2026-10-21: Bria releases Fibo 2.0 (Oct 20): an 8B model that does text-to-image and editing in one set of weights, plus Edit Turbo for 4-step edits with masks and up to 4 references. #1 open-weights editing model on Artificial Analysis. License is CC BY-NC 4.0, so commercial use needs a paid Bria agreement. ComfyUI supports it natively on master; LoRA training works with Bria's own script, the diffusers DreamBooth scripts (open PR) and AcademiaSD LoRAlab (from 8 GB). ai-toolkit, OneTrainer, musubi-tuner, kohya and SimpleTuner have no support yet · huggingface.co/briaai/Fibo-2.0-dev · huggingface.co/briaai/Fibo-2.0-edit-turbo · github.com/Comfy-Org/ComfyUI/pull/17240 · github.com/huggingface/diffusers/pull/15084 · guide
2026-10-21: Fibo 2.0: BriaFibo2Pipeline is now in diffusers main (no release yet, so install from source); the diffusers LoRA scripts PR is still open. Bria's Fibo-2.0-dev repo has its own single-GPU LoRA trainer for text-to-image and edit LoRAs. First LoRAs on Dev show likeness by about step 300, JSON and plain captions give the same result, and realism trains better than anime. Edit mode is weak at small edits and tends to redraw the whole head, so describe the hair · github.com/huggingface/diffusers/pull/14991 · huggingface.co/briaai/Fibo-2.0-dev/blob/main/fine_tuning/REA · guide
2026-10-21: MOSS Image speed LoRAs: the Base 8-step DMD2 LoRA is final (update 3000; stock Euler, 4-8 steps, CFG 1), and Kromoss gets an 8-step DMD2 LoRA (4 steps also work, CFG 1). To judge your own LoRA's likeness, test without them: 50 steps, CFG 2, LoRA strength ~1.5, at 1.5 MP or more · huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd · huggingface.co/lodestones/Kromoss · guide
2026-10-20: kohya sd-scripts (released in v0.12.2) can cache several caption variants per image for MOSS Image and picks one at random each step; before, only the first was kept. Other models now stop with an error on multi-variant captions instead of silently using the first one · github.com/kohya-ss/sd-scripts/pull/2466 · guide
2026-10-19: Kromoss (Apache 2.0): a full fine-tune of MOSS Image Base that fixes Base's blurriness and drops in wherever MOSS Image SFT is used. Ships an 8-step turbo LoRA (8 steps, CFG 1), a ComfyUI conversion script and a ~5.4 GB int8 file for smaller GPUs. Still being updated (update 11000 on Oct 20) · huggingface.co/lodestones/Kromoss · guide
2026-10-19: OneTrainer: 20-30% faster LoRA training on 8-bit quantized models (FLUX.2 16 GB preset 1.30x, Krea 2 16 GB preset 1.20x) and faster 8-bit math on RTX 50-series cards. Merged into the preview-2 branch, not master yet; DoRA does not get the speed-up · github.com/Nerogar/OneTrainer/pull/1680 · github.com/Nerogar/OneTrainer/commit/f9e9778 · guide