All stories
MOSS-Image-Edit-HQ Momentum Community reports

MOSS-Image-Edit-HQ recommends v5 over its newer experiment

The community image-editing LoRA publishes a practical ComfyUI choice, with 50-step final edits and an 8-step draft mode that trades away precision.

Sources (3)

Why now The October 22 project appeared in our October 27 Hugging Face trending discovery, alongside fresh October 26-27 workflow and comparison uploads. This is a current checkpoint recommendation, not a new upstream model launch or proof of measured growth.

For ComfyUI users trying MOSS-Image-Edit-HQ, the author’s current recommendation is armV5, even though newer experimental artifacts are available. The unofficial MOSS-Image-SFT editing LoRA’s English card says armV6 failed to improve its comparisons: some edits stayed similar, while color casts appeared and an object-addition case regressed. That makes the recommended checkpoint a more useful starting point than simply choosing the highest version number.

The project was created October 22 and appeared in our October 27 Hugging Face trending discovery. Its October 26-27 history includes ready workflows and fresh comparison uploads. Those are reasons to revisit the adapter, not evidence that the underlying MOSS model launched today or that the project has achieved a measured rate of growth.

Two installation routes matter. The native ComfyUI route uses the ai-toolkit-format LoRA, a build containing native MOSS editing support, and the experimental Text Encode MOSS-Image Edit node with edit_format = ai-toolkit. The alternative Chenkin-x plugin route uses the separate armV5-moss-edit-lora.safetensors file in models/model_patches/, with ready single-reference and multi-reference workflows. The card names ComfyUI v0.42.3 as its tested plugin baseline. Mixing those formats or folders can produce loading errors.

The card also separates final quality from quick iteration. Its recommended final-edit recipe uses 50 steps and CFG 3. A separate community DMD speed LoRA can be stacked for an 8-step, CFG 1 draft mode, but the author reports softer object additions and worse color containment. These are the author’s visual comparisons, not AITrending benchmark results.

This is not a small standalone model download: the base model and pipeline components are still required. The maintainer estimates roughly 20GB or more for the bf16 pipeline and explicitly rules out a 6GB VRAM setup. The page also contains older, conflicting version-table and example references, so use the top-level armV5 recommendation and the matching file format when choosing what to test. The adapter is a community fine-tune, with no claimed OpenMOSS endorsement.