All stories
MOSS-Image Character Lab Launches Community reports

MOSS-Image gets a browser character lab with four switchable LoRAs

A new community ZeroGPU app puts Alya, Emilia, Frieren and Marin adapters beside a Base/SFT switch, adjustable strength and a session contact sheet.

Sources (4)

Why now Hugging Face records the Space's creation on October 22 at 08:35 UTC and the adapter collection at 08:47 UTC, establishing newly public repositories for this character workflow.

MOSS-Image now has a community character playground that puts four LoRAs and two model checkpoints behind browser controls. Tachiyumi’s new Hugging Face Space offers Alya, Emilia, Frieren and Marin adapters, with adjustable strength and a contact sheet that restores a selected frame’s settings.

Hugging Face records the Space’s creation on October 22 at 08:35 UTC and its companion adapter repository at 08:47 UTC. Those are the public repositories’ creation times, not proof of when each adapter was trained. This is a separate community workflow, rather than a new OpenMOSS base-model release.

Pick a character, then choose Base or SFT

The English documentation describes rank-32 adapters trained with ai-toolkit against MOSS-Image-Base. Users can switch to the aesthetic-tuned SFT checkpoint, but the author explicitly calls that an approximate landing for adapters trained on Base.

The suggested workflow is natural-language prompting with a character trigger, rather than a long string of appearance tags. The app adds a missing trigger automatically. Its documented defaults include 50 steps, CFG 3.0 and a one-megapixel square canvas; the recommended adapter strength is 1.0 except for Emilia at 0.85.

According to the author, adapters apply as forward deltas without merging, allowing strength changes without reloading weights. The browser controls expose those choices in one place, making this a practical entry point for experimenting with the supplied character recipes.

GPU time is part of the workflow

The Space runs on ZeroGPU and processes one run at a time. Its reservation ticket estimates the GPU time needed before starting, with a documented default ceiling of 120 seconds. The author gives two 1024-by-1024 images at 50 steps as an approximately 150-second request that the default configuration refuses; reducing image count, canvas or steps is the suggested adjustment. These are the project’s estimates, not an AITrending speed test or a guaranteed service time.

The README also estimates roughly 35 GB of resident memory for its bf16 setup. Browser access should not be mistaken for a claim that the same configuration fits a small local GPU.

Check rights separately from access

The adapter collection says these are not official OpenMOSS releases and that its source character frames remain the property of their rights holders. The Space metadata and English documentation also differ in their code-license labels. Access to the app therefore does not establish one permissive license covering code, model weights and character adapters.

AITrending read the English root documentation and inspected the running app’s controls. It did not generate images or independently test character consistency, memory consumption or reservation accuracy.