Fizgig 6.8.1 ships Krea 2 slider LoRAs, Ultra mode, and Qwen Image 2.1 full fine-tuning
shootthesound · reddit · 2026-10-02
Open-source image training tool Fizgig released 6.8.1 with major updates:
- Slider LoRAs on Krea 2: one LoRA acts as a dial between two looks (sad ↔ happy), trainable from a few photo pairs or three prompts. A new Ultra mode trains only composition blocks, letting sliders run cleanly at strength 20 in ComfyUI.
- Full fine-tuning for Qwen Image 2.1: joins Krea 2 and MiniMax H3. Runs on a single consumer GPU (16 GB works, 24/32 GB recommended), taking 3-4x as long as the Ultra Fast Krea 2 LoRA preset. A single-character fine-tune took 30 minutes on a 5090; multi-character datasets take a few hours. Converting to a LoRA via Checkpoint to LoRA gives far better concept separation.
- Repair Studio: no strength limit, bigger sliders with ±0.1 nudge, per-block donor LoRA mixing.
- New driver system: describe a model once and training, sliders, Repair Studio, Profiler and Extract all work with it.
Open source on GitHub with PRs welcome.
More from Multimodal
- Physics YouTuber automates his own videos: Claude scripts, voice-clones, and edits end-to-end — danbri · 2026-10-02
- One prompt, 8h 21min: Claude produces an 8-minute documentary, 96 takes, fully autonomous — GCWebDesigner · 2026-10-02
- ComfyUI Ships Comfy Agent to All Comfy Cloud Users: AI Plans, Builds and Fixes Workflows — cpaik · 2026-10-02
- Open-source .char format packs face, clothes and body into one portable consistent character file — ashishsanu · 2026-10-02
- One prompt recreates a 2021 bobcat-attack CCTV clip as a cinematic video — aitrendz_xyz · 2026-10-02
- YUI: first animated short made with Comfy Agent, a three-minute apocalypse story — PurzBeats · 2026-10-02