ComfyUI-SUPIR

SUPIR

• SUPIR_conditioner

  Creates the conditioning for the sampler. Caption input is optional, when it receives a single caption, it's added to the positive prompt. If a list of caption is given for single input image, the captions need to match the number of tiles, refer to the SUPIR Tiles node. If a list of captions is given and it matches the incoming image batch, each image uses corresponding caption.

• SUPIR_decode
• SUPIR_encode
• SUPIR_first_stage

  SUPIR "first stage" processing. Encodes and decodes the image using SUPIR's "denoise_encoder", purpose is to fix compression artifacts and such, ends up blurring the image often which is expected. Can be replaced with any other denoiser/blur or not used at all.

• SUPIR_model_loader

  Old loader, not recommended to be used. Loads the SUPIR model and the selected SDXL model and merges them.

• SUPIR_model_loader_v2

  Loads the SUPIR model and merges it with the SDXL model. Diffusion type should be kept on auto, unless you have issues loading the model. fp8_unet casts the unet weights to torch.float8_e4m3fn, which saves a lot of VRAM but has slight quality impact. high_vram: uses Accelerate to load weights to GPU, slightly faster model loading.

• SUPIR_model_loader_v2_clip

  Loads the SUPIR model and merges it with the SDXL model. Diffusion type should be kept on auto, unless you have issues loading the model. fp8_unet casts the unet weights to torch.float8_e4m3fn, which saves a lot of VRAM but has slight quality impact. high_vram: uses Accelerate to load weights to GPU, slightly faster model loading.

• SUPIR_sample

  - **latent:** Latent to sample from, when using SUPIR latent this is just for the noise shape, it's actually not used otherwise here. Identical to feeding this comfy empty latent. If fed anything else it's used as it is, no noise is added. - **cfg:** Linearly scaled CFG is always used, first step will use the cfg_scale_start value, and that is interpolated to the cfg_scale_end value at last step. To disable scaling set these values to be the same. - **EDM_s_churn:** controls the rate of adaptation of the diffusion process to changes in noise levels over time. Has no effect with DPMPP samplers. - **s_noise:** This parameter directly controls the amount of noise added to the image at each step of the diffusion process. - **DPMPP_eta:** Scaling factor that influences the diffusion process by adjusting how the denoising process adapts to changes in noise levels over time. No effect with EDM samplers. - **control_scale:** The strenght of the SUPIR control model, scales linearly from start to end. Lower values allow more freedom from the input image. - **restore_cfg:** Controls the degree of restoration towards the original image during the diffusion process. It allows for dome fine-tuning of the process. - **samplers:** EDM samplers need lots of steps but generally have better quality. DPMPP samplers work well with lower steps, good for lightning models. Tiled samplers enable tiled diffusion process, this is very slow but allows higher resolutions to be used by saving VRAM. Tile size should be chosen so the image is evenly tiled. Tile stride affects the overlap of the tiles. Check the SUPIR Tiles -node for preview to understand how the image is tiled.

• SUPIR_tiles

  Tiles the image with same function as the Tiled samplers use. Useful for previewing the tiling and generating captions per tile (WIP feature)

• SUPIR_Upscale