<div data-youtube-video="https://youtu.be/UGZQmF2GuJM"><iframe src="https://www.youtube.com/embed/UGZQmF2GuJM" width="853" height="480" allowfullscreen></iframe></div> 
 
Restore and reinterpret an input image with MOSS-Image SFT, then create a larger reconstructed result with FLUX 3 Klein. This ComfyUI workflow saves both the MOSS wash result and the later enlarged image, so you can compare the two stages. 
 
Main features: 
- An input image is scaled for MOSS-Image SFT. A RunningHub LLM node uses Gemini 3.1 Flash-Lite Preview to inspect the image and write a grounded reconstruction prompt. The JSON Layout switch is enabled in this supplied file and can be changed to a normal caption. 
- The MOSS stage uses moss_image_sft_int8_convrot.safetensors, qwen3.5_9b_bf16.safetensors, and flux2-vae.safetensors. Its connected Euler sampler uses 20 steps, CFG 3, and the sgm_uniform scheduler. 
- The second stage uses RealESRGAN_x4plus.pth, scales the long edge to 3840 pixels while preserving aspect ratio, and reconstructs with flux3_image_klein_sd_int8_convrot.safetensors and qwen3vl_4b_int8_convrot.safetensors. Its Euler sampler uses 16 steps, CFG 1, and the simple scheduler. 
- A second Gemini 3.1 Flash-Lite Preview node describes the MOSS result for the FLUX reconstruction. Both LLM nodes are connected; these API calls may incur charges. 
- Separate SaveImage nodes produce the MOSS result and the enlarged FLUX result. No LoRA is connected. 
 
Suggested workflow: 
1. Install the named MOSS, FLUX, text encoder, VAE, and RealESRGAN files, along with the nodes used by the graph. Configure access for the two RH LLM calls before running. 
2. Replace the LoadImage input with your own picture. Review the editable prompt instructions and choose caption or JSON Layout mode. 
3. Run the workflow, inspect the MOSS output first, and then compare the final FLUX result. Verify small details and text manually: reconstruction can change them, and a 3840-pixel long edge does not imply a fixed 3840×2160 canvas. 
 
RunningHub AI workflow: https://www.runninghub.ai/zh-cn/post/2113992913247076354?inviteCode=rh-v1111 
RunningHub signup: https://www.runninghub.ai/?inviteCode=rh-v1111 
Bilibili video: https://www.bilibili.com/video/BV16WHM6LE5V/

MOSS-Image Restore + FLUX 3 Klein 4K

<div data-youtube-video="https://youtu.be/UGZQmF2GuJM"></div> 
 
Restore and reinterpret an input image with MOSS-Image SFT, then create a larger reconstructed result with FLUX 3 Klein. This ComfyUI workflow saves both the MOSS wash result and the later enlarged image, so you can compare the two stages. 
 
Main features: 
- An input image is scaled for MOSS-Image SFT. A RunningHub LLM node uses Gemini 3.1 Flash-Lite Preview to inspect the image and write a grounded reconstruction prompt. The JSON Layout switch is enabled in this supplied file and can be changed to a normal caption. 
- The MOSS stage uses moss_image_sft_int8_convrot.safetensors, qwen3.5_9b_bf16.safetensors, and flux2-vae.safetensors. Its connected Euler sampler uses 20 steps, CFG 3, and the sgm_uniform scheduler. 
- The second stage uses RealESRGAN_x4plus.pth, scales the long edge to 3840 pixels while preserving aspect ratio, and reconstructs with flux3_image_klein_sd_int8_convrot.safetensors and qwen3vl_4b_int8_convrot.safetensors. Its Euler sampler uses 16 steps, CFG 1, and the simple scheduler. 
- A second Gemini 3.1 Flash-Lite Preview node describes the MOSS result for the FLUX reconstruction. Both LLM nodes are connected; these API calls may incur charges. 
- Separate SaveImage nodes produce the MOSS result and the enlarged FLUX result. No LoRA is connected. 
 
Suggested workflow: 
1. Install the named MOSS, FLUX, text encoder, VAE, and RealESRGAN files, along with the nodes used by the graph. Configure access for the two RH LLM calls before running. 
2. Replace the LoadImage input with your own picture. Review the editable prompt instructions and choose caption or JSON Layout mode. 
3. Run the workflow, inspect the MOSS output first, and then compare the final FLUX result. Verify small details and text manually: reconstruction can change them, and a 3840-pixel long edge does not imply a fixed 3840×2160 canvas. 
 
RunningHub AI workflow: https://www.runninghub.ai/zh-cn/post/2113992913247076354?inviteCode=rh-v1111 
RunningHub signup: https://www.runninghub.ai/?inviteCode=rh-v1111 
Bilibili video: https://www.bilibili.com/video/BV16WHM6LE5V/

Files with SHA256: a8847136426da9bcb2e74d22fc917f08ff4d9adeee9deb4cfc4da1ae5457a06e