MOSS-Image Edit HQ, a second community edit LoRA, reaches v6 with workflows and a demo
MOSS-Image Edit HQ is a free, community-made LoRA that teaches the open MOSS-Image-SFT model to edit a photo from a text instruction, with a focus on keeping everything you did not ask to change. Its author publishes every training round, including the ones that lost, plus two ComfyUI workflows and a demo you can try in the browser.
What it does
It covers attribute changes, object removal, relighting, facial and text edits on one image. v5 (25 October) raised the reference budget from 144 to 256 tokens and trained with masks so unedited areas stay close to the source; v6, now the latest, adds captions that describe the target state with a separate edit instruction and a region-weighted loss for tighter preservation. Weights ship as an ai-toolkit LoRA that loads natively in ComfyUI and as a file for the Chenkin-x ComfyUI-Moss-Image-Edit node pack.
Results and limits
The card's recommended settings are Euler, shift 5.0, 50 steps and CFG 3.0 (3.5 to 4 for weak edits), output up to 1024 px. Stacking the community 8-step DMD speed LoRA makes edits about six times faster, but the author's own verdict is that the 50-step pass clearly wins: at 8 steps added objects weaken and a color cast creeps in, so the fast path is for drafts. The first training round lost to chinoll's earlier edit LoRA on identity, which the card says openly.
Worth knowing
It needs about 20 GB of VRAM in bf16 and does not run on 6 GB. It is trained on MOSS-Image-SFT (Base and Kromoss are unverified) and on single-image edits, so putting a person from one photo into another is not what it does. The native ComfyUI route needs a current nightly build, because the edit node merged after the v0.42.5 release. It is an unofficial Apache 2.0 fine-tune, not an OpenMOSS release.
Try it yourself
fix the overexposed lighting, natural balanced exposure, keep everything else the same