Experiment with style changes by applying a rank-256 SFT difference LoRA to the MOSS-Image SFT base workflow. This is an editable text-to-image recipe for comparing visual character, not a separate anime model.
Main features:
- Active MOSS-Image SFT INT8 ConvRot model and moss_image_sft_lora_rank_256_bf16.safetensors at strength 1.49.
- Qwen3.5 9B BF16 text encoder, FLUX.2 VAE, and a 1024 x 1024 latent canvas.
- Euler sampler at 25 steps, CFG 3, simple scheduler, fixed seed 9527; one SaveImage output.
Suggested workflow:
1. Install the SFT model, rank-256 LoRA, text encoder, and VAE, then load the JSON.
2. Generate with the included prompt and strength 1.49 to establish a baseline.
3. Change prompt and LoRA strength in controlled steps to study color, texture, and style differences.
LoRA rank alone does not imply better image quality. The older note inside the JSON mentions strength 1.0, but the connected LoRA is set to 1.49. The paired cover is promotional artwork, not a measured output of this JSON.
RunningHub AI workflow: https://www.runninghub.ai/zh-cn/post/2114399087264362498?inviteCode=rh-v1111
RunningHub signup: https://www.runninghub.ai/?inviteCode=rh-v1111
Bilibili video: https://www.bilibili.com/video/BV1UYnU6qEGK/
Related model files: https://pan.quark.cn/s/9ec095b95838
Description
Experiment with style changes by applying a rank-256 SFT difference LoRA to the MOSS-Image SFT base workflow. This is an editable text-to-image recipe for comparing visual character, not a separate anime model.
Main features:
- Active MOSS-Image SFT INT8 ConvRot model and moss_image_sft_lora_rank_256_bf16.safetensors at strength 1.49.
- Qwen3.5 9B BF16 text encoder, FLUX.2 VAE, and a 1024 x 1024 latent canvas.
- Euler sampler at 25 steps, CFG 3, simple scheduler, fixed seed 9527; one SaveImage output.
Suggested workflow:
1. Install the SFT model, rank-256 LoRA, text encoder, and VAE, then load the JSON.
2. Generate with the included prompt and strength 1.49 to establish a baseline.
3. Change prompt and LoRA strength in controlled steps to study color, texture, and style differences.
LoRA rank alone does not imply better image quality. The older note inside the JSON mentions strength 1.0, but the connected LoRA is set to 1.49. The paired cover is promotional artwork, not a measured output of this JSON.
RunningHub AI workflow: https://www.runninghub.ai/zh-cn/post/2114399087264362498?inviteCode=rh-v1111
RunningHub signup: https://www.runninghub.ai/?inviteCode=rh-v1111
Bilibili video: https://www.bilibili.com/video/BV1UYnU6qEGK/
Related model files: https://pan.quark.cn/s/9ec095b95838