anr2me/klein-build-small-starter

Update UI + add missing i2i example outputs; fix 21:9 preset

Refactor: drop _LORA_LOADED global (stateless load/unload, fixes cross-tab LoRA bleed); hoist imports to module top

README: 4B-only framing, klein family as reference

4B-only: drop multi-model machinery; fix description font (override Quicksand with Inter)

README: clarify generate picker is distilled-only

Inference picker = distilled only (4B+9B); bases stay as LoRA-training targets

README: document all four klein variants

Expose full klein 2x2 family (4B/9B x distilled/base), on-demand download, cap-1 RAM

Prefetch 9B weights in background to avoid GPU-budget timeout on first 9B call

README: document 4B+9B model variants

Add 9B model toggle + iPhone-style 9B LoRA

Frame ai-toolkit as one community trainer of several, not the only one

Add About & links tab + GitHub/docs links in footer, README, and Train tab

Drop the 'Off the Grid badge' line from the header

Image→Image: add clickable examples (bundled photos + restyle prompts) so people can try i2i in one click

LoRA tab: show before→after example gallery for image-edit LoRAs so people see what they do

Fix i2i LoRA example preload: return PIL image not hf-cache path (Gradio serve-path error)

LoRA tab: correct per-card triggers + t2i/i2i types; image-edit LoRAs reveal an input box pre-filled with the repo example

Add peft to requirements — diffusers needs the PEFT backend for load_lora_weights (fixes LoRA tab crash)

Plainer copy: drop the LLM-ish phrasings (no-cloud-API flourish, demo-judges-remember, em-dash punchlines)

Make Train-a-LoRA tab scannable: 4 collapsible steps + full guide tucked in an accordion

Surface training guide as its own '🛠️ Train a LoRA' tab (rendered from TRAIN_A_LORA.md)

Add Hub LoRA dropdown (zero-setup) + training path intro (RunPod/UI/Modal)

FLUX.2 [klein] 4B Build Small starter — t2i / i2i / LoRA-compare, ZeroGPU

initial commit