请克隆仓库而非下载 ZIP，因为更新器依赖 git 拉取；模型文件需在 Preferences 标签页手动点击下载；AMD ROCm 用户需另看 docs/INSTALL.md 配置。
Fizgig
shootthesoundApache-2.0活跃维护一站式 LoRA 与微调工作室，支持 Klein 9B、Qwen Image 2.1、Z-Image Turbo 等图像、视频、音乐模型，可训练、微调、分析、修复并提取 LoRA/LoKR。
Fizgig — LoRA & Fine-tune Studio for Klein 9B, Krea 2, MiniMax H3, Qwen Image 2.1, Anima & Z-Image Turbo and Ace-Step XL: train, fine-tune, profile, repair and extract LoRAs & LoKR for Image, Video & Music Models
v9.5.1发布于 2026-11-10 22:32 UTC714star +27 / 7 天72fork同步于 2026-11-10 23:54 UTC
git clone https://github.com/shootthesound/Fizgig.git
cd Fizgig
install_fizgig.bat 硬件要求按 README 说明
需要 NVIDIA RTX 30/40/50 系列或支持 ROCm 的 AMD Radeon，Windows 10/11 或 Linux，Python 3.11–3.13，建议 32 GB 系统内存；显存最低可到 8 GB。
需要 CUDA支持 ROCm / AMD
面向消费级显卡的 LoRA 与微调工作台，训练、修复、提取一条龙
一站式 LoRA 与微调工作室，支持 Klein 9B、Qwen Image 2.1、Z-Image Turbo 等图像、视频、音乐模型，可训练、微调、分析、修复并提取 LoRA/LoKR。
核心能力
- 在低至 8 GB 显存的消费级显卡上微调基础模型
- 支持 Klein 9B、Krea 2、Qwen Image 2.1、SDXL 等图像视频音乐模型
- 无需重训即可修复损坏的 LoRA
- 通过预览图目测挑选最佳 epoch
- 内置 Profiler、Studio、Extract 等工作台工具
容易踩的坑
内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README
- v9.5.1Fizgig v9.5.1: Element boost, and new MOSS-Image settings最新正式版2026-11-10 · 2026-11-10 22:32 UTC
Fizgig v9.5.1: Element boost, and new MOSS-Image settings
Element boost (experimental), for MOSS-Image, in Training Parameters. Early in a run, one part of each picture counts much more than the rest, so it's learned sooner. By default it's the face and hair, which brings likeness in faster, most of all when faces are small in the frame (medium and full-length shots). It runs three boosted stretches then one normal, so the body and background keep learning, for the first 12 epochs or 600 steps, whichever comes first; after that the run trains as normal. It works with or without masks (a mask trims it to the person), and a picture where nothing is found trains as normal. It's one tick: there's nothing to tune.
What it boosts: Face and hair (the default) or Face only, for a subject whose hair changes between pictures. The other choices are the ones Make masks offers (the person, the clothes, the body, the background and so on), found by the same person finder. The clothes choice suits an outfit LoRA rather than a person one.
It's on in MOSS-Image's Fast, Standard and Turbo Train presets, and off in Style, Slider and Fine-tune, where a single face isn't the point. Element boost is for MOSS-Image only for now.
MOSS-Image's Clean-end share now defaults to
11:7:40:8:3. That's 11% of steps on nearly clean images until epoch 7 (about the model's own share), then 40% from epoch 8, with every 3rd epoch after that back at 11% at half the learning rate. Muon, every preset's optimizer, trains differently from AdamW, so the settings that worked best before aren't the sweet spot any more. Starting at 11 lets composition settle first. In our tests it also stopped the hair drifting darker late in a run, and likeness late in the run was higher. It was tuned together with Element boost: the boost does the early face work, and the lower clean-end share at the start leaves room for the rest of the picture. The Fine-tune preset keeps11:4:30:8:2. The card's "Click here to try them" puts the earlier30:7:60:8:3in the boxes. Load a ✨ preset to pick up the new values; saved settings keep what they had.Also in this release:
- Training tab: related settings sit together in one box each (Save state, Gradients, Learning-rate schedule, Buckets, Metadata, Preview size, Turbo LoRA, and a fine-tune's Rotations and Regularisation). LR Scheduler and Warmup steps move from Other Options to the Optimizer section. The Metadata box fits about three to a line, one on a phone.
- Training maps: Learned since the start can also show each part's share of its own starting error, and that share on one scale for every run and epoch, so two runs compare by eye. Amount stays the default.
- Studio: a Primary + donor, 50/50 preset on every model.
- In the app window, F5 and Ctrl+R reload the page as in a browser, and if the previews gallery can't refresh it now says why instead of freezing at its last picture.
- AMD (ROCm): JoyCaption in 4-bit works on ROCm 7.15 (#316). Thanks @AlexWolfgangDRG for the diagnosis.
- The Fizgig 8 welcome no longer shows at start-up.
- Install guide: a new entry for training that's suddenly several times slower on Windows, which is usually part of it spilling into shared GPU memory, and how to stop it in the NVIDIA Control Panel.
For command-line users: Element boost is
--face_weight 9 --face_area head --face_cycle 3:1 --face_until 12:600.--face_areatakesface,head, or a Make masks choice (person,no_clothes,body_clothes,body,clothes,not_faces,background). See the training guide.Updating: run the updater as usual. Nothing new installs. The person finder Element boost uses for its non-face choices is the 16 MB model Make masks already uses; it downloads the first time if you've never made masks.
- v9.5.0Fizgig v9.5.0: JoyCaption and booru tags on the Captions tab2026-11-10 · 2026-11-10 02:18 UTC
查看更新说明
Fizgig v9.5.0: JoyCaption and booru tags on the Captions tab
Two new captioners on the Captions tab:
- JoyCaption Beta One: an 8B model made for writing training captions. It's good with people (it tells men from women, and isn't fooled by mirrors and reflections) and with every style. It follows the same instructions as Qwen3-VL, and also offers JoyCaption's own caption types (Descriptive, Straightforward, Stable Diffusion prompt, MidJourney prompt, Art critic) with a Length, plus extra options to tick: lighting, camera angle, shot type and more. The box under the ticks takes your own instructions, and the line below shows exactly what it's told. It runs in full precision with about 19 GB free, otherwise in 4-bit, which needs about 8 GB. Thanks @mart-hill for asking (#313).
- WD EVA02 tagger: booru tags (1girl, long hair, smile…) for anime and illustration, which is what Anima and anime SDXL models were trained on. Choose how sure a tag must be to keep it, and list tags never to write. Thanks @axymeus for asking (#315).
Preferences has a new Captioning models card. It shows the main captioning model (Krea 2's text encoder), the optional bigger Qwen3-VL (8B), and the other captioning models, each with a Download button. Each one also downloads itself the first time you use it (JoyCaption is 16 GB, the tagger 1.3 GB). Download ahead to have it ready or to caption offline. No Hugging Face token is needed.
Auto-recaption is unchanged. During training it always uses Qwen3-VL (the Krea 2 text encoder, or Qwen3-VL 8B if you've set it in Preferences), whichever model the Captions tab uses. The new captioners never load inside a training run, so training memory and speed are exactly as before.
Adding a captioner is now like adding a model: a description and a small class in
src/fizgig/captioners/, and Fizgig does the rest. Start with a post in New models & fine-tuning, then see Adding a captioner.Also in this release:
- Make masks (with SAM 3.1), the Look Consistency Filter and Extract now free the caption model's memory first, as training and Studio already did, so a smaller card doesn't run out.
- Detail crops follow their picture's mask. A mask you edit, add or delete updates the crops, and a picture replaced by one too small for a crop no longer leaves its old crop training. Thanks @mabseyuk.
- Batch size above 1 on MiniMax H3: a clip or a voice trains on its own step again, instead of stopping the run. Thanks @mabseyuk.
For command-line users:
batch_caption.pynow takes--captioner <name>(qwen3vlwith--model_file,florence2_promptgen,joycaption_beta_one,wd_eva02_large_v3…) instead of--backend,--text_encoderand the--florence_*flags. See the CLI guide.Updating: run the updater as usual. Nothing new installs, and nothing downloads until you pick a new captioner. Pods you use from your own Fizgig follow your computer's version when they're free.
- v9.4.0Fizgig v9.4.0: Maps in Studio, for any LoRA, and two new themes2026-11-09 · 2026-11-09 19:05 UTC
查看更新说明
New: Maps in Studio. The training maps' heatmaps now work on any LoRA Studio can load, one you didn't train included. Under the preview, pick a map and it's drawn over the Tweaked picture in a few seconds, then made again after every render:
- Where the LoRA acts: the LoRA as you've set it, against no LoRA. For a person LoRA it should gather on the face and body; spread over clothes or the background, the LoRA changes those too.
- What your tweaks change: move any slider and see where your change acts, against the Baseline.
- Trigger word: where the trigger word's attention goes (read from the LoRA, or type it and press Show). A made-up trigger word often lands on lettering in the picture, a sign or a logo, as well as the subject.
- Blocks: where each group of the LoRA's blocks acts, a button per group. On MOSS-Image, the Identity group (blocks 24–35) lights up the face while the last blocks work on edges and texture across the whole picture: the block map you'd otherwise read in the Profiler's numbers, drawn on the picture. A group that lights up the background or clothes is the one to turn down there.
All levels or one noise level, and Opacity, sit under the buttons. Picture models only; Trigger word on every picture model but SDXL. Studio guide: Maps
Two new themes: Cosmic Pigeon (space navy, nebula purple, pigeon-neck green) and Deep Space (near-black, a sparing nebula violet), with glowing buttons and star-specked, shimmering progress bars (off when your system asks for reduced motion). From @Murf-Code's Fizgig Advanced. Pick them under Preferences → Theme.
Updating: run the updater as usual. Pods you use from your own Fizgig follow your computer's version when they're free.
- v9.3.2Fizgig v9.3.2: The trigger-word map on every still-picture model2026-11-09 · 2026-11-09 17:36 UTC
查看更新说明
The Training maps' Trigger word map now works on Klein 9B, Flux 3 Image Klein, Fibo 2.0, Z-Image Turbo and Base, and Anima, as well as MOSS-Image, Krea 2 and Qwen Image 2.1. It shows how much of each area's attention goes to your trigger word inside the model, epoch by epoch: for a person it should settle on the face and body. Before training, most models also put a made-up trigger word on any lettering in the picture, such as a logo or watermark, since they read it as text; watch it move onto your subject as the LoRA learns. SDXL doesn't have it: its map is too coarse at training sizes to show anything useful. Training maps guide
The Trigger word map is easier to read. Some models spread a new word's attention almost evenly over the picture, which drew the whole picture in one colour. The map now shows where the attention is higher than in the rest of the picture.
Docs: the command-line guide covers the training maps flags, every optimizer and the fine-tune window flags; the guides for adding a model cover training maps and the trigger-word map.
Updating: run the updater as usual. Pods you use from your own Fizgig follow your computer's version when they're free.
- v9.3.1Fizgig v9.3.1: Training maps for several subjects, and Fizgig LoRA Isolation fixes2026-11-09 · 2026-11-09 15:25 UTC
查看更新说明
Training maps cover every subject. With Multi Concept or Fizgig LoRA Isolation, each subject now gets at least one of the map pictures, so four subjects get four. An Isolation run also includes one of each subject's side-by-side pictures, the subject beside a stranger: on the Trigger word map, the trigger should land on your subject and stay off the stranger. That's a direct look at whether Isolation is keeping other people themselves. When a caption names more than one subject's trigger, the Trigger word map shows each trigger on its own too. Training maps guide
Fizgig LoRA Isolation: the side-by-side pictures are found again. Every subject's side-by-side pictures share file names, and they're built in the cache folder rather than your training folder. So with two subjects, their pictures were treated as one, and the Per-image loss window couldn't show them. Each subject's are now kept apart and found. Resuming an Isolation run started before this update starts the side-by-side pictures' per-image history afresh; training itself is unaffected.
Training maps: a Map opacity slider makes the map fainter or stronger over the picture. 100% is the usual look; double-click to put it back.
Updating: run the updater as usual. Pods you use from your own Fizgig follow your computer's version when they're free.
用户点评
登录用户的真实使用记录，可按版本和硬件筛选共 0 条
还没有人点评。第一个说出真实体验的人最有价值。
📰 Latest news
- Fizgig 9.2 — faster training on PyTorch 2.15, with less memory. A MOSS-Image LoRA at batch 1 on an RTX 5090 takes 20 minutes instead of 49, batch training matches OneTrainer's speed with less memory, and likeness is unchanged. Muon is 12× faster, LoKR 4–6× faster, and telemetry is off. Needs an NVIDIA driver with CUDA 13 (580+). Thanks @tsk5cb-cmyk (#221) and @Nessa2026691 (#307). Release notes
- Fizgig 9.0.8 — Muon for every picture model. Improved likeness and training performance across most of the models: every picture model's built-in presets now train on the Muon optimizer, at their existing learning rates (thanks @Murf-Code). Load a preset again to get it. Release notes
- Fizgig 9.0.7 — Vimoo 1.0 Lite. Vimoo-AI's 9.5B Apache 2.0 video model trains in Fizgig from photos and clips: LoRAs, LoKR, sliders, full fine-tunes and masked training, with clip previews in 12 steps and image-to-video in Studio, on a 16 GB card. In our tests it sits closer to Wan 2.2 5B than to newer video models; we support Lite because it's light. Also Batch size Auto on 40 GB+ cards, and Studio waits for Update. Release notes
- Fizgig 9.0.6 — Identity mode for Extract. Tick Identity mode and a character's identity blocks keep 4× the rank, so a fine-tune extracted as a LoRA keeps its likeness at a small size (MOSS-Image and Qwen Image 2.1; MOSS also gets Fast Identity Mode and Studio's identity presets). Full fine-tunes can train in batches, and the per-image loss watch now works at any Batch size (thanks @Murf-Code). Release notes
- Fizgig 9.0.5 — Batch size. Train up to 8 pictures a step on every model except ACE-Step (thanks @Murf-Code and @tsk5cb-cmyk). It pays most with block swap (Qwen Image 2.1 2.7× faster at batch 3) and on smaller models (Anima 2.6×, SDXL 2.3× on a 4090), and the Training tab says what the batch needs on your card. Move to… can batch a run on a bigger pod and bring it back at your own size. Also fixed: Flux 3 Image Klein's download, and Klein's VAE sharing a name with Z-Image's. Release notes
- Fizgig 9.0.4 — Fibo 2.0. BRIA's 8.2B model for text-to-image and editing trains in Fizgig: LoRAs, edit LoRAs, slider LoRAs and full fine-tunes from a 12 GB card, 4-step Turbo previews, masked training, and every workbench tool. Everything it saves loads in ComfyUI. Brought to Fizgig by @spaghettified26. Release notes · Fibo 2.0 guide
- Fizgig 9.0.1–9.0.3 — transparent pictures, new optimizers, H3 masks. Transparent PNGs train with their transparency on Qwen Image 2.1, and the previews keep it (thanks @CognitiveDiffusion); on every other model they're laid on white, as the captioner sees them, instead of on the colour hidden behind them. Three new optimizers, Prodigy, Muon and Automagic v4 (thanks @Murf-Code), masked training for MiniMax H3 stills (thanks @spaghettified26), Studio settings kept inside a downloaded picture, a right-click menu in the app window, the Pods tab on RunPod's new API, and Start downloads the model if it's missing. Release notes
- Fizgig 9 — refinements, and Detail crops for styles. The 63 releases since 8.0 brought together: an app that explains itself, runs that survive things going wrong, and pods that work like your own computer. New in 9, Detail crops train each picture as a square crop at its own scale too, so brushwork, grain and texture train pixel for pixel, and a Style score in the Samples gallery gives a style LoRA a number to watch epoch by epoch (thanks @Murf-Code). Caption tasks now cover people, characters, objects, products, styles and concepts, and Adaptive LR judges plateaus on a 3-epoch average. FIBO 2.0 is on its way (#250, @spaghettified26). Release notes
- Fizgig 8.9.7–8.9.98 — Z-Image Base, masks from words, and more. Z-Image Base trains as its own model, with previews on Turbo (8.9.95, thanks @girzai), and a bigger 8B captioning model can be picked in Preferences (thanks @Dwuuds). SAM 3.1 makes masks of anything you name, and Edit masks fixes them by hand (8.9.9, 8.9.91). Train on sub-folders (8.9.92), eight themes (8.9.93), Start at a smaller size, caption dropout and a Run summary (8.9.98), and your pods log you in by themselves (8.9.97). All releases
- Fizgig 8.9.6 — masked training. Train on the part of each picture that matters: the person, not the room around them. A black-and-white mask beside a picture says which part counts, and Image Prep makes them for a whole folder in one click (the person, just the face and hair, the body without the face, and more). In our tests of character LoRAs, masks significantly raised both the best likeness a run reached, by eye and in likeness scores, and how quickly it got there, on every recent model. It seems especially effective on MOSS-Image with its default Clean-end share, Slide. Release notes · Masked training guide
- Fizgig 8.9.5 — faster training on 40 GB+ cards. On most pods (A100, A40, A6000, L40S, H100, RTX PRO 6000) Auto now turns gradient checkpointing off when the run fits with room to spare: 13–27 % faster per step on Krea 2, Qwen Image 2.1, MOSS-Image, Z-Image and MiniMax H3. Where it doesn't all fit, the last blocks go without it: 2–21 % faster. Previews keep at least 4 GB free, and cards of 32 GB and under train as before. Release notes
- Fizgig 8.9.0 — your other GPUs, at home or in the cloud, speeding up your run. A helper is another computer running Fizgig: a second PC or laptop, a second GPU in this computer, or a RunPod or Vast.ai pod. It renders your previews so training never stops for them, takes over a run while you need your computer, or trains a share of each epoch alongside your GPU. In a share, each GPU trains its own copy of the LoRA on its part of the images and the copies' changes are added into one LoRA at every combine: in testing its likeness matched a one-GPU run. On a fast network, combine every epoch for the closest result. It pays most on long runs. Training on a pod? Your home GPU can help it, and its checkpoints come home as they're saved. Release notes · ▶ Watch the video · Helpers guide
- Fizgig 8.8.1 — MOSS-Image likeness sooner, and previews with the SFT look. A new Clean-end share default, Slide, brings likeness in early and keeps it: on our most difficult dataset, the best of eight settings we tried, around 70% likeness from epoch 20 to epoch 60. Base previews can take on much of MOSS-Image SFT's look with no model swap, every enlarged preview says how it was made, and Krea 2's Auto precision no longer picks a slow setup on 16 GB cards. 8.8.2 stops the web UI freezing when you open it during a long run. Release notes · 8.8.2 · ▶ Watch the video · MOSS-Image guide
- Fizgig 8.5 — MOSS-Image. OpenMOSS's new 5B open image model trains in Fizgig: LoRAs, LoKRs, sliders, full fine-tuning, Multi Concept and Fizgig LoRA Isolation, with previews on MOSS-Image SFT, the model most people generate with, and every workbench tool. Brought to Fizgig by @Murf-Code. 8.5.1 fixes the GPU readout on AMD cards (thanks @scryptio); 8.5.3 fixes Linux launching; 8.5.4 makes MOSS-Image previews render as ComfyUI does; 8.5.5 trains MOSS-Image at 1 MP; 8.5.8 fixes MOSS-Image training with block swap; 8.5.9 makes MOSS-Image Studio renders faster; 8.6.0 improves MOSS-Image likeness and adds an option for training on SFT; 8.6.1 fixes noisy Qwen Image 2.1 previews; 8.7 brings MOSS-Image likeness much sooner with a new Clean-end share option, 30% by default since 8.7.1, fine-tunes included in 8.7.2, since 8.7.3 a fine-tune fits a 16 GB card, and 8.7.4 adds an 8-step Turbo LoRA for fast previews on Base, 8.7.5 lets MOSS-Image train on computers with 16 GB of RAM, 8.7.6 fixes the cache on a resume and for a refilled folder, 8.7.7 drops Kromoss, 8.7.8 adds preview and save schedules, 8.7.9 is a maintenance release, and 8.8.0 fixes resuming a run on Python 3.10. Release notes · MOSS-Image guide
- Fizgig 8.4 — Fizgig LoRA Isolation. Train two people into one LoRA and every stranger keeps their own face: with a normal LoRA 48 of 50 strangers came out looking like the trained person, with LoRA Isolation 0 of 50, at the same likeness. DOP from @ostris's AI-Toolkit at its base, with Fizgig's own techniques on top, set up on the Training tab's new Subjects card, on every model except ACE-Step. 8.4.2 adds a trigger word checker on the same card. Demo · How to set it up · Release notes
- Fizgig 8.1 — Flux 3 Image Klein, from day one. Black Forest Labs' new 5.6B text-to-image and edit model trains in Fizgig the day it came out: LoRAs, LoKRs, edit LoRAs from before/after pairs, slider LoRAs and full fine-tuning, from an 8 GB card, with every workbench tool, prompt travel and edit chains in LoRA Royale included. Release notes · Flux 3 guide
- Fizgig 8 — the new app (yes, the web UI). Fizgig opens in its own window or your browser, and your phone can watch a run. One Pods tab opens your other computers' GPUs, found on your network by themselves, and RunPod pods, logged in for you. What's training and what's up next sit in the sidebar on every tab, and Runs & outputs keeps every run. Updating keeps your settings, presets, folders and queued runs; a welcome tour on the first start shows where everything went. Release notes · The app
What Fizgig is
A trainer and workbench for image, video and music models, in its own window on your computer, in your browser, or on a rented GPU — and you can watch a run from your phone. Every trainer makes LoRAs; Fizgig is also built around what you do with them afterwards:
- Fix a LoRA block by block without retraining: drag a slider per block, save a new
.safetensors. - Explore variations: 🎲 I'm feeling lucky in Studio rolls new block strengths within your limits; keep the ones you like.
- Find the best epoch by eye: LoRA Royale renders every epoch on one seed with a crossfade, and exports shareable loops.
- Profile which blocks carry identity, style and detail before you touch anything.
- Train beyond LoRAs: LoKR, or the full base model itself: Anima on an 8 GB card, the larger models on 16 GB. A LoRA extracted from a fine-tune came out better than one trained directly at the same rank.
Memory plans itself: precision, block swap and previews size to your free VRAM, and if a preview can't fit, training keeps going. Fizgig loads kohya, PEFT, OneTrainer, AI-Toolkit and LyCORIS LoRAs, and saves .safetensors that drop straight into ComfyUI. Free and open source.
https://github.com/user-attachments/assets/2db82d0d-e7ed-4f60-a79c-d2ac1f4984ac
Supported models
| Model | Trains on | LoRA | LoKR | Slider / Edit LoRA | Full fine-tune | Smallest card | Guide |
|---|---|---|---|---|---|---|---|
| Flux 2 Klein 9B | photos | ✅ | ✅ | ✅ slider + edit | ✅ experimental | 10 GB | Klein 9B |
| Flux 3 Image Klein (5.6B), experimental | photos | ✅ | ✅ | ✅ slider + edit | ✅ experimental | 8 GB | Flux 3 Image Klein |
| Fibo 2.0 (8.2B), experimental | photos | ✅ | — | ✅ slider + edit | ✅ experimental | 12 GB | Fibo 2.0 |
| Krea 2 (12.9B) | photos | ✅ | ✅ | ✅ slider | ✅ experimental | 8 GB | Krea 2 |
| MiniMax H3 (33B) | photos, video clips, sound, voice | ✅ | ✅ | ✅ slider | ✅ experimental | 16 GB | MiniMax H3 |
| Vimoo 1.0 Lite (9.5B), experimental | photos, video clips | ✅ | ✅ | ✅ slider | ✅ experimental | 16 GB | Vimoo 1.0 Lite |
| Qwen Image 2.1 | photos | ✅ | ✅ | ✅ slider + edit | ✅ experimental | 10 GB | Qwen Image 2.1 |
| MOSS-Image (5B), experimental | photos | ✅ | ✅ | ✅ slider | ✅ experimental | 12 GB | MOSS-Image |
| Z-Image Turbo (6B), experimental | photos | ✅ | ✅ | ✅ slider | ✅ experimental | 12 GB | Z-Image Turbo |
| SDXL (any checkpoint), experimental | photos | ✅ | ✅ | ✅ slider | ✅ experimental | 8 GB | SDXL |
| Anima (2B), experimental | photos | ✅ | ✅ | ✅ slider | ✅ experimental | 8 GB | Anima |
| ACE-Step 1.5 XL (4B), experimental | songs | ✅ | — | ✅ slider | ✅ experimental | 16 GB | ACE-Step |
Every image and video model gets all four workbench tools; ACE-Step has Studio and Extract. Qwen Image 2.1 trains on the base model and previews on Qwen's own Turbo (8 steps, CFG 1); your LoRAs work on both. MOSS-Image trains on Base and previews on MOSS-Image SFT, the model most people generate with. MiniMax H3 also makes RefMods, tuned against H3 itself rather than a plain encode (how do I…?). Fizgig's training adapters for Qwen Image 2.1 and Z-Image Turbo are free on Hugging Face for any trainer. Full fine-tuning has its own guide.
The workbench
Each tool works on your own runs or any LoRA you've downloaded, and they hand off to each other. More detail.
- Studio (was Repair Studio) — a live slider per block with a side-by-side preview, donor-LoRA blending, 🎲 I'm feeling lucky for random variations within your limits, and an exact baked save. On H3 the previews are clips with sound, with a background block library and first/last-frame pinning. Guide
- LoRA Royale — every epoch on one seed, an optional likeness score, and MP4/GIF exports: epoch morphs, seed/prompt/strength travels, comparison sheets.
- Profiler — a colour-coded per-block report that Studio reads inline.
- Extract — shrink any LoRA to a lower rank.
Training features
- Presets per model: pick a ✨ preset on the Training tab and go.
- Slider LoRAs (Klein 9B, Flux 3 Image Klein, Fibo 2.0, Krea 2, MiniMax H3, Vimoo 1.0 Lite, Qwen Image 2.1, MOSS-Image, Z-Image Turbo, SDXL, Anima): a LoRA whose strength is a dial between two looks, from photo pairs or three prompts, and a slider from prompts can be steered while it trains (change its ends, it trains on toward them). Krea 2's Ultra mode trains the composition blocks only, so the dial holds up at much higher strengths. It works best for sliders trained from prompts, and with photo pairs when the change is compositional.
- Edit LoRAs (Klein 9B, Flux 3 Image Klein, Fibo 2.0, Qwen Image 2.1): teach an edit from pairs of original and edited photos, then apply it to any photo.
- Fizgig LoRA Isolation, on the Subjects card (every model except ACE-Step): up to four people in one LoRA, each from their own folder. Fizgig's own techniques combined with DOP keep your subjects to their trigger words: leave the trigger out and you get the model's own people, and in a group picture your subjects, strangers and people the model already knows each keep their own face. See it in action · How it works · How to set it up (video) · How to caption for it.
- Masked training (every image model): train on the part of each picture that matters, the person and not the room around them. Image Prep makes the masks for a whole folder in one click, or you paint your own. In our tests of character LoRAs it significantly raised both the best likeness and how quickly runs got there. How it works
- Adaptive LR: a plateau tracker that raises or lowers the rate within your Min/Max, with rollback on instability.
- Weight averaging (EMA), on by default where it's measured to help.
- Context LoRA: train on top of a frozen, active LoRA so the two coexist (a face on a style, an outfit on a character). No other trainer does this.
- Pause and resume with full state, plus a training queue for back-to-back runs.
- The trainer curates your dataset while it trains (every model except MiniMax H3). Every image starts as useful training data; the watch follows each one's loss, and only when an image stops teaching the model does it step in: throttle it, recaption it, and as a last resort set it aside. It also tells you when the whole run has plateaued. No other trainer does this.
- Fast previews in a few steps on a turbo or distilled model where one exists: Klein's and Flux 3 Image Klein's distilled models, the Turbo LoRAs for Fibo 2.0, Krea 2, MiniMax H3, Qwen Image 2.1 and Anima, Vimoo's distilled LoRA, and Qwen Image 2.1's Turbo model.
- A sample gallery that scores likeness live, and a run visualiser to scrub epochs.
- Dataset prep: AI captions with Qwen3-VL or Florence-2, bilingual captions, face crops, and a Look Consistency Filter.
- Face backends for every kind of character: face crops, the Look Filter and likeness scores work on anime, furry and fursuit characters as well as people (see the section below).
- Gizmo (MiniMax H3): cut clips with scene detection, crop to the subject, and record or segment a voice dataset with Whisper transcription.
Full list: docs/TRAINING.md. Everything also runs headless: docs/CLI.md.
One app, wherever you are
- In its own window (
Launch Fizgig.bat), or in your browser (Launch Fizgig in browser.bat) — the same app either way. - The queue in sight: what's training and what's queued sit in the sidebar on every tab; a held queue says why, with Start queue right there. Continue on any run in Runs & outputs picks it up again from its last save, so you can pause one project, train another and come back.
- On your phone or another computer: set a password in Preferences and open it over your network, or scan the Start tab's Open on your phone QR code and you're straight in; add it to your phone's home screen and it opens like an app.
- On your other computer's GPU: train on your desktop's card from your laptop. Fizgig finds your other computers running it, and the Pods tab opens one as it does a pod (how).
- On a rented GPU: the Pods tab finds the card that's cheapest per run among those RunPod has free right now (one card or two), deploys it, opens the pod in the same window (logged in for you), and sends a training folder from your computer straight to it. Prefer Vast.ai? Fizgig has a template there too, with https addresses and a stop when the run finishes (Vast.ai).
- Your other GPUs helping: a second computer, a second GPU in this one, or a pod renders your previews so training never stops for them, trains a share of each epoch alongside yours, or takes over a run while you game or someone else needs the computer, then hands it back. A pod's checkpoints can come home as they're saved, and the pod can be deleted once they're all here (Helpers).
- Runs & outputs keeps every run with its loss graph, notes, previews and checkpoints, and starts any of them again with the same settings.
More: docs/WEB_UI.md.
Face backends
Face crops, the Look Consistency Filter and every likeness score can work on anime, furry and drawn characters and fursuits too, not just people.
The defaults are InsightFace (face finder) and ArcFace (look match), the same as before. People's sets need nothing changed. Everything else is opt-in: open Image Prep → Anime/furry/drawn characters and pick:
- Face finder: Anime/furry/drawn (YOLO model) for characters, or Both for a set with people and characters in it.
- Look match: Anime/furry/drawn for characters, or Auto (Decide for each set). Each look match has its own thresholds, and saved scores record which one made them.
Setup (which face model to use and how to get it), numbers and limitations are in docs/FACE_BACKENDS.md.
Install
Needs an NVIDIA RTX 30/40/50-series or an AMD Radeon with ROCm, Windows 10/11 or Linux, Python 3.11–3.13, and 32 GB of system RAM recommended. Full requirements, AMD ROCm and model downloads: docs/INSTALL.md.
Clone the repo rather than downloading the ZIP (the updater pulls with git):
git clone https://github.com/shootthesound/Fizgig.git
cd Fizgig
- Windows (NVIDIA): double-click
install_fizgig.bat, then launch withLaunch Fizgig.bat(its own window) orLaunch Fizgig in browser.bat(in your web browser); update withupdate_fizgig.bat. - Linux (NVIDIA):
python install_fizgig.py, then./launch_fizgig.shor./launch_fizgig_browser.sh; update with./update_fizgig.sh. - AMD ROCm: see docs/INSTALL.md.
Model files download from the Preferences tab: one Download models for me button per model.
No GPU? Rent one
Nothing to install on your computer: the whole app runs on the pod, in your browser.
- ⚡ Deploy on RunPod, and add a 300 GB Volume disk before Deploy Pod: RunPod doesn't add one for you (how).
- Wait for Ready in the pod's Logs tab; the password is printed under it.
- Connect → HTTP port 8081, and log in with that password. (Opened from Fizgig on your own computer, with your RunPod API key on its Pods tab, the pod logs you in by itself.)
- Preferences → Download models for me, upload your training folder on the Start tab, and press Start on the Training tab.
Prefer Vast.ai? The same image runs there: Fizgig on Vast.ai (guide).
Step by step: Quick start. The pod also has a drag-and-drop file manager for big datasets and fine-tunes (the File manager button in the sidebar), one-click model downloads and an optional auto-stop when training finishes. Opened from your own Fizgig, a pod runs the same version as your computer (which Fizgig a pod runs). The full pod guide.
Mid-job on the old pod with the remote desktop? It stays available as Fizgig Classic v7.x, and pods you already have keep running as they are.
Getting started
Work down through the numbered tabs:
- Start — set your training image folder.
- Image Prep (optional) — resize, face-crop, and filter out images that don't match the look.
- Captions — trigger-word or AI captions.
- Samples — the preview prompts that render during training.
- Training — pick a model and a preset, click Start Training.
Tip: give one sample prompt something your dataset doesn't have, such as a specific pose, angle or setting that none of your photos show (for example "looking over her shoulder from a low angle" when your photos are all straight-on portraits). Early epochs follow it. When a LoRA starts to overtrain, that is the first thing to fail: the previews drift back to your dataset's framing even though the prompt asks for something else. The last epoch that still follows it is a good one to keep.
The unnumbered tabs are Runs & outputs, the workbench (Profiler, Studio, LoRA Royale, RefMod Toolkit, Extract, Metadata), Preferences, and Pods. The video tutorial walks through all of it.
Updating from Fizgig 7
Run update_fizgig.bat as always. Fizgig 7's desktop app is replaced by the new one: start it with Launch Fizgig.bat from now on (run_fizgig.bat is gone, so point a shortcut at the new file). Everything of yours moves into one user_data folder: your settings, presets, models, cache and queued runs come across, Fizgig opens on the model you last trained, and the welcome on the first start shows where everything went.
Community translation: Korean (한국어) by @ssain3d-lgtm, an unofficial add-on that translates the UI at runtime without touching Fizgig's files and uninstalls with one script. If you hit a bug with it installed, uninstall and reproduce before reporting here; translation issues go to its repo.
Support the project
Fizgig is free. If it saves you time or makes your LoRAs better, a coffee or a GitHub sponsorship keeps it going:
Renting through the RunPod link supports development too, at no extra cost to you. Add a 300 GB Volume disk on RunPod's page before Deploy Pod (how).
Documentation
| Guide | What's in it |
|---|---|
| Install and requirements | Hardware, system RAM, AMD ROCm, model downloads, VRAM tables |
| Klein 9B · Flux 3 Image Klein · Fibo 2.0 · Krea 2 · MiniMax H3 · Qwen Image 2.1 · MOSS-Image · Z-Image Turbo · SDXL · Anima · ACE-Step | Each model: presets, features, memory |
| Training and the workbench | The shared features and tools in detail |
| Face backends | Face finder (InsightFace / YOLO) and look match (ArcFace / DINOv2) for people, anime and furry |
| Full fine-tuning · How do I…? | Training the base model itself |
| RefMods | RefMods for MiniMax H3 |
| Command line | Every trainer, headless |
| Fizgig LoRA Isolation | How it works, the recipe and the results |
| Masked training | Train on the part of each picture that matters; making masks, or painting your own |
| The app | Starting Fizgig, its window and browser, your phone, your other computers, the port and password, pods |
| Helpers · Previews on a helper · Training share · Move a run | Another computer or a pod renders your previews, trains a share of each epoch, or takes over a run while you use your computer |
| RunPod · Vast.ai | The cloud image, on RunPod and on Vast.ai |
License
Fizgig is open source under the Apache License 2.0 — free to use, modify, and redistribute, including commercially, with attribution and no warranty. Third-party components under compatible permissive licenses (and other terms where noted) are listed in THIRD_PARTY_NOTICES.md.
The Automagic v3 and v4 optimizers, the Krea 2 MMDiT backbone and flow-matching sampler, and Ostris's MiniMax H3 training adapter all come from @ostris — the first two from AI-Toolkit under the MIT licence, the adapter downloaded as a model rather than bundled. The default H3 training adapter comes from circlestone-labs, also downloaded as a model. Anima's model code is adapted from kohya-ss's sd-scripts under the Apache licence, Z-Image's from Tongyi-MAI's reference implementation, and Flux 3 Image Klein's from Black Forest Labs' reference implementation, both Apache too. The Prodigy optimizer is Konstantin Mishchenko's prodigy (MIT), and Muon is PyTorch's torch.optim.Muon (BSD).
A few features came from looking at other trainers: starting a run at a smaller size, the run summary, the sidebar's disk and job meters and the captions table were inspired by ZenTrainer by @Nynxz.
Copyright © 2026 Peter Neill.
Model weights are not covered by this license — each model carries its own terms from its publisher (see the Download links in Preferences).
Working with me
I'm available for consulting on local AI training pipelines, custom workflow tooling, and private model work — the same engineering that's in Fizgig, applied to your studio's hardware and IP. Get in touch: [email protected].