MOSS-Image LoRA training fits under 6GB VRAM at ~1s/step on RTX 5080
AcademiaSD · reddit · 2026-10-17
AcademiaSD LoRAlab Trainer Studio added LoRA training for the Apache 2.0 open text-to-image model MOSS-Image, its 12th supported model family alongside Qwen-Image 2.1, FLUX.2 Klein 9B, Z-Image, SDXL and others.
Key details:
- Trains on MOSS-Image Base or SFT, which share a text encoder so one pre-cache works for both
- The 5B transformer runs in NF4: 1s/step at 512x512, rank 8, under 6GB VRAM on an RTX 5080
- Text encoder (Qwen3.5-9B) defaults to BF16 CPU offload, with INT8/NF4 options
- darask0's 4-step turbo adapter cuts previews from 43s (28 steps + CFG) to 4s, unmerged so it never touches the training LoRA
- Auto-captions via Qwen3-VL-8B with exact copying of visible text
- LoRAs load directly in ComfyUI nightly with the standard loader
- Defaults: 500 steps, LR 0.0004, cosine with 100 warmup; good results around step 300 in testing
Download is 9.9–20.2GB depending on text encoder choice (original model is 27.2GB), with Windows/Linux support, a RunPod template, and 8 interface languages.
More from Multimodal
- MOSS-Image, an Apache-2.0 image generation model, trends on Hugging Face — Comfy-Org · 2026-10-18
- Stunning AI short film by Douyin creator MetaAI echoes 'Ripples of Life' style — xiaosun86 · 2026-10-18
- Studio H3: open-source AI film studio with full director workflow — Exciting-Income-5840 · 2026-10-18
- 'GoPro video' in prompts: a simple trick for realistic everyday AI footage — Puzzleheaded_Ebb8352 · 2026-10-18
- Kling AI 4.0's Omni Reference blends up to 15 references for cinematic AI scenes — Meer_AIIT · 2026-10-17
- Bindu Reddy: everyone is fine-tuning models for 3D games, DeepSeek will do it 30x cheaper — bindureddy · 2026-10-17