screen → Connect (no password by default).
2. Landscape + fullscreen (⤢) + tap once so input reaches Blender.
3. Chrome ⋮ → Add to Home screen for an app-like launcher icon.
4. Touch mapping (noVNC): one finger = left mouse drag (orbit/move). No mouse wheel — zoom with the viewport toolbar (magnifier +/−, ⌂ frame), pan with the 2D-pan arrows or SHIFT+middle-equivalent via keyboard layout. Menus: tap. For more precision use the noVNC toolbar (top-right of the page).

What's in the repo

File	Purpose
notebook.ipynb	The full Kaggle notebook (5 cells + instructions)
setup.sh	Idempotent installer: arch check, apt (xvfb/x11vnc/novnc/websockify/X11 libs), GPU detect, cloudflared + Blender (version auto-detect, cached)
launch.sh	Starts Xvfb → x11vnc → noVNC → cloudflared → Blender; prints + saves the phone URL
startup_blender.py	Blender --python startup config: Cycles backend OptiX → CUDA → CPU, all GPUs on, prints device summary
render_test.py / run_render_test.sh	Headless Cycles smoke test → /kaggle/working/test-render.png
test_components.sh	PASS/FAIL check of every component incl. a real RFB handshake
backup_to_hf.sh	Push all .blend to a HF dataset (<user>/blender-backups)
autosave_loop.sh	Every 10 min: local archive + HF push (if token set)
teardown.sh	Stop everything cleanly

GPU notes (read this)

• 1× or 2× T4, or 1× P100: all supported. Cycles renders on all enabled GPUs (real multi-GPU). startup_blender.py prefers OptiX (fastest on both Turing and Volta with a recent driver) and falls back to CUDA, then CPU.
• Kaggle driver quirk: if Cycles reports no GPU devices even though nvidia-smi works, setup.sh already adds /usr/local/nvidia/lib64 to LD_LIBRARY_PATH (see blender-env/env.sh). Re-run launch.sh.
• Viewport ≠ render: the viewport you see over VNC is CPU software GL. "Cycles" viewport shading will still compute on GPU but display through the software GL pipeline — for phone use, model in Solid, hit Render (F12) for GPU results (save renders to /kaggle/working).
• Multi-GPU ≠ 2× speed everywhere: Cycles partitions work across GPUs, but viewports, geometry booleans, sculpting etc. stay single-core/CPU.

Persistence strategy (Hugging Face is your storage)

What	Where	Notes
Setup scripts + Blender 4.5.2 tarball (377 MB)	amer21/blender-kaggle-hub (public dataset)	The notebook fetches everything from here first — no blender.org download, works on any Kaggle session
Your .blend files	amer21/blender-backups (private dataset)	Cell 4 + autosave loop push every version every 10 min
During a session	/kaggle/working	Normal Blender Save As; also becomes the notebook output dataset
Local archive	/kaggle/working/backups/	Timestamped copies of every .blend
Code mirror	GitHub amerameryou1-blip/blender-kaggle-desktop	Fallback fetch source if HF is down

To update the hub (after editing scripts): push to GitHub, then re-upload the changed files to amer21/blender-kaggle-hub (one upload_file call per file).

Troubleshooting

Symptom	Fix
Phone shows a directory listing (app/, core/, vnc.html…)	The tunnel is alive — you just hit noVNC's root. Use the full URL printed by launch.sh (ends with /vnc_lite.html?autoconnect=1…), or re-run launch.sh (newer builds add an auto-redirect landing page, so even the bare URL works)
VNC screen is black after connecting	Blender still starting — wait 10–20 s; check blender-env/logs/blender.log
Viewport doesn't fit the phone	Make sure the URL contains resize=scale (all URLs printed since the 2026-09-16 fix do)
Cell 1 says "No internet"	Kernel settings ⚙ → Internet: On → Re-run
Tunnel URL never printed	Retry Cell 1 (cloudflared is rate-limited at times); check blender-env/logs/cloudflared.log
Cycles backend: CPU on a GPU notebook	Check nvidia-smi runs in a cell; re-run launch.sh (LD_LIBRARY_PATH fix); look at the printed device summary
Viewport is sluggish on phone	Lower resolution: edit RES_W=1280 RES_H=720 in blender-env/env.sh, re-run launch.sh; use Solid shading
apt fails on Kaggle	Rare; re-run Cell 1 (idempotent)
Blender crash on start	cat /kaggle/working/blender-env/logs/blender.log — most often OOM on 16 GB T4 with huge scenes