Fine-Tuned Gemma-4 Extends Prompts for Qwen-Image-Edit, Trained with Kimi as Reward Model
pmttyji · reddit · 2026-08-19
New HF release Wuli-art/Gemma-4-for-Qwen-Image-Edit-2511-Prompt-Extend: a prompt-extension model fine-tuned from gemma-4-12B-it that takes an editing prompt plus input images and generates an enhanced editing prompt, helping Qwen-Image-Edit-2511 better understand editing intent and produce more stable results. Trained with Prompt Extend Reinforcement Learning (PERL) using the ROLL framework, with Kimi K2.6 as the reward worker evaluating edit quality.
More from Multimodal
- Prompt template: "Temporal Glitch Loop" traps subjects in repeating visual corruption — LudovicCreator · 2026-08-19
- SpAItial generates high-quality 3D worlds from iPhone recordings — risingodegua · 2026-08-19
- AI Crosses SpongeBob with Attack on Titan, Surprisingly Epic Results — eyishazyer · 2026-08-19
- Stimma 1.0.13 releases with ComfyUI manager integration and LTX-2.5 video support — stimma · 2026-08-19
- Two days on a 5090: first Minimax H3 short via ComfyUI hits YouTube — Routine_Ad_3391 · 2026-08-19
- Kuaishou's Kling AI tops 850M yuan quarterly revenue, up over 200% YoY — 智东西 · 2026-08-19