Fizgig 9.2: Train a MOSS-Image LoRA in 20 Minutes (Install, Settings, VRAM and ComfyUI)

Training your own LoRA used to mean a long wait and a big GPU. Fizgig 9.2, the free LoRA studio from Peter Neill (shootthesound), cuts a MOSS-Image LoRA from 49 minutes to about 20 on an RTX 5090, according to its release notes.

Memory drops even more. A MOSS-Image run at batch 3 fell from 21.7 GB to 7.1 GB, which brings serious training to 8-12 GB cards.

This guide covers what changed in 9.2, every supported model, a clean install on Windows, Linux and RunPod, a step-by-step first LoRA, how Fizgig compares with OneTrainer, ZenTrainer and KohakuMoss, and fixes for the problems people hit after updating.

What Is Fizgig?

Fizgig is a free LoRA and fine-tuning studio by Peter Neill, who publishes as shootthesound on GitHub. The same developer made the ComfyUI Realtime LoRA Trainer node, so the tool grew out of the ComfyUI community.

It runs in its own window, in your browser, or on a rented GPU pod. You point it at a folder of images (or clips or songs), it preps and captions them, then trains a LoRA, LoKR, slider or full fine-tune.

A quick version history

The repo had about 586 stars and 60 forks when I checked on 9 November 2026, with 8 open issues.

What Is New in Fizgig 9.2

The 9.2 release is mostly about speed and memory, with the developer saying likeness matches v9.0.8 or better. Here are the changes that matter to you.

• PyTorch 2.15: with gradient checkpointing on, memory drops sharply, even at batch 1. This is a roughly 3 GB download during the update.
• Muon 12x faster: Muon now processes a LoRA’s same-shaped matrices together. Results are identical, and this explains most of the batch-1 gain.
• Faster batches: different caption lengths no longer slow attention when batch size is above 1.
• LoKR 4-6x faster on every model, after a community report (issue #307) that LoKR Context LoRA training was about 5x slower than it should be.
• Adaptive LR ceiling: if a learning-rate rise is followed by a cut within three epochs, later rises stay below that rate.
• Faster MiniMax H3 previews: “Show early” now uses taeh3, a 22.7 MB decoder by madebyollin, about 10x faster than the 4.5 GB full decoder. Final clips still use the full VAE.
• Telemetry removed: recent onnxruntime builds on Linux send device details to Microsoft by default. Fizgig now switches that off, plus Hugging Face usage reporting, on every platform.
• Optimizer switch fix: moving a preset from Muon or Prodigy to AdamW now clears their optimizer args, so AdamW no longer silently falls back to a slower plain version.

The 9.2.5 MOSS-Image Turbo Train preset

Version 9.2.5 adds an experimental Turbo Train preset for MOSS-Image. It takes the Fast preset (rank 16, 0.25 MP, Muon) and adds Adaptive LR between 2e-4 and 4e-4 with per-image tracking.

The developer frames it as a quick preview before a long run, not a final-quality preset. It also changes the MOSS-Image Clean-end share to 30:7:60:8:3: 30% of steps on nearly clean images until epoch 7, then 60% from epoch 8.

The Speed and Memory Numbers

All figures below are developer-reported from the 9.2.0 release notes and GitHub issue #221. I have not run these tests myself.

How it stacks up against OneTrainer

The update came out of a public speed comparison. In issue #221, a user showed OneTrainer finishing a 39-image MOSS-Image LoRA on an RTX 3090 in 17 min 43 s, while Fizgig with gradient accumulation took 36 min 17 s.

After the fixes, the developer reported these side-by-side results:

• A40, batch 3: Fizgig 1.07 s/step at 6.9 GB vs OneTrainer 1.07 s/step at 7.4 GB.
• Rented RTX 3090 on Linux, INT8 for both: Fizgig 0.99 s/step vs OneTrainer 1.12 s/step.
• RTX 5090, same settings: Fizgig 0.47 s/step vs OneTrainer 0.49 s/step.
• RTX 5090, AdamW at batch 1: OneTrainer was still ahead, 0.241 vs 0.29 s/step.

The user who opened the comparison re-tested on 9.2 under Windows 11 and got 26 min 3 s at 7.4 GB peak with Muon and INT8 Auto precision. That is far better than before, but still slower than their OneTrainer run, and the Windows gap was left open.

Note: Linux results were faster than Windows in the same thread. If you train on Windows, expect numbers closer to the user’s re-test than to the developer’s Linux and RunPod figures.

Supported Models and Minimum VRAM

The README lists these models. Every model except ACE-Step supports both LoRA and LoKR, and full fine-tuning is marked experimental for all of them.

The 9.2 release notes also mention testing Fibo 2.0 and Vimoo, so check the Training tab for the current list. Fizgig sizes precision, block swap and previews to your free VRAM automatically.

What you need

• GPU: NVIDIA RTX 30, 40 or 50 series, or an AMD Radeon card with ROCm.
• NVIDIA driver: 580 or newer (CUDA 13) for 9.2.
• OS: Windows 10/11 or Linux. No macOS support is listed.
• Python: 3.10 to 3.13.
• System RAM: 32 GB recommended, more helps with block swap on big models.
• Disk: room for base models plus the 3 GB PyTorch update. The RunPod template asks for a 300 GB disk.

How to Install or Update Fizgig

Clone the repo with git rather than downloading the ZIP, because the built-in updater uses git.

git clone https://github.com/shootthesound/Fizgig.git
cd Fizgig

Windows (NVIDIA)

1. Update your driver. Install NVIDIA driver 580 or newer first, or Fizgig 9.2 may not see your GPU.
2. Run the installer. Double-click install_fizgig.bat in the Fizgig folder.
3. Launch it. Use Launch Fizgig.bat for its own window or Launch Fizgig in browser.bat for a browser tab.
4. Download models. Open the Preferences tab and press “Download models for me” next to each model you want.

Linux (NVIDIA)

python install_fizgig.py
./launch_fizgig.sh          # own window
./launch_fizgig_browser.sh  # browser
./update_fizgig.sh          # later updates

AMD users should follow docs/INSTALL.md in the repo. In 9.2, ROCm installs keep their own bundled PyTorch, so nothing changes for them.

RunPod (no local GPU)

1. Deploy the template from the README link and keep the 300 GB disk.
2. Wait for “Ready” in the pod’s Logs tab. The login password is printed there.
3. Open Connect, then HTTP port 8081, and log in.
4. Download models in Preferences, upload your image folder on the Start tab and press Start on the Training tab.

Tip: the pod has an optional auto-stop when training finishes, which saves money on long runs. Existing pods pick up 9.2 after a restart. The README link is a referral link that supports the developer.

Updating from an older version

1. Run the updater (update_fizgig.bat or ./update_fizgig.sh).
2. Let the check finish. It shows a 10-second countdown; press Space to go now.
3. Wait for PyTorch 2.15 (about 3 GB) to download.
4. If it will not start or train, run the updater again and press R. This rebuilds the Python environment and keeps your models, datasets, LoRAs and settings.

Train Your First LoRA Step by Step

Fizgig’s tabs are numbered in the order you use them. Here is a first MOSS-Image character or style LoRA, using the presets.

1. Start: pick your image folder. 20-100 varied, sharp images is a sensible range; the developer’s MOSS-Image test used 94.
2. Image Prep: resize, face-crop if it is a person, and run the Look Consistency Filter to drop images that do not match.
3. Captions: choose trigger-word captions or AI captions with Qwen3-VL or Florence-2. Bilingual captions are supported.
4. Samples: write two or three preview prompts using your trigger word, so you can watch progress.
5. Training: pick MOSS-Image and a preset. Try Turbo Train (9.2.5) as a quick preview, then a normal preset for the final run.
6. Check results: use LoRA Royale to render every epoch on one seed and choose the best one, not just the last.

Handy settings to know

The workbench adds three more tools: Studio for per-block sliders with side-by-side previews, Profiler for per-block analysis, and Extract to shrink a LoRA to a lower rank.

Use the LoRA in ComfyUI

Fizgig saves kohya-format .safetensors files that load with ComfyUI’s stock LoRA loader. No custom node is needed.

1. Copy the file from Fizgig’s Runs & outputs folder into ComfyUI/models/loras/.
2. Restart or refresh ComfyUI so the file appears in the list.
3. Add a “LoraLoaderModelOnly” or “Load LoRA” node between your model loader and the sampler.
4. Start at strength 0.8-1.0 and use your trigger word in the prompt.

For the base MOSS-Image files and folders, see my MOSS-Image in ComfyUI guide. If you get black images in fp16, the fix is in MOSS-Image-heal.

Note: Fizgig can also load LoRAs from kohya, PEFT, OneTrainer, AI-Toolkit and LyCORIS, so you can open older LoRAs in Studio or Extract.

Fizgig vs OneTrainer, ZenTrainer and KohakuMoss

Which should you pick?

• You train many model types: Fizgig, because one interface covers images, H3 video and ACE-Step songs.
• You live in ComfyUI: ZenTrainer, so you never leave your workflow.
• You have a 6-8 GB card and only want MOSS-Image: KohakuMoss.
• You want the fastest AdamW batch-1 runs: OneTrainer was still slightly ahead in the developer’s own 5090 test.

Troubleshooting

Fizgig cannot see my GPU after updating

Version 9.2 needs NVIDIA driver 580 or newer for CUDA 13. Update the driver, reboot and launch again.

It will not start or train after the update

Run the updater and press R during the countdown. It deletes and reinstalls the Python environment and keeps your data.

My saved preset still uses the old schedule

Saved presets keep the previous Clean-end share. Reload the built-in preset to pick up 30:7:60:8:3.

Batch size is greyed out

Edit LoRAs, sliders, fine-tunes, DOP, LoRA Isolation and per-image loss watch read one picture per step, so Fizgig forces batch 1 for them. This is by design.

AdamW feels slower than before

Before 9.2, switching from Muon or Prodigy to AdamW could keep their optimizer args and fall back to plain AdamW. In 9.2 this is fixed; re-select AdamW or AdamW8bit so the args are cleared.

Windows is slower than the published numbers

In issue #221, Linux runs were consistently faster than Windows. Use Auto precision, keep gradient checkpointing on, and compare like with like (INT8 vs INT8) before deciding a tool is slow.

Licence and Privacy Notes

• Fizgig: Apache 2.0, so you can use it commercially. Third-party parts are listed in THIRD_PARTY_NOTICES.md.
• Base models: not covered by Fizgig’s licence. MiniMax H3, for example, uses the MiniMax H3 community licence; check each model’s terms before selling LoRAs.
• Privacy: 9.2 turns off onnxruntime and Hugging Face usage reporting on all platforms.
• People: only train on photos you have the right to use, and get consent before training a LoRA of a real person.

Warning: the update rebuilds parts of your Python environment. Back up any custom presets you edited by hand before updating.

My Take

What I like most about this release is how it happened. A user posted hard OneTrainer numbers, the developer reproduced them, found three separate causes (compiler memory, INT8 gradients and a slow Muon step), and shipped fixes within days.

The memory drop is the real story for most readers. Going from 21.7 GB to 7.1 GB at batch 3 means a MOSS-Image LoRA is now realistic on a 12 GB card like the RTX 3060 12GB, not just on a 5090.

I would still treat the 20-minute figure as a best case on Linux with a top card. On Windows, plan for something closer to the 26 minutes the community tester reported, and use Turbo Train only as a preview.

Frequently Asked Questions

Is Fizgig free?

Yes. Fizgig is open source under Apache 2.0 and free to download from GitHub. You only pay if you choose to rent a GPU on RunPod.

Is Fizgig a ComfyUI custom node?

No. It is a standalone trainer with its own window or browser UI. The LoRAs it makes load in ComfyUI with the normal LoRA loader from ComfyUI/models/loras/.

How much VRAM do I need for a MOSS-Image LoRA?

The README lists 12 GB as the smallest card for MOSS-Image. The developer reports about 7 GB at batch 3 with gradient checkpointing on 9.2.

How long does a MOSS-Image LoRA take in Fizgig 9.2?

About 20 minutes for 94 images and 70 epochs on an RTX 5090, according to the release notes. A community tester on Windows reported about 26 minutes for a smaller set on an RTX 3090 at batch 3.

Does Fizgig work on Mac?

The README only lists Windows 10/11 and Linux with NVIDIA or AMD ROCm GPUs. On a Mac, the RunPod template is the practical option.

Can I train MiniMax H3 video LoRAs with Fizgig?

Yes. MiniMax H3 (33B) trains on photos, video clips, sound and voice, from a 16 GB card with block swap. The Gizmo tool cuts clips and transcribes voice data with Whisper.

Why does Fizgig need driver 580?

Version 9.2 moved to PyTorch 2.15 built for CUDA 13, which needs NVIDIA driver 580 or newer. AMD ROCm installs are not affected.

Sources

• Fizgig v9.2.0 release notes (GitHub)
• Fizgig README: models, requirements, install (GitHub)
• Issue #221: batch size and OneTrainer comparison (GitHub)
• ComfyUI Wiki: Fizgig 9.2 cuts MOSS-Image LoRA training to 20 minutes
• ComfyUI Wiki: Fizgig v7