MOSS-Image Gouache LoRA GGUF

Separate F16 GGUF conversion of the colorful vintage-gouache LoRA from axiomofmind/Moss-image-gouache-lora. It is designed to be applied at inference time to axiomofmind/Moss-Image-SFT-GGUF, keeping the adapter strength adjustable without creating a merged model for each base quantization or scale.

The style favors painterly shapes, warm editorial color, simplified forms, textured scenes, and bold illustrative compositions. This repository contains the LoRA adapter only. It does not contain the base diffusion transformer, Qwen3.5 text encoder, tokenizer, or VAE.

File

The adapter uses rank 16 and alpha 16. Its trigger phrase is moss_vgouache_style; place it at the beginning of the prompt. Scale 1.0 was selected after local comparisons of checkpoints 1000, 1500, and 2000 at scales 0.7, 1.0, and 1.2.

Required components

• A compatible base transformer from Moss-Image-SFT-GGUF. Q8_0 was validated; Q4_K_S uses the same MOSS architecture and LoRA targets.
• A MOSS-enabled stable-diffusion.cpp build.
• A Qwen3.5-9B text encoder, the official MOSS tokenizer.json, and the official MOSS / FLUX.2 VAE. See the base GGUF repository for links and setup notes.

Usage

Put the base GGUF and this LoRA GGUF in accessible locations. The <lora:...> tag selects the adapter file and scale; moss_vgouache_style is the learned trigger sent to the text encoder.

sd-cli.exe `
  --diffusion-model MOSS-Image-SFT-Q8_0.gguf `
  --llm text_encoder/model.safetensors.index.json `
  --vae vae/diffusion_pytorch_model.safetensors `
  --tokenizer tokenizer/tokenizer.json `
  --lora-model-dir . `
  --lora-apply-mode at_runtime `
  --cfg-scale 3 `
  --steps 30 `
  --sampling-method euler `
  --scheduler discrete `
  --flow-shift 5 `
  --seed 42 `
  --rng cpu `
  --sampler-rng cpu `
  -W 1024 -H 1024 `
  -p "moss_vgouache_style, white farmhouse with a red roof sits on a golden hillside surrounded by pink and orange trees, with a winding path through colorful wildflowers leading toward distant mountains under a soft pink sky. <lora:MOSS-Image-Gouache-LoRA-F16:1.0>" `
  -o gouache-gguf.png

Quantized transformers automatically select runtime LoRA application. --lora-apply-mode at_runtime is explicit here because it is the validated path. Ready-to-edit PowerShell and Bash wrappers are provided under examples/.

Controlled validation

The adapter was tested with MOSS-Image-SFT-Q8_0.gguf. Both images below used the same description, base model, text encoder, VAE, 1024×1024 resolution, 30 steps, CFG 3, Euler/discrete sampling, flow shift 5, and seed 42.

• The stock prompt omitted both the trigger phrase and LoRA tag.
• The adapter prompt added moss_vgouache_style and <lora:MOSS-Image-Gouache-LoRA-F16:1.0>.
• stable-diffusion.cpp reported all 504 / 504 LoRA tensors applied.

Timings were measured on an NVIDIA RTX PRO 6000 Blackwell Max-Q and are correctness checks, not general performance benchmarks.

Conversion

Converted independently from pytorch_lora_weights.safetensors using stable-diffusion.cpp commit b13a482:

sd-cli.exe -M convert `
  -m pytorch_lora_weights.safetensors `
  -o MOSS-Image-Gouache-LoRA-F16.gguf `
  --type f16 `
  --convert-name `
  -t 16

The source adapter was trained against MOSS-Image-Base and selected at 1,500 steps from a 2,000-step run. Its transformer targets are compatible with MOSS-Image-SFT, which produced the preferred evaluation results. See conversion.json for machine-readable conversion and validation metadata.

License

The adapter and this conversion are provided under Apache License 2.0, matching the source release and base model. This is an independent community conversion and is not affiliated with or endorsed by OpenMOSS.