Images Instead of Prompts and Transparent PNG in MOSS-Image

Two weeks after the release of the open model MOSS-Image, third-party developers released a node that feeds image tokens into the Qwen3.5-9B encoder instead of a text prompt, a LoRA for PNG with an alpha channel directly from the sampler, and a sharpness slider.

Compiled by Sergey KostenchukPublished 2026-10-29Updated 2026-10-29

2026-10-29 Coding HuggingFace

🛠 Images Instead of Prompts and Transparent PNG in MOSS-Image

Third-party tools for the open model MOSS-Image (conditioning only via the Qwen3.5-9B encoder): the cond2img node feeds image tokens into the encoder instead of text, up to 16 images with masks, and the paralaif LoRA outputs PNG with an alpha channel directly from the sampler.

🌍 The ecosystem around the open Apache-2.0 model is growing rapidly: image conditioning via the text channel, transparency directly in sampling, and regional prompting for up to 5 character-LoRAs. These are demonstrations, not a measured comparison with closed editors.

👤 Works in ComfyUI: cond2img is useful when training others' LoRAs, transparent PNG eliminates the need for chroma keying, and sharpness is adjusted by the LoRA strength from 1 to −1 (noticeable only on MOSS Base). For commercial use, it is worth checking the license of the RGBA VAE.

Source 1: https://github.com/thezveroboy/ComfyUI-MossTextEncoder Source 2: https://huggingface.co/paralaif/MOSS-Manga-Alpha-Image-LoRA

Sources