A three-branch MOSS-Image SFT comparison for exploring a standard 50-step run against 8-step DMD and 4-step accelerated paths. Each active branch has its own sampler and SaveImage output.
Main features:
- One MOSS-Image SFT INT8 ConvRot loader with Qwen3.5 9B BF16 text encoder and FLUX.2 VAE.
- The 50-step branch uses CFG 3; the 8-step and 4-step branches use CFG 1. All three use Euler, the simple scheduler, and fixed seed 4242.
- The 8-step branch connects MOSS_DMD_8STEP_QKV_FORGE_FIX.safetensors; the 4-step branch connects moss_image_sft_4step_v2.0.safetensors. Both LoRAs are enabled at strength 1.
Suggested workflow:
1. Install the SFT model, text encoder, VAE, and both acceleration LoRAs, then load the JSON.
2. Run the three connected branches with the same prompt and seed.
3. Compare their saved images and actual run times on your own hardware; change prompts to test the trade-off across subjects.
The included example prompt describes a cliffside greenhouse. This graph has three reachable outputs; it does not supply benchmark timings or guarantee a particular speedup.
RunningHub AI workflow: https://www.runninghub.ai/zh-cn/post/2113906849406201857?inviteCode=rh-v1111
RunningHub signup: https://www.runninghub.ai/?inviteCode=rh-v1111
Bilibili video: https://www.bilibili.com/video/BV1ZHHP6FEVZ/
Related model files: https://pan.quark.cn/s/20e19307e8ef
Description
A three-branch MOSS-Image SFT comparison for exploring a standard 50-step run against 8-step DMD and 4-step accelerated paths. Each active branch has its own sampler and SaveImage output.
Main features:
- One MOSS-Image SFT INT8 ConvRot loader with Qwen3.5 9B BF16 text encoder and FLUX.2 VAE.
- The 50-step branch uses CFG 3; the 8-step and 4-step branches use CFG 1. All three use Euler, the simple scheduler, and fixed seed 4242.
- The 8-step branch connects MOSS_DMD_8STEP_QKV_FORGE_FIX.safetensors; the 4-step branch connects moss_image_sft_4step_v2.0.safetensors. Both LoRAs are enabled at strength 1.
Suggested workflow:
1. Install the SFT model, text encoder, VAE, and both acceleration LoRAs, then load the JSON.
2. Run the three connected branches with the same prompt and seed.
3. Compare their saved images and actual run times on your own hardware; change prompts to test the trade-off across subjects.
The included example prompt describes a cliffside greenhouse. This graph has three reachable outputs; it does not supply benchmark timings or guarantee a particular speedup.
RunningHub AI workflow: https://www.runninghub.ai/zh-cn/post/2113906849406201857?inviteCode=rh-v1111
RunningHub signup: https://www.runninghub.ai/?inviteCode=rh-v1111
Bilibili video: https://www.bilibili.com/video/BV1ZHHP6FEVZ/
Related model files: https://pan.quark.cn/s/20e19307e8ef