Change the performance field to GPU Memory Usage instead, and use the 'balanced' profile by default, since it's just 5% slower than 'high', and uses nearly 50% less VRAM

ui/media/js/auto-save.js

@@ -36,7 +36,7 @@ const SETTINGS_IDS_LIST = [
     "save_to_disk",
     "diskPath",
     "sound_toggle",
-    "performance_level",
+    "vram_usage_level",
     "confirm_dangerous_actions",
     "metadata_output_format",
     "auto_save_settings",

ui/media/js/main.js

@@ -602,7 +602,7 @@ function onTaskCompleted(task, reqBody, instance, outputContainer, stepUpdate) {
                             <b>Suggestions</b>:
                             <br/>
                             1. If you have set an initial image, please try reducing its dimension to ${MAX_INIT_IMAGE_DIMENSION}x${MAX_INIT_IMAGE_DIMENSION} or smaller.<br/>
-                            2. Try picking a lower performance level in the '<em>Performance Level</em>' setting (in the '<em>Settings</em>' tab).<br/>
+                            2. Try picking a lower level in the '<em>GPU Memory Usage</em>' setting (in the '<em>Settings</em>' tab).<br/>
                             3. Try generating a smaller image.<br/>`
                 }
             } else {
@@ -887,7 +887,7 @@ function getCurrentUserRequest() {
             width: parseInt(widthField.value),
             height: parseInt(heightField.value),
             // allow_nsfw: allowNSFWField.checked,
-            performance_level: perfLevelField.value,
+            vram_usage_level: vramUsageLevelField.value,
             //render_device: undefined, // Set device affinity. Prefer this device, but wont activate.
             use_stable_diffusion_model: stableDiffusionModelField.value,
             use_vae_model: vaeModelField.value,

ui/media/js/parameters.js

@@ -94,18 +94,18 @@ var PARAMETERS = [
         default: true,
     },
     {
-        id: "performance_level",
+        id: "vram_usage_level",
         type: ParameterType.select,
-        label: "Performance Level",
+        label: "GPU Memory Usage",
         note: "Faster performance requires more GPU memory<br/><br/>" +
+              "<b>Balanced:</b> almost as fast as High, significantly lower GPU memory usage<br/>" +
               "<b>High:</b> fastest, maximum GPU memory usage</br>" +
-              "<b>Medium:</b> decent speed, uses 1 GB more memory than Low<br/>" +
+              "<b>Low:</b> slowest, force-used for GPUs with 4 GB (or less) memory",
-              "<b>Low:</b> slowest, for GPUs with 4 GB (or less) memory",
         icon: "fa-forward",
-        default: "high",
+        default: "balanced",
         options: [
+            {value: "balanced", label: "Balanced"},
             {value: "high", label: "High"},
-            {value: "medium", label: "Medium"},
             {value: "low", label: "Low"}
         ],
     },
@@ -227,7 +227,7 @@ function initParameters() {
 
 initParameters()
 
-let perfLevelField = document.querySelector('#performance_level')
+let vramUsageLevelField = document.querySelector('#vram_usage_level')
 let useCPUField = document.querySelector('#use_cpu')
 let autoPickGPUsField = document.querySelector('#auto_pick_gpus')
 let useGPUsField = document.querySelector('#use_gpus')

ui/sd_internal/__init__.py

@@ -6,7 +6,7 @@ class TaskData(BaseModel):
     request_id: str = None
     session_id: str = "session"
     save_to_disk_path: str = None
-    performance_level: str = "high" # or "low" or "medium"
+    vram_usage_level: str = "balanced" # or "low" or "medium"
 
     use_face_correction: str = None # or "GFPGANv1.3"
     use_upscale: str = None # or "RealESRGAN_x4plus" or "RealESRGAN_x4plus_anime_6B"

ui/sd_internal/app.py

@@ -110,7 +110,7 @@ def setConfig(config):
     except:
         log.error(traceback.format_exc())
 
-def save_to_config(ckpt_model_name, vae_model_name, hypernetwork_model_name, performance_level):
+def save_to_config(ckpt_model_name, vae_model_name, hypernetwork_model_name, vram_usage_level):
     config = getConfig()
     if 'model' not in config:
         config['model'] = {}
@@ -124,7 +124,7 @@ def save_to_config(ckpt_model_name, vae_model_name, hypernetwork_model_name, per
     if hypernetwork_model_name is None or hypernetwork_model_name == "":
         del config['model']['hypernetwork']
 
-    config['performance_level'] = performance_level
+    config['vram_usage_level'] = vram_usage_level
 
     setConfig(config)
 

ui/sd_internal/device_manager.py

@@ -128,7 +128,7 @@ def needs_to_force_full_precision(context):
     device_name = context.device_name.lower()
     return (('nvidia' in device_name or 'geforce' in device_name) and (' 1660' in device_name or ' 1650' in device_name)) or ('Quadro T2000' in device_name)
 
-def get_max_perf_level(device):
+def get_max_vram_usage_level(device):
     if device != 'cpu':
         _, mem_total = torch.cuda.mem_get_info(device)
         mem_total /= float(10**9)
@@ -136,7 +136,7 @@ def get_max_perf_level(device):
         if mem_total < 4.5:
             return 'low'
         elif mem_total < 6.5:
-            return 'medium'
+            return 'balanced'
 
     return 'high'
 

ui/sd_internal/model_manager.py

@@ -25,9 +25,9 @@ DEFAULT_MODELS = {
     'gfpgan': ['GFPGANv1.3'],
     'realesrgan': ['RealESRGAN_x4plus'],
 }
-PERF_LEVEL_TO_VRAM_OPTIMIZATIONS = {
+VRAM_USAGE_LEVEL_TO_OPTIMIZATIONS = {
+    'balanced': {'KEEP_FS_AND_CS_IN_CPU', 'SET_ATTENTION_STEP_TO_4'},
     'low': {'KEEP_ENTIRE_MODEL_IN_CPU'},
-    'medium': {'KEEP_FS_AND_CS_IN_CPU', 'SET_ATTENTION_STEP_TO_4'},
     'high': {},
 }
 
@@ -125,9 +125,24 @@ def resolve_model_paths(task_data: TaskData):
     if task_data.use_upscale: task_data.use_upscale = resolve_model_to_use(task_data.use_upscale, 'gfpgan')
 
 def set_vram_optimizations(context: Context):
+    def is_greater(a, b): # is a > b?
+        if a == "low": # b will be "low", "balanced" or "high"
+            return False
+        elif a == "balanced" and b != "low": # b will be "balanced" or "high"
+            return False
+        return True
+
     config = app.getConfig()
-    perf_level = config.get('performance_level', device_manager.get_max_perf_level(context.device))
+
-    vram_optimizations = PERF_LEVEL_TO_VRAM_OPTIMIZATIONS[perf_level]
+    max_usage_level = device_manager.get_max_vram_usage_level(context.device)
+    vram_usage_level = config.get('vram_usage_level', 'balanced')
+
+    if is_greater(vram_usage_level, max_usage_level):
+        log.error(f'Requested GPU Memory Usage level ({vram_usage_level}) is higher than what is ' + \
+                  f'possible ({max_usage_level}) on this device ({context.device}). Using "{max_usage_level}" instead')
+        vram_usage_level = max_usage_level
+
+    vram_optimizations = VRAM_USAGE_LEVEL_TO_OPTIMIZATIONS[vram_usage_level]
 
     if vram_optimizations != context.vram_optimizations:
         context.vram_optimizations = vram_optimizations

ui/sd_internal/task_manager.py

@@ -341,7 +341,7 @@ def get_devices():
             'name': torch.cuda.get_device_name(device),
             'mem_free': mem_free,
             'mem_total': mem_total,
-            'max_perf_level': device_manager.get_max_perf_level(device),
+            'max_vram_usage_level': device_manager.get_max_vram_usage_level(device),
         }
 
     # list the compatible devices

ui/server.py

@@ -134,7 +134,7 @@ def render(req: dict):
 
         render_req.init_image_mask = req.get('mask') # hack: will rename this in the HTTP API in a future revision
 
-        app.save_to_config(task_data.use_stable_diffusion_model, task_data.use_vae_model, task_data.use_hypernetwork_model, task_data.performance_level)
+        app.save_to_config(task_data.use_stable_diffusion_model, task_data.use_vae_model, task_data.use_hypernetwork_model, task_data.vram_usage_level)
 
         # enqueue the task
         new_task = task_manager.render(render_req, task_data)