CivArchive
    Senko-san_NAI-V5-Curated_HD_ULTRA (仙狐さん) - Base V1 - 2000 steps
    Preview 149880568
    Preview 149878853
    Preview 149881237

    High-fidelity character LoRA for Senko-san from The Helpful Fox Senko-san, trained on OpenMOSS MOSS-Image-SFT.

    Built from 1,000 native 1024×1024 NovelAI 5.0 Curated generations, this LoRA brings the Senko-san HD ULTRA dataset to MOSS Image. It captures her amber eyes, short layered golden-blonde hair, fox ears, fluffy golden tail, and hair ornament, with clean anime linework and flexible clothing and scene prompts.

    This ULTRA version combines the 598 base compositions used in my NovelAI 4.5 Curated series with 402 activity scenes for wider scene and pose coverage. The images for this dataset were generated with NovelAI 5.0 Curated.

    The Trigger Word

    senko_san
    

    Quick Generation Settings

    • Base checkpoint — OpenMOSS-Team/MOSS-Image-SFT
    • LoRA version — SFT V1, final 2,000-step checkpoint
    • LoRA weight — 1.0
    • Prompt style — Descriptive natural language with senko_san
    • Sampler — Flow-matching Euler
    • Denoising steps — 50
    • CFG scale — 3
    • Flow shift — 5
    • Unconditional branch skip layer — 22, using zero-based indexing
    • Negative prompt — Empty
    • Resolution — 1024 × 1024
    • Example seed — 38141
    • Text encoder — Qwen3.5-9B, BF16
    • VAE — OpenMOSS bundled small VAE, FP32

    These are the settings used for the saved previews. Start with LoRA weight 1.0, then adjust it to suit your prompt. Use a MOSS workflow that exposes the flow shift and unconditional branch skip settings.

    Prompting

    Describe clothing, pose, composition, lighting, and environment in natural language. The trigger senko_san anchors the character. The examples below are the exact prompts from the training previews, each generated at seed 38141 with 50 denoising steps, CFG 3, and LoRA weight 1.0. All four used an empty negative prompt.

    1. Teal Raincoat in a Courtyard

    senko_san, full-body anime illustration, taking a step across a stone courtyard in a teal raincoat and plain boots, fluffy fox tail visible, overcast light.
    

    Negative prompt left empty.

    2. Window-Light Portrait

    senko_san in a head-and-shoulders anime illustration, smiling gently in window light, both fox ears visible, clean line art and flat cel shading.
    

    Negative prompt left empty.

    3. Setting Out Rice and Chopsticks

    senko_san setting a bowl of rice and chopsticks on a low wooden table with both hands, medium-wide view, warm morning light, clean anime line art.
    

    Negative prompt left empty.

    4. Sharing Tea

    senko_san on the left, sharing tea with an adult human woman with short black hair on the right, two distinct faces, cozy room at dusk, cel-shaded anime illustration.
    

    Negative prompt left empty.

    Dataset and Captions

    Training used all 1,000 image-caption pairs from Senko HD ULTRA V5 Captions, with images at their native 1024×1024 resolution. Each image has a matching .txt caption containing a descriptive natural-language caption. Captions were kept intact, with caption shuffling disabled and caption dropout set to zero.

    Dataset available on Hugging Face.

    Training Specifications

    • Training engine — ostris/ai-toolkit 0.13.42
    • Starting checkpoint — OpenMOSS-Team/MOSS-Image-SFT
    • Model architecture — MOSS Image
    • Adapter — Standard LoRA
    • Trigger word — senko_san
    • Training steps — 2,000 optimizer updates
    • Batch size — 2
    • Gradient accumulation — 1
    • Approximate dataset passes — 4
    • Network rank / alpha — 32 / 32
    • Optimizer — Automagic
    • Configured learning rate — 0.0001
    • LR scheduler — Cosine
    • Training precision — BF16
    • Diffusion model weights — Unquantized BF16
    • Text encoder — Frozen, unquantized BF16 Qwen3.5-9B
    • VAE computation — FP32, OpenMOSS small VAE
    • Saved LoRA precision — BF16
    • Gradient checkpointing — Enabled
    • Maximum gradient norm — 1.0
    • Training scheduler — Flow matching with shifted timestep sampling
    • Training resolution — 1024 × 1024
    • Training image-caption pairs — 1,000
    • Caption shuffling — Disabled
    • Caption dropout — 0
    • Horizontal flipping — Disabled
    • Latent and text embedding caching — Enabled
    • Checkpoint and preview interval — Every 250 training updates
    • Training hardware — NVIDIA RTX 6000 Ada, 48 GB, on RunPod

    The uploaded adapter is the final SFT V1 checkpoint at step 2,000, selected after comparing all four preview prompts across the saved checkpoints and the two MOSS Base runs. The rice and tea scenes showed especially clear action and interaction in this version.

    Description

    Base V1 — final 2,000-step checkpoint

    The original HD ULTRA Base V1 character LoRA for Senko-san, trained on OpenMOSS-Team/MOSS-Image-Base using all 1,000 native 1024×1024 NovelAI V5 Curated HD ULTRA image-caption pairs. Select this version when using a MOSS-Image-Base generation workflow.

    Trigger word

    senko_san
    

    Training settings

    • Starting checkpoint — OpenMOSS-Team/MOSS-Image-Base.
    • Dataset — Senko HD ULTRA V5 Captions, comprising 598 core character compositions and 402 activity scenes.
    • Training engine — ostris/ai-toolkit 0.13.42.
    • Training schedule — 2,000 optimizer updates, batch size 2, gradient accumulation 1, approximately four dataset passes.
    • Adapter — standard LoRA, rank 32, alpha 32.
    • Optimizer — Automagic, configured learning rate 0.0001, cosine scheduler.
    • Training computation — BF16, with the diffusion model and frozen text encoder loaded as INT8 ConvRot weights.
    • Text encoder — Qwen3.5-9B, frozen during training.
    • VAE — FLUX.2 full VAE, used in BF16 in this original run.
    • Exported adapter — BF16 safetensors.
    • Caption handling — natural-language captions kept intact, shuffling disabled, caption dropout 0.
    • Horizontal flipping — disabled.
    • Gradient checkpointing — enabled, maximum gradient norm 1.0.
    • Checkpoint and preview interval — every 250 training updates.
    • Training hardware — NVIDIA RTX 6000 Ada 48 GB on RunPod.

    Teal raincoat example

    The attached image is the original preview saved at training step 2,000. It used the following prompt.

    senko_san, full-body anime illustration, taking a step across a stone courtyard in a teal raincoat and plain boots, fluffy fox tail visible, overcast light.
    
    • Generation checkpoint — MOSS-Image-Base.
    • LoRA weight — 1.0.
    • Sampler — flow-matching Euler.
    • Preview denoising steps — 30.
    • CFG scale — 3.
    • Flow shift — 5.
    • Unconditional branch skip layer — 22, using zero-based indexing.
    • Resolution — 1024×1024.
    • Seed — 38141, fixed across preview prompts.
    • Negative prompt — empty.

    The adapter is the original Base V1 release. The saved example records its original 30-step preview settings. The 2,000 training updates are the optimizer schedule for the LoRA.