v1.3.18 - launcher fixes

Full Changelog: v1.3.17...v1.3.18

v1.3.17 - Rename pypi packages

Full Changelog: v1.3.16...v1.3.17

v1.3.16 - Character Finetuning

What's Changed

• Finetune Character node, reference re-testing, and wardrobe scoring by @imprsnst in #66

Full Changelog: v1.3.15...v1.3.16

v1.3.15 - Character: API Nodes(Nano Banana, Seedance & Minimax H3)

What's Changed

• character: support for api nodes, minimax, seedance & nanobanana by @imprsnst in #64

Full Changelog: v1.3.14...v1.3.15

v1.3.14

What's Changed

• feat: workflow popup & import in-app by @imprsnst in #63

Full Changelog: v1.3.13...v1.3.14

v1.3.13 - Character body & clothing references

What's Changed

• Characters carry body and clothing references, not just face
• Compile References gains Body and Cloth ports; roles are stored in the .char, so existing files still load
• MiniMax H3 divides its nine reference slots 2:1:1 face/body/cloth, with a per-render budget and a face-only filter
• Generation nodes show one slot for the render that has not landed: editing, rendering, cancelled or failed
• That slot holds the settings it ran with, so browsing take history no longer loses them
• Take history is protected server-side; a stale browser write could previously drop a finished render
• MiniMax H3 memory: the render is staged in four phases, so the VAE is only on the card for the two that read it
• Quantised weights are sized by their real storage, which was over-reporting the H3 denoiser by 2x
• Load Assets nodes resize to their content
• Fal-backed nodes get their own icon

Character body and clothing references, and MiniMax H3 memory fixes by @ashish-aesthisia in #62

Full Changelog: v1.3.12...v1.3.13

v1.3.12 - Edit dataset name and trigger word

Full Changelog: v1.3.11...v1.3.12

v1.3.11 - Minimax H3 .char support

What's Changed

• Added character support for all minimax nodes t2v, i2v & r2v
• Support for NVFP4 text encoder
• Memory fixes
• Minimax H3: Default to FP8 & NVFP4 for efficient resource usages
• Fix resource node

MiniMax H3 character support by @imprsnst in #60

Full Changelog: v1.3.1...v1.3.11

v1.3.1 - training & installation fix

Changes

• Renamed clip length to max clip length
• Fix: Import HF dataset from sub dirs
• App update: Outdated warning & auto update frontend on --install --extras all

Full Changelog: v1.3.0...v1.3.1

v1.3.0 - Merged trainer tab into generation, node graph refactor

What's new in v1.3.0?

Sumamry

• Merged training & generation tab
• Node graph updates: Run progress, export graph, new nodes
• Models: Models are stored now in Registry, missing model popup
• Workflows: New workflows page for exported node graphs & detailed guide on how to run.
• Unified Character: Single .char now supports both flux family & Krea2

Related Workflows

• Flux 2: Portable consistent characters, without LoRA training
• Flux 2 + Krea 2: Multi model portable consistent characters, Training only
• Krea 2: Generate consistent images with unified .char model
• Flux 2 Klein: Generate consistent images with unified .char model

sample .char model is available on Huggingface

Characters

• A character is a .char file holding one identity plus a payload per model, so adding a model
  adds a payload instead of rebuilding the character
• Seven nodes replace the character side panel: Encode Character, Compile References,
  Character to Dataset, Attach Adapter, Write .char, Load Character, Edit Character
• Encode Character stores references, crops each face with YuNet, and computes SFace (128-d) and
  DINOv2 (768-d) embeddings per reference
• Krea 2 applies a character through its trained adapter, which is the only route on a model with
  no reference channel
• FLUX.2 applies one through compiled references, or through an adapter when the character has one
• Attach Adapter reads the architecture, base, rank, steps and resolution from the adapter's own
  metadata, and has a Strength setting that rides in the file
• The description doubles as the trigger: it is stored in the file, used as the caption on every
  training image, and prepended to the prompt at generation
• Character encoders are pickable per node from models/annotators/ instead of being resolved in
  the background

One canvas

• The Trainer tab is gone. Load Dataset, Caption, Train LoRA and Graph are nodes on the Studio canvas
• One Run walks the whole graph in order: encode, build a dataset, train, attach, write
• Train LoRA outputs both the trained adapter and its step/loss series, so it wires into Attach
  Adapter and into the loss curve
• Cancelling a graph stops the trainer at its next checkpoint and leaves the run resumable

Models

• Missing-models popup/
• A model registry contain most of the models required by the app.
• Hugging Face token can be set in the app settings for gated repos

Graph export and import

• Recipe format v2: each param states its type, and each node lists the models it needs with a
  download link
• A graph rebuilds on a machine that has none of the models, with the popup offering every one
• Params driven by a wire are left to the loader driving them, so a graph names each file once
• PNG metadata carries the same recipe, so a dropped image reports its missing models
• Training nodes survive export and re-import
• New workflow page for exported required workflows

Canvas

• Run state is per node: green follows what is executing, red stays on what stopped the run
• Search in the add-node menu, across all tabs
• Port dots aligned, coloured by kind, and labelled on hover
• Minimum node heights derived from port count, so a dot never sits outside the card
• Node settings sidebar no longer closes on its own first keystroke
• Params can appear on the node face as well as in Adjust

Fixes

• Write .char kept only the last payload wired to it, dropping a trained adapter and cutting the
  training branch out of the run
• A trained adapter reached Attach Adapter as a path relative to the models root, so it read as an
  adapter with no provenance
• FLUX.2 checkpoint problems are surfaced instead of failing mid-load
• A queued model download never started
• The registry served a cached list indefinitely

Upgrading

cd Inline-Studio
git pull
cd core
./webui.sh --install --extra all   # .\webui.bat on Windows

What's Changed

• Refactor: Portable characters & one canvas for training by @imprsnst in #57

Full Changelog: v1.2.74...v1.3.0