CivArchive
    ← All articles
    Published October 23, 2026by 7err4

    Fumika – a free local studio for comics, images and LoRA training

    177 views3 reactions1 comments on CivitAI7 collected
    comfyuimangaannouncementcomictool

    Write a short story, get finished comic pages. Fumika is a free desktop app for Windows and Linux (NVIDIA GPU) that turns a few lines of text into manga / comic pages with panels and speech bubbles – rendered on your own PC. Around it: an image studio, character assets, and trainers for your own LoRAs, so your cast looks the same on every page.

    What it does

    • Series and chapters. A project is a series – cast, style and model settings shared by every chapter – with as many chapters as you like, just like Civitai Comics. Each chapter can get a drawn cover with the title on it, the reader shows a chapter or the whole series, and every chapter exports as PDF or CBZ (with ComicInfo.xml for Komga, Kavita, YACReader, Mihon).

    • Story → pages. One line per panel: what happens | speech bubble. Fumika lays out the page with MOSS-Image, renders three versions and keeps the one with the cleanest lettering. A story editor with panel cards, scenes (place, light, who wears what) and reading direction left-to-right or right-to-left.

    • Characters stay themselves. Describe your cast once, add a character LoRA – or just a reference picture: CharRef v2 draws your main character from one image, no training needed.

    • Image Studio. Text to image, image to image, inpainting and upscaling (Real-ESRGAN, SeedVR2 or a MOSS redraw) with MOSS-Image, Anima (1.0, 2.9B, 3.8B) or SDXL / Illustrious / Pony. Every picture keeps its settings in the history.

    • Asset Studio. Expression sets, outfits, turnarounds, profile sheets, chibi and pixel sprites, transparent VN sprites.

    • Train your own LoRAs. Trainers for MOSS-Image, Anima and SDXL (LoRA, LyCORIS, embeddings) plus a slider trainer for Anima. Pick a folder, press Check (dataset, steps, memory estimate), then Start; compare saved epochs side by side.

    • Your models, any source. The library takes checkpoints and LoRAs from Civitai, Hugging Face or your own training and recognises the architecture from the file itself, whatever it is called.

    • Plugins. Pick what you want at first start; only those models are downloaded. Every model's licence is shown first.

    • Local and private. Fumika installs its own ComfyUI with its own Python into one folder and touches nothing else. A render queue that survives restarts and an optional local MCP server so AI agents can queue pages and pictures.

    • Marked as AI-generated. Every picture, PDF and CBZ carries the standard IPTC "AI-generated" metadata (EU AI Act Art. 50).

    A look inside

    A chapter of the example series: the drawn cover, five finished pages and the panel editor.

    Reader: one chapter or the whole series in order, with the series picture first.

    Each chapter gets a drawn cover with the series title and chapter number on it - three versions, pick one, use it as the series picture.

    Image Studio: text to image, image to image, inpaint and upscale with MOSS-Image, Anima or SDXL - every picture keeps its settings in the history.

    Asset Studio: nine expressions of the same character in one image (also outfits, turnarounds, profile sheets, chibi and pixel sprites).

    Trainer: pick a folder, choose the base model (Anima 1.0, 2.9B or 3.8B, MOSS-Image, SDXL), Check estimates steps and memory, then Start.

    Guided setup: pick your plugins - only their models are downloaded, every licence is shown first.

    The example story

    Fumika, the app's mascot, has seven pages due tomorrow. Her editor Sora is not amused. The pages were made with the app exactly as it ships: MOSS-Image Base, SFT look, Fumika's character LoRA, quality mode.

    CharRef v2 – a character from one picture

    The reference-picture option is its own LoRA, free under Apache-2.0 on Hugging Face. It also works in plain ComfyUI.

    Requirements

    • Windows 10/11 or Linux (x86-64), NVIDIA GPU – 24 GB VRAM recommended for MOSS-Image at full size; an int8 model for smaller cards is included

    • ~25 GB free disk space for engine + MOSS models (more for each further model family)

    Download & licence

    • GitHub: github.com/7err4/fumika – Windows installer (current user, no admin rights), Linux AppImage and .deb

    • Fumika is free to use, commercial work included – Apache License 2.0 with the Commons Clause: modify and share it freely, only selling the software itself is not allowed. What you make with it is yours. The bundled ComfyUI node packs are GPL-3.0. Models keep their own licences – the app shows them before every download (Anima uses the CircleStone Labs non-commercial licence).

    • The installer is not signed yet: Windows SmartScreen may warn – "More info" → "Run anyway".

    Feedback and bug reports are welcome on GitHub.

    by 7err4

    Archived from CivitAI · Updated October 23, 2026View source