MOSS Serenity

MOSS Serenity is a full finetune of MOSS-Image Base that teaches the model a set of people by name, while keeping what the base model already knows. The current release, v6, was trained for 210,000 steps on a single RTX 5090 with maltrainer, in seven stages that each added more people (Versions).

It is a drop-in replacement for MOSS-Image Base: the same architecture, text encoder (Qwen3.5-9B) and VAE (FLUX.2), so any MOSS-Image workflow in ComfyUI runs it unchanged.

The v6 gallery shows all 179 newest samples (faces, showcases, styles) with their prompts and shared settings. More in samples/: v6 (faces, showcases, styles), v5 and v4 (film portraits of earlier models), v3 and v2 (portraits and 50-step scenes), portraits (8-step street portraits), scenes (8-step film scenes), scenes-50steps (Full HD, 50 steps), pairs (two people), trios (three people) and no-people.

Files

The text encoder and VAE are MOSS-Image's own: Qwen3.5-9B (qwen3.5_9b_int8_convrot.safetensors, or qwen3.5_9b_bf16.safetensors) in models/text_encoders, and the FLUX.2 VAE (flux2_vae.safetensors) in models/vae.

Versions

Each version continues the previous one from its saved state (weights, optimizer and EMA), with more people added. The people in the Characters table are the ones trained up to v6.

Later versions know everyone earlier ones did, but a person trained long ago and rehearsed little since can be weaker in a later version than in the one where they were new: compare two versions on the same seed if one person matters to you (samples/v4, samples/v5).

Quick start (ComfyUI)

1. Put serenity_base_v6_210000_int8_convrot.safetensors (the latest, half the size and faster; or serenity_base_v5_170000_bf16.safetensors for bf16) in ComfyUI/models/diffusion_models/.
2. Drag one of the workflows into ComfyUI and pick the model in Load Diffusion Model.
3. Prompt photo of <trigger>, a woman, ... (see Prompting).

Shift needs no node: ComfyUI runs MOSS-Image at shift 5. Full details in docs/generating.md.

bf16 or INT8 ConvRot? Same model, same workflows: pick the file in Load Diffusion Model. The INT8 file stores each block's big Linears as int8 in a Hadamard-rotated basis with one scale per row (ComfyUI's INT8 ConvRot, as in Comfy-Org's moss_image_*_int8_convrot releases), so it is half the size and runs on comfy-kitchen's int8 kernels. Its weights are within 0.9% of the bf16 ones (mean relative error per Linear), and on the same seed it renders a near-identical image: the turbo workflow took 8.7 s including loading, against ~16 s for bf16, on an RTX 5090. Keep the bf16 file for finetuning, merging and LoRA extraction.

Prompting

The model learned each person from captions of one shape, so prompts work best in that shape:

photo of <trigger>, a <woman|man>, <one sentence: look, framing, clothing, setting, light>

photo of gilliananderson, a woman with shoulder-length straight auburn red hair, a serious focused
expression, medium close-up from the chest up, holding a small flashlight raised beside her face,
wearing a dark charcoal 1990s pantsuit jacket over a white blouse, standing in a dark abandoned
warehouse corridor at night, cold blue moonlight, 1990s film grain, cinematic film still

• Keep faces large. Chest-up and closer gives the best likeness; full-body shots with a small face can come out wrong.
• Describe the look, not the title. The model knows the people, not their films: "a long brown leather duster coat in the cargo bay of a worn frontier spaceship" works, a film title does not.
• Two people: a man and a woman work well; two of the same gender are hit-or-miss. See Two people in one shot and Three people.
• Without a trigger, photo of a woman, ... gives an ordinary person, as in the base model.

Characters

Each trained person is prompted with a trigger: their name in lower case, without spaces, put first after photo of, then the class word they were trained with:

photo of keanureeves, a man, ...

The table lists everyone the model was trained on up to v6 (210,000 steps). Samples marked v6, v5 and v4 are film portraits from those models; v2 and v3 samples come from the earlier versions; a - means there is no sample in this repository yet.

The samples below were made with the workflows in workflow/: (50) marks the full model at 50 steps and 1920 × 1088; the rest are the 8-step turbo workflow at 1584 × 1776.

Scenes without people, at 50 steps: alpine valley, macro dragonfly, Tokyo alley, rural village, vintage cars.

Two people in one shot

The training data had no group shots, yet two triggers in one prompt work well for a man and a woman: each keeps their own face. Name both with their class word, put them on a side, and describe each separately:

photo of nathanfillion, a man, and morenabaccarin, a woman, two people in the cockpit of a worn
frontier spaceship: on the left nathanfillion, a man with short dark hair and a wry smile, wearing a
brown leather duster coat over a red shirt; on the right morenabaccarin, a woman with long dark hair
pinned up in curls, wearing a flowing silk robe in deep jewel tones, ...

Samples (50 steps): Firefly, helicarrier, Dune, ship's bow 1, ship's bow 2, diner.

Two people of the same gender are hit-or-miss. Some work: the stadium duet and the palace audience keep two distinct women. Others don't: in this library scene the second woman has the right braids, collar and dress, yet a face pulled toward the first, and two men on a beach do the same (the second loses most of his likeness). Both triggers share a class word (a woman or a man), which is likely where they meet. Costumes and hair always stay apart; the faces are what can drift. Try several seeds, or render each person separately.

Three people in one shot

The same rule holds with three: faces stay distinct across class words, and drift within one.

• One woman, two men (a rainy city alley): all three keep their own faces.
• One man, two women (a spaceship cargo bay): the man is solid; the two women keep their costumes and hair, but their faces move partly toward each other.
• Three women (a castle dining hall): braids, red hair and house ties are all right, yet all three faces converge on the strongest identity of the three.

So for groups, mix genders where you can, and render same-gender groups with several seeds, or person by person.

The film scenes describe each role's look (costume, hair, setting) rather than naming the film, which the model does not know: see docs/generating.md.

The LoRAs

The lora-extract files hold the finetune as a LoRA for MOSS-Image Base: the SVD of each layer's change, so they load on the base model and stack with other LoRAs. Use them when you'd rather keep one base model and switch LoRAs. The workflow lora-extract-turbo-lora.json is that setup: moss_image_base_bf16 → serenity_v6_lora_r256 (1.0) → the DMD turbo LoRA → 12 steps, CFG 4, euler / simple.

They keep only part of each layer's change (by size, r64 drops 96%, r128 93%, r256 87%): at batch size 1 much of a full finetune's change is step-to-step noise that no low rank captures. That noise may not matter for likeness, but compare a LoRA with the full model on the same seed before relying on it. The full model keeps everything.

Training

Full finetune of MOSS-Image Base with maltrainer, summarised here and in full in docs/training.md:

Responsible use

This model can produce the likeness of real people. Do not use it to deceive, to impersonate, to harass, or to make sexual or demeaning images of anyone. Label generated images as generated.

License

Apache 2.0, as MOSS-Image. The DMD turbo LoRA used in the turbo workflows is lodestones' (lodestones/MOSS-Image-Base-8-steps-dmd) and is not part of this repository.