AnimaForge
Free, local, one-click LoRA trainer built only for Anima. Point it at a folder. Name it. Forge.
Training a LoRA shouldn't mean an evening lost in Kohya's wall of knobs. AnimaForge is a free Windows app that does the whole run from one screen: pick your image folder, name the LoRA, hit Train. It auto-captions (WD14 tags + JoyCaption descriptions), calculates sane step counts for your dataset size, and trains with settings hand-tuned for one model — Anima. That focus is the whole point.
You watch it learn. Preview images render every epoch while training runs. Flip on Compare epochs to see them side by side, spot the earliest epoch that already nails your character, and stop right there — no over-training guesswork, no re-runs. Stopped too soon or crashed? Resume from the last checkpoint.
It has opinions, so you don't need any. Pick a preset — Person, Object / Concept, or Style — and the subject handling, optimizer, network size, and step budget are set. Power users can save their own presets, switch optimizers (Prodigy auto-LR or classic AdamW8bit), and preview the exact training config before launch. Guard rails catch the classics before they burn a run: empty captions (one-click fix), low VRAM, even duplicate images hiding in your dataset.
It finishes the job. One click copies the trained LoRA into your Forge or ComfyUI models folder — with the trigger word baked into the filename, so anyone you share it with knows exactly what to type. A batch queue captions and trains multiple LoRAs unattended, and survives a restart.
Runs on your machine. Your images, captions, and LoRAs never leave your computer. No subscriptions, no queues, no uploads.
WHAT'S NEW — v2.9.3
• Stray USE_LIBUV in your own Windows environment? Fix it without leaving AnimaForge — Setup → Training Environment adds a checkbox that forces AnimaForge's setting even when your environment already has one, plus a button that clears it from Windows outright. No more manual System Properties surgery
• See exactly what's in play — a live status line in Setup shows the effective USE_LIBUV and where it's coming from, so you're never guessing whether the fix applies to you
• Reset Training Environment — one click re-verifies PyTorch from scratch when something about your setup looks stale, without losing your last known-good result unless the fresh check genuinely can't confirm it
• Clear now tells you whether it worked — success or failure is logged right there instead of the button just going grey
Full history: github.com/SillySilk/AnimaForge
WHAT YOU'LL NEED
• Windows + NVIDIA GPU (CUDA) — includes RTX 50-series. No CPU/AMD path.
• ~16 GB VRAM comfortable; low-VRAM mode reaches smaller cards (8 GB practical floor).
• The three Anima files (Setup auto-detects them): DiT checkpoint, Qwen3 text encoder, Qwen-Image VAE — from Hugging Face circlestone-labs/Anima.
GET IT → github.com/SillySilk/AnimaForge
install.bat (one-time — builds a self-contained .venv with the full stack), then launch.bat.
100% free and open source (MIT). Built for the Anima community. Forge something great.
------------------------------ ② VERSION DESCRIPTION (v2.9.3) — see changelog-2026-08-03.md ------------------------------
v2.9.3 — self-service fixes for the training crash
• Stray USE_LIBUV in your own Windows environment? Fix it without leaving AnimaForge — Setup → Training Environment adds a checkbox that forces AnimaForge's setting even when your environment already has one, plus a button that clears it from Windows outright. No more manual System Properties surgery.
• See exactly what's in play — a live status line in Setup shows the effective USE_LIBUV and where it's coming from, so you're never guessing whether the fix applies to you.
• Reset Training Environment — one click re-verifies PyTorch from scratch when something about your setup looks stale, without losing your last known-good result unless the fresh check genuinely can't confirm it.
• Clear now tells you whether it worked — success or failure is logged right there instead of the button just going grey.
Full detail: github.com/SillySilk/AnimaForge
------------------------------ ③ VERSION HISTORY (bottom of description) ------------------------------
(Insert a CODE BLOCK via the </> toolbar button, then paste the table so the columns hold.
If you'd rather not use a code block, use Variant B in version-history.md instead.)
VERSION DATE WHAT CHANGED
------- ---------- ------------------------------------------------------------
v2.7 2026-07-10 LM Studio removed entirely — captioning is fully local (WD14
+ JoyCaption, merged mechanically), no external LLM server
needed or contacted · fixes the bug where some .nl captions
randomly came out as raw tags instead of prose · Combine now
de-dupes and tidies booru tags with no model in the loop ·
Describe gains "Redo all" — re-caption everything with a
live per-image stream · steadier shutdown
v2.6 2026-07-10 Optional "check for updates at startup" (Setup) notifies you
when a newer build has landed · the check runs safely off the
UI thread · icon/caption fixes
v2.5 2026-07-09 Captioning never overwrites existing caption files without
asking (keep / overwrite / set it once) · caption runs resume
where they stopped, and a killed training run comes back even
if you moved its folder · the batch queue captions each set
then trains it, settling every conflict up front · Add to
Batch is back on the front page; Restart from Top re-runs a
finished queue · find/replace rules fix misgendering and ban
tags · re-running Auto-Tag no longer duplicates every tag
v2.3 2026-07-06 Fixes a "division by zero" crash when aspect-ratio bucketing
met an image with a side under 64 px (it killed the run at
exit code 1) — now caught before launch with a one-click fix,
and auto-handled on unattended/batch runs · corrected
epoch/preview step math · live sample-preview UI overhaul
v2.2 2026-07-04 Pick your own UI font (Setup → App Defaults): forge faces,
system font, or any installed font · fixes non-Latin
readability (CJK, Cyrillic) · applies live, no restart ·
default look unchanged
v2.1 2026-07-03 Project autosave under the LoRA name at both caption
milestones · Restore captions from Saved Sets · low-VRAM
warning auto-continues after 10 s on unattended runs ·
in-app updater (Updates on the front page)
v2.0 2026-07-02 New brand badge across the app · preset picker shows each
intent's step math · one launch point (Options configures,
the front launches) · config preview beside Start ·
monospace version digits · spacing polish
v1.9 2026-07-02 Training presets (Person/Object/Style + save your own) ·
compare epochs side-by-side · Deliver to ComfyUI · trigger
word in the delivered filename · duplicate-image catcher ·
readable UI text (CJK-friendly) · accurate speed/ETA dials ·
previews every epoch · side-by-side installs stay separate
v1.1 2026-07-01 The forge redesign: everything runs from one screen, with
analog Epoch/Loss/Speed/ETA dials · optimizer presets
(Prodigy / AdamW8bit) · RTX 50-series support · empty-caption
guard with one-click fix · config preview · per-epoch
preview schedule on the progress bar
v1.0 2026-06-30 Launch: one-click caption→train for Anima (WD14 + JoyCaption)
· auto step counts tuned to dataset size · batch queue that
survives restarts · crash recovery + resume · low-VRAM mode
· Name Cast for consistent character triggers
Description
<!--
CIVITAI VERSION UPDATE — AnimaForge v2.3 (bucketing crash fix)
HOW TO USE: Civitai → your AnimaForge model → Add Version.
• Version name: v2.3
• File: marketing/Zip file 2026-07-06/AnimaForge.zip (built from
public main @ f6cf0da, app version 2.3.0 — replace the old zip)
• Paste ONLY the short blurb below as the version description.
• Also swap the "What's new" section of the model description and add the
v2.3 row to the version-history code block.
-->
## v2.3 — bucketing that doesn't blow up
- **Fixes the "division by zero" crash** — an image with a side under 64 px made aspect-ratio bucketing fail and killed the whole run at exit code 1. AnimaForge now catches it *before* launch and tells you exactly which files are too small.
- **One-click fix** — turn bucketing off (those images get center-cropped) and keep going, or bow out to enlarge/remove them first.
- **Walk-away safe** — unattended and batch runs auto-handle it so a stray thumbnail can't strand an overnight queue.
- **Truer progress** — corrected epoch/preview step math so the dials and preview schedule line up with what's actually happening.
- **Nicer previews** — live sample-preview UI overhaul.
Full detail: **github.com/SillySilk/AnimaForge**
Comments (12)
It's an amazing UI - thank you!
About Anima training - is using lower (512x512) and/or higher (1536/1536 and <1536 buckets) resolutions of the source images worth to achieve better Anima LoRA quality? What about the "@" character before artist style in the base model - does it have a meaning in Anima LoRA training?
the @ Is a style in anima. The program is designed based off of deep research using AI from two different sources as far as the structure of Anima. I will check to see about the resolution presets and make sure that those are changeable so people can experiment with different variations. Those were the settings that were discovered to be the most optimal but that doesn't mean that people don't have different case usage that might require different settings.
@PanPDX Thank you for reply! The UI is very comfortable to use, especially when VRAM size is an issue. I love the calculator, too, showing, with no BS, how many times the image will be "seen" by the procedure, how many steps and epochs will the training take... And it's VERY soothing to the eyes. :)
@mikmod I'm definitely noticing a few problems, at least on my screen. I think I'm going to reduce the preview images to three so they better fit for some reason it has the first epoch as zero, which is fine because it hasn't actually been trained on anything yet, but then they make the next one two and never call one epoch one. I don't really use those low epoch so it doesn't really matter but it just bugs me a little. The Batch functionality is wired in but I don't even see the buttons to make it work. I think I lost that in one of the upgrades so I'm going to need to wire that back in.
@PanPDX Thanks for the work you're doing. :)
I'd love a warning about .txt filed being overwritten, when user points at a folder with images AND existing tag files. :)
And also, in the hindsight, the 'jobs' system, similar to this trainer. It will probably require a bit of re-coding of the entire project though. ^_^
I've also noticed it's hard to fix the tagging model's 'misgender-ism'. I'd love to have 'search and replace' function for tags, with the ability to mark the images, which will be re-tagged. :)
I will try and fold that into this. Update should be done in about an hour.
@PanPDX I'm testing 2.5, and I noticed one bug - when in dataset section I click 'Validate Names', the captions I had get overwritten with the name of the character and the keyword is written inside the new txt files (as I renamed the images and .txt outside the program, before starting it). :)
@mikmod thank you. That's exactly the kind of thing I'm looking for! Myself I only validate names if I'm doing a character and because I use the Anima sorter before I almost never have a bad file name. If you've captioned it from another program, yep, I could see how that would do it. I will add that into the fix but right now I have 2.6 about coming out, which just has some cosmetic fixes and a few little tweaks. I'll see if I can add this to 2.7. Actually fuck it. I'll fix that right now and hopefully that'll be in 2.6.
@PanPDX There's a TagGUI program (not updated since October 2025, but it still works), that supports tagging with various models, including JoyCaption one. It also has searching for empty tags, repeated tags, has simple query capabilities; BUT it doesn't have a search for mutually exclusive tags. :)
Like, for example "black hair, blonde hair", or "censored, uncensored" tags in the same .txt file. I've seen such tagging with the tagging model your program, and TagGUI uses. :)
@PanPDX I had the captions done with TagGUI program, as it's the nicest one I can find, outside yours GUI (as it uses the same models, as TagGUI). :)
The file names were 'original' I had, hence I didn't give much attention to them, until now. Hence my renaming 'scheme' with Total Commander and its Ctrl+M function.
@PanPDX I wonder about one thing - would it be possible to have your GUI save the LoRA it is training, when the loss is at it's lowest in a given timeframe or epoch? I believe it's best for the LoRA, if the loss is the lowest it can be, right?
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.