latent-diffusion-upscale n-steps

Signed-off-by: Vladimir Mandic <mandic00@live.com>

CHANGELOG.md

@@ -48,6 +48,7 @@ And if you're a ROCm user, this release brings much faster compile times on Linu
       supports text, image and video prompts with or without input image  
       *note*: if input image is provided, model should be left at default `gemma-3-4b-it` as most other LLMs do not support hybrid workflows  
 - **Fixes**
+  - Latent Diffusion Upscale
   - Model load: support SDXL safetensors packaged without VAE  
   - ROCm: disable cuDNN benchmark, fixes slow MIOpen tuning with `torch==2.7`  
   - Extensions: use in-process installer for extensions-builtin, improves startup performance  

modules/postprocess/sdupscaler_model.py

@@ -55,7 +55,7 @@ class UpscalerDiffusion(Upscaler):
             'prompt': '',
             'negative_prompt': '',
             'image': img,
-            'num_inference_steps': 20,
+            'num_inference_steps': shared.opts.upscaler_latent_steps,
             'guidance_scale': 7.5,
             'generator': generator,
             'latents': None,

modules/shared.py

@@ -884,6 +884,7 @@ options_templates.update(options_section(('postprocessing', "Postprocessing"), {
 
     "postprocessing_sep_upscalers": OptionInfo("<h2>Upscaling</h2>", "", gr.HTML),
     "upscaler_unload": OptionInfo(False, "Unload upscaler after processing"),
+    "upscaler_latent_steps": OptionInfo(20, "Upscaler latent steps", gr.Slider, {"minimum": 4, "maximum": 100, "step": 1}),
     "upscaler_tile_size": OptionInfo(192, "Upscaler tile size", gr.Slider, {"minimum": 0, "maximum": 512, "step": 16}),
     "upscaler_tile_overlap": OptionInfo(8, "Upscaler tile overlap", gr.Slider, {"minimum": 0, "maximum": 64, "step": 1}),
 }))