v2026.07.13 — Beta

Everything that landed since v2026.07.12: first-run fixes from your bug reports, a new FLUX.1 training family, expert training controls, and a big cloud-training maturation pass.

✨ Headline

⚡ FLUX.1 training family (local, early)

Train character LoRAs on FLUX.1-dev alongside Z-Image / SDXL / Krea 2. Prose captions, best-practice defaults from ai-toolkit's own reference (flowmatch, rank/alpha 16, guidance 4 / 20-step previews); the trained LoRA deploys straight into loras/flux. FLUX.1-dev is a gated 12B model, so ~24 GB VRAM is the comfort zone (drop to 768 to fit smaller cards). Local training first; cloud + in-app Test Studio for Flux follow.

🔬 Expert "last-mile" training levers

For the last 10% on small datasets: decoupled alpha, network dropout, timestep weighting, optimizer (adamw8bit / adafactor / automagic / prodigy), LR schedule + warmup, and effective batch (gradient accumulation) — each with a plain-English why/how, all defaulting to the current behaviour so nothing changes unless you touch them.

☁️ Cloud training, matured

• Concurrent runs with a configurable cap, a monthly budget guardrail and a worst-case cost confirm.
• Host quality layer — stop renting the flakiest cheap hosts: reliability floor, disk-bandwidth filter, bait-price exclusion, and a short blacklist for machines that die mid-run.
• Resume-capable checkpoint downloads (HTTP Range retry) with size verification, and the newest checkpoint is mirrored locally during the run.
• A dedicated Cloud runs hub page (watch / stop / download every run in one place) + honest per-tier GPU prices with a cap warning.

🔄 Automatic update detection

The app now checks for new releases and surfaces an update available banner + a Check for updates button next to Settings, with one-click update & restart.

🐛 First-run fixes (thanks for the bug reports, @GroxicTinch 🙏)

• Dataset generation now runs on a stock ComfyUI — the image workflow used a custom easy int node a fresh ComfyUI doesn't have, so the prompt was rejected and no image was produced (which also looked like "generated images aren't saved to the dataset folder"). Now uses the native PrimitiveInt. (#1, #2)
• "Train" no longer fails silently — if the ai-toolkit run crashed on start, the panel went quiet after the green "Training started" toast. It now shows an error banner with the exit code, the log tail, and the common first-run causes (venv packages, or the gated base model still downloading / needing a HF token). (#3)

🎛️ Also

• Test Studio strength choices now extend down to 0 (a proper LoRA-off control column).
• Guide: per-model steps & convergence guidance — where "good results" start for each family.

Full changelog: v2026.07.12...v2026.07.13