Fizgig adds per-image loss tracking, auto-recaptioning with Qwen3-VL, and resume-safe LoRA training
shootthesound · reddit · 2026-07-24
Fizgig, a Krea 2 training tool, adds a set of practical features for image-model finetuning and dataset management.
- Per-image loss tracking with adaptive handling: each image gets its own status such as easy, suspect, stuck, or exhausted.
- Auto-recaptioning: stuck samples are rewritten mid-run by Qwen3-VL, re-encoded, and retried.
- Automatic exclusion after repeated failed recaption attempts, with safety rails to prevent healthy images from being dropped.
- Dataset intelligence includes ArcFace-based look-consistency filtering, gentle warm-up for unusual-but-real images, and live preview scoring.
- Operational features include pause/resume with zero quality loss, fp8 residency tuned for modest GPUs, context LoRA stacking, repair sliders, and ComfyUI-compatible output.
More from coding & agent
- TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27
- Precompute golden answers and match at runtime for exact wording control — colinmcnamara · 2026-10-27
- Dev argues there's still alpha in working with your coding agent interactively — alex_frantic · 2026-10-27
- Researchers launch AMA on patterns for safer AI agents used by citizen developers — Loose_Rip359 · 2026-10-27
- Gradio Adds In-Chat Human-in-the-Loop So ML Intern Pauses Before Training — Gradio · 2026-10-27
- Building Halo Infinite Forge maps with local LLMs, DCjson and Blender MCP — swagonflyyyy · 2026-10-27