← ImageLoRA · All LoRA guides

MOSS Image LoRA training with kohya sd-scripts: setup guide

Setting up kohya sd-scripts to train a MOSS Image LoRA: support status, the settings known to work, and links to the official docs. Status in kohya sd-scripts: supported (v0.12.1, Oct 18; LoRA incl. ConvRot INT8).

About kohya sd-scripts

The long-standing kohya training scripts (also used by the kohya_ss GUI); strongest for SD1.5/SDXL-era models. Sources: github.com/kohya-ss/sd-scripts

VRAM

Official (kohya sd-scripts guide; 1024px, batch 1, LoRA dim 16, AdamW, gradient checkpointing, text encoder outputs cached): ~13 GB training with bf16 DiT, ~9 GB with ConvRot INT8 DiT + text encoder, ~5-7 GB adding block swap (needs ~15 GB system RAM). Caching text encoder outputs needs ~8.3 GB even in INT8, so 8 GB GPUs are not supported yet. Training the text encoder LoRA too: ~17 GB INT8, ~32 GB bf16. SimpleTuner ships 24, 32, 48, 80 and 144 GB recipes at 512px. Community report (ai-toolkit): ~15 GB at batch 2, 1024px, AdamW [Discord community reports (AI Toolkit / banodoco / lodestone_rock servers), Oct 2026]

Settings that work for MOSS Image (any trainer)

• Learns fast [Discord community reports (AI Toolkit / banodoco / lodestone_rock servers), Oct 2026]
• Tokenizer loading bug fixed same day (Oct 16) in ai-toolkit; update before training
• SFT variant: stripe artifacts reported, unresolved [Discord community reports (AI Toolkit / banodoco / lodestone_rock servers), Oct 2026]
• sd-scripts: fp8_base is not supported for MOSS Image; use --convrot_int8 (and --convrot_int8_text_encoder) instead
• sd-scripts: --timestep_sampling sigma --weighting_scheme logit_normal --logit_mean -1.09 --logit_std 0.95 --discrete_flow_shift 1.0 reproduces pre-training; simpler alternative: --timestep_sampling shift --discrete_flow_shift 3.0
• sd-scripts example: LR 1e-4 constant at network_alpha 1 (lower LR if you raise alpha)
• SimpleTuner recipes: rank-32 LoRA, LR 1e-4, REPA, 4,000 steps, 512px-area training; validate at 50 steps, CFG 3, 1024px
• Edit/reference LoRAs: SimpleTuner model_flavour sft-edit or base-edit with matching target and reference folders; ai-toolkit also trains Base and SFT as edit models
• Base LoRA with ai-toolkit defaults (adamw8bit, shift timesteps, rank 32) gave a good style by ~1,500 steps; Base-trained LoRAs also work on SFT [Discord community reports (AI Toolkit / banodoco / lodestone_rock servers), Oct 2026]
• May need a higher LoRA rank than usual (smaller hidden size than Krea 2); character LoRAs reported as hard to get right so far [Discord community reports (AI Toolkit / banodoco / lodestone_rock servers), Oct 2026]
• Full fine-tunes on small datasets collapsed after a few epochs; use LoRA unless you have a very large dataset [Discord community reports (AI Toolkit / banodoco / lodestone_rock servers), Oct 2026]
• ComfyUI "lora key not loaded": ai-toolkit now saves unfused q/k/v LoRA keys (Oct 17); users fixed it with the official Comfy-Org MOSS files [Discord community reports (AI Toolkit / banodoco / lodestone_rock servers), Oct 2026]
• ai-toolkit low_vram + cache_text_embeddings could OOM at startup because the text encoder stayed on the GPU; fixed in ai-toolkit on Oct 17, so update first
• Inference: an SFT LoRA at ~0.6 strength with the Base model at 30 steps, CFG 3 works well; 8-step DMD LoRAs exist for SFT and Base but can weaken LoRA effect [Discord community reports (AI Toolkit / banodoco / lodestone_rock servers), Oct 2026]

General tips

• Gated Hugging Face models: click 'Agree and access' on each model page, or you'll get 403 errors
• Steps scale with dataset size (steps = epochs x images at batch 1)
• Common safe start: LoKr with auto rank (~64), batch/grad accumulation 4 [Discord community reports (AI Toolkit / banodoco / lodestone_rock servers), Oct 2026]

Compare with other trainers

• ai-toolkit vs kohya sd-scripts for MOSS Image
• kohya sd-scripts vs SimpleTuner for MOSS Image
• MOSS Image LoRA with ai-toolkit
• MOSS Image LoRA with SimpleTuner
• MOSS Image (overview)

Where to rent a 16 GB+ GPU

Cheapest 16 GB+ cards (2026-10-16 14:43 UTC) from RunPod and Vast.ai. Prices move constantly; see the live price table.

Sources

• huggingface.co/OpenMOSS-Team/MOSS-Image-SFT
• huggingface.co/OpenMOSS-Team/MOSS-Image-Base
• github.com/ostris/ai-toolkit/commit/e05a3a4
• github.com/ostris/ai-toolkit/commit/c6a9b56
• github.com/ostris/ai-toolkit/commit/29f5ff4
• github.com/kohya-ss/sd-scripts/blob/main/docs/moss_image_train_network
• github.com/kohya-ss/sd-scripts/releases/tag/v0.12.1
• github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/MOSS_
• github.com/bghira/SimpleTuner/releases/tag/v4.10.0
• github.com/ostris/ai-toolkit/commit/b7cd8a1
• github.com/ostris/ai-toolkit/commit/82e6311
• github.com/ostris/ai-toolkit/commit/9a1828f
• github.com/ostris/ai-toolkit/commit/8cf43dd
• huggingface.co/Comfy-Org/MOSS-Image
• huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd
• huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd
• github.com/kohya-ss/sd-scripts

FAQ

Does kohya sd-scripts support MOSS Image?

Yes: supported (v0.12.1, Oct 18; LoRA incl. ConvRot INT8) (see sources below).

How much VRAM do I need for a MOSS Image LoRA?

Official (kohya sd-scripts guide; 1024px, batch 1, LoRA dim 16, AdamW, gradient checkpointing, text encoder outputs cached): ~13 GB training with bf16 DiT, ~9 GB with ConvRot INT8 DiT + text encoder, ~5-7 GB adding block swap (needs ~15 GB system RAM). Caching text encoder outputs needs ~8.3 GB even in INT8, so 8 GB GPUs are not supported yet. Training the text encoder LoRA too: ~17 GB INT8, ~32 GB bf16. SimpleTuner ships 24, 32, 48, 80 and 144 GB recipes at 512px. Community report (ai-toolkit): ~15 GB at batch 2, 1024px, AdamW [Discord community reports (AI Toolkit / banodoco / lodestone_rock servers), Oct 2026]

Data updated 2026-10-16. Disclosure: RunPod and Vast.ai links are referral links; ImageLoRA may earn a commission at no cost to you. unknown = not yet verified from a source.