Pixal sits beside the ComfyUI you already run and drives it — every checkpoint, LoRA, video model and upscaler on your disk. You describe it, you correct it, it re-renders. Nothing to re-wire, nothing to re-download.
The caption under each is the recipe Pixal ran to get it — family, sampler, steps, canvas. All of it is inspectable, and all of it is overridable. It's in front of you when you want it and out of the way when you're working.
Same launcher, same flags — wrapped, not replaced. The bar is calibrated to your machine's last cold boot, so it never lies about progress.
Every phase timed, VRAM on a meter, an errors-only log. Close this window and ComfyUI closes with it.
No graph, no nodes. Model, style, character and canvas ride along as pills — touch them or don't.
What ran is on the frame — model, seed, size — and the next move is one click.
"change her hair to platinum blonde" — same scene, one change. Edits run a dedicated image-edit model, not a re-roll.
Animate hands the frame to a video model — the clip lands in the same chat, sound baked in. Click it to hear it.
One session: Pixal boots your ComfyUI, you ask in a sentence, the card lands finished — then you correct it in words and ask for motion. Both windows below are coded 1:1 from the product's own source.
██████╗ ██╗██╗ ██╗ █████╗ ██╗ ██╔══██╗██║╚██╗██╔╝██╔══██╗██║ e n g i n e r o o m ██████╔╝██║ ╚███╔╝ ███████║██║ ██╔═══╝ ██║ ██╔██╗ ██╔══██║██║ run_nvidia_gpu_fast_fp16_accumulation.bat ██║ ██║██╔╝ ██╗██║ ██║███████╗ ╚═╝ ╚═╝╚═╝ ╚═╝╚═╝ ╚═╝╚══════╝ ComfyUI 0.32.0 · NVIDIA GeForce RTX 5090 · torch 2.10.0+cu130 · sage attention ╭─ boot ───────────────────────────────────────────────────────╮ │ ✓ waking Python 7.1s│ │ ✓ prestart hooks 4.5s│ │ ✓ loading node packs 28.2s│ │ ✓ final checks 1.9s│ │ ✓ starting the web server 0.6s│ ╰──────────────────────────────────────────────────────────────╯ boot ████████████████████████████████ 100% 42.3s of ~42.3s vram ███████████████████░░░░░░░░░░░░░ 19.2/31.8 GB 11% 46°C [rgthree-comfy] Loaded 48 magnificent nodes. 🎉 ⚡ SeedVR2 optimizations check: SageAttention ✅ | Flash Attention ✅ | Triton ✅ 🔧 Conv3d workaround active: PyTorch 2.10.0, cuDNN 91200 (fixing VAE 3x memory bug) 📊 Initial CUDA memory: 30.12GB free / 31.84GB total 🔧 torchaudio.save/load globally patched (scipy for WAV files, no TorchCodec) To see the GUI go to: http://127.0.0.1:8188 1 warning C:\Pixal\logs\comfy.log E error log L full log V raw output Q quit - stops ComfyUI
Running a stranger's installer is a real risk and worth being careful about. Here is everything I would want to know first, answered straight.
Yes, and you don't have to run it to do so. Pixal installs as plain Python — nothing compiled, nothing obfuscated — and the installer payload is a git archive of the repository, so the code on your disk after setup is the code. The full source is going public as a repository, with that exact tree attached to each release as a source zip carrying its own sha256. Unpacking the .exe itself won't work today — it's Inno Setup 6.4, which neither 7-Zip nor innoextract 1.9 can open — so the zip is the route. Read it, change it for yourself; the licence asks only that you don't pass it on.
By default, nothing. Pixal listens on 127.0.0.1; chats, characters, settings and history stay in the Pixal folder; renders stay under output\pixal_dm. Two exceptions, both yours to switch on: an API chat brain receives the messages you send it, and LAN access puts the app on your own network behind an access key.
Free. No account, no queue, $0.00 a picture. The only thing that can cost money is optional: point the chat brain at an API provider and that provider bills your key, under your account with them. The local brain has neither.
It installs beside it, never inside it, and refuses to install over one. It starts your existing install with your existing launcher and flags, adds only the node packs it needs, and leaves a ComfyUI carrying local edits alone. Uninstall it and nothing about your setup has changed.
Yes, and you decide how. The chat brain can stay resident for instant replies, unload after every reply, or run on the CPU — one setting. Settings → Compute has free VRAM and free brain for eviction on demand, and a render that runs out of card clears it and retries a size down rather than dying.
Point Pixal at your existing folders and rescan — nothing is moved or copied. Your checkpoints appear in the model pill if they belong to a family Pixal has a profile for (Krea 2, Z-Image and Anima today), and your LoRAs join the chain, filtered to the ones the selected model can actually take.
No. Pixal builds and queues the graphs; you never see a node. If you want the full editor it is one setting away, or at the ComfyUI address any time — but nothing in the normal flow asks for it.
Windows 11 x64 and an NVIDIA card on driver 580 or newer. 24 GB and up runs everything, video included; 16 GB runs every still model and LTX video; 8–16 GB wants quantized builds. Under the comfortable tier things still run — they stream weights and slow down rather than refuse.
Windows is the supported target today. The Linux port is written but unreleased — LINUX.md is the runbook and makes no promises. There is no Mac build.
The pictures are yours; the licence claims no rights in your output, and you may run Pixal commercially. Three components in the box carry non-commercial terms of their own — the Anima model, FLUX.2 Klein and the PiD upscaler — and those ride on the checkpoint, not on your pictures.
Longer answers — install, settings, troubleshooting, the full FAQ — live in the manual, which also ships in the box.
One installer. Every model fetched and verified byte for byte, resumed if the line drops. It lands beside your ComfyUI, never inside it — remove it and nothing changed. No account, no queue, no upload. Ever.
Pre-release · Windows first
Kimi K3
Claude
ChatGPT