diff --git a/modules/ui_definitions.py b/modules/ui_definitions.py index 1808f62a7..d52a142b6 100644 --- a/modules/ui_definitions.py +++ b/modules/ui_definitions.py @@ -711,7 +711,7 @@ def create_settings(cmd_opts): "lora_stack_sep": OptionInfo("

Stacking options

", "", gr.HTML), "lora_stack_mode": OptionInfo("sum", "LoRA stack mode", gr.Dropdown, {"choices": ["sum", "ties", "dare_ties", "dare_linear", "magnitude_prune", "klora", "estlora"]}), "lora_stack_density": OptionInfo(0.5, "LoRA stack density", gr.Slider, {"minimum": 0.05, "maximum": 1.0, "step": 0.05}), - "lora_stack_alpha": OptionInfo(1.5, "LoRA stack ramp", gr.Slider, {"minimum": 0.0, "maximum": 3.0, "step": 0.1}), + "lora_stack_alpha": OptionInfo(0.0, "LoRA stack ramp", gr.Slider, {"minimum": 0.0, "maximum": 3.0, "step": 0.1}), "lora_stack_discrepancy": OptionInfo(0.5, "LoRA stack discrepancy", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.05}), "lora_meta_sep": OptionInfo("

Metadata

", "", gr.HTML), diff --git a/ui/locale/locale_en.json b/ui/locale/locale_en.json index f52b58a91..3fb030aa1 100644 --- a/ui/locale/locale_en.json +++ b/ui/locale/locale_en.json @@ -867,9 +867,9 @@ {"id":"","label":"LoRA quantized host rank","localized":"","hint":"Maximum rank used to carry adapter types that are not natively low-rank (LoKR, LoHA, OFT, DoRA) alongside the quantized weights instead of merging them in.
Higher values retain more of the adapter at proportionally more memory. Plain LoRA files are carried exactly at their own rank.

Applies only to SDNQ models quantized below 8 bits, where merging erases most of the adapter; at 8 bits and above merging retains it and hosting is skipped.

0 disables hosting and merges every adapter into the quantized weights.

Default is 256.","ui":"settings_lora"}, {"id":"","label":"LoRA quantized host calibration","localized":"","hint":"Collects per-channel activation statistics from the model's own generations and uses them to focus hosted-adapter truncation on the channels with the strongest activations.
Statistics accumulate in the background on models quantized below 8 bits, persist per checkpoint, and raise delivered adapter fidelity at the same LoRA quantized host rank, most at low ranks.

Capture is skipped while the model is compiled; previously cached statistics still apply.

Enabled by default.","ui":"settings_lora"}, {"id":"","label":"LoRA quantized host cache","localized":"","hint":"Disk space in GB for caching computed hosting factors.
A cached set skips the truncation math on the next load; least recently used entries are evicted once the budget is exceeded.

0 disables the cache.

Default is 10.","ui":"settings_lora"}, - {"id":"","label":"LoRA stack mode","localized":"","hint":"How multiple networks targeting the same layer are combined:
- sum: adds all contributions
- ties: keeps each network's strongest elements and merges only where signs agree
- dare_ties: randomly drops elements, rescales the survivors, then merges where signs agree
- dare_linear: randomly drops elements, rescales the survivors and sums
- magnitude_prune: keeps each network's strongest elements and sums
- klora / estlora: assign each layer to one of exactly two networks, the first in the prompt as subject and the second as style, shifting from subject toward style over the sampling steps

Each layer is given to a single network at a time, so a subject and a style that both need sustained strength can end up under-applied. For reliable blending of two strong networks, sum, ties and dare_ties apply every network throughout and combine more fully.

Kept fractions are set by LoRA stack density; the subject-to-style shift by LoRA stack ramp and LoRA stack discrepancy.

Applies to the native load path; other load methods and text encoder networks always combine as sum. Selection modes fall back to sum unless exactly two networks are loaded, or when model compile is active.

Default is sum.","ui":"settings_lora"}, + {"id":"","label":"LoRA stack mode","localized":"","hint":"How multiple networks targeting the same layer are combined:
- sum: adds all contributions
- ties: keeps each network's strongest elements and merges only where signs agree
- dare_ties: randomly drops elements, rescales the survivors, then merges where signs agree
- dare_linear: randomly drops elements, rescales the survivors and sums
- magnitude_prune: keeps each network's strongest elements and sums
- klora / estlora: assign each layer to one of exactly two networks, the first in the prompt as subject and the second as style; LoRA stack ramp optionally shifts layers toward style over the sampling steps

Each layer is given to a single network at a time, so a subject and a style that both need sustained strength can end up under-applied. For reliable blending of two strong networks, sum, ties and dare_ties apply every network throughout and combine more fully.

Kept fractions are set by LoRA stack density; the subject-to-style shift by LoRA stack ramp and LoRA stack discrepancy.

Applies to the native load path; other load methods and text encoder networks always combine as sum. Selection modes fall back to sum unless exactly two networks are loaded, or when model compile is active.

Default is sum.","ui":"settings_lora"}, {"id":"","label":"LoRA stack density","localized":"","hint":"Fraction of elements each network keeps under the ties, dare_ties, dare_linear and magnitude_prune stack modes.
Lower values keep only the strongest contributions and reduce interference between networks at the cost of per-network detail. The dare variants drop at random and rescale the survivors to preserve expected strength.

Default is 0.5.","ui":"settings_lora"}, - {"id":"","label":"LoRA stack ramp","localized":"","hint":"Slope of the subject-to-style shift across the sampling steps in the klora and estlora stack modes.
Higher values shift layers to the style network earlier and more broadly; lower values keep the subject network dominant for longer.

0 keeps the balance fixed for the whole generation.

Default is 1.5.","ui":"settings_lora"}, + {"id":"","label":"LoRA stack ramp","localized":"","hint":"Slope of the subject-to-style shift across the sampling steps in the klora and estlora stack modes.

0 keeps the layer assignment fixed for the whole generation: each layer stays with the network that is more salient there, which preserves the subject while the style keeps its own layers. Higher values hand layers to the style network progressively, ending in a style takeover; on few-step models the handover happens early enough to override the subject.

Default is 0.","ui":"settings_lora"}, {"id":"","label":"LoRA stack discrepancy","localized":"","hint":"Stand-in for the measured style separation the estlora stack mode would otherwise derive from data.
Higher values keep layers with the subject network longer; lower values let the style network take layers earlier.

Layer scores are balanced by each network's overall strength, so a louder network does not take layers on magnitude alone.

Applies only when LoRA stack mode is estlora.

Default is 0.5.","ui":"settings_lora"}, {"id":"","label":"LoRA auto-apply tags","localized":"","hint":"Automatically add trigger words/tags from LoRA metadata to your prompt.
Set to the number of tags to auto-apply, e.g., 3 = add top 3 trigger tags.
Set to 0 to disable, -1 to add all available tags.","ui":"settings_lora"}, {"id":"","label":"LoRA memory cache","localized":"","hint":"How many LoRAs to keep in network for future use before requiring reloading from storage","ui":"settings_lora"},