feat: add top-n-sigma sampler

src-tauri/src/chat_manager/execution/mod.rs

@@ -1197,6 +1197,24 @@ pub(super) fn resolve_llama_xtc_threshold(
         .or(settings.advanced_model_settings.llama_xtc_threshold)
 }
 
+pub(super) fn resolve_llama_top_n_sigma(
+    session: &Session,
+    model: &Model,
+    settings: &Settings,
+) -> Option<f64> {
+    session
+        .advanced_model_settings
+        .as_ref()
+        .and_then(|cfg| cfg.llama_top_n_sigma)
+        .or_else(|| {
+            model
+                .advanced_model_settings
+                .as_ref()
+                .and_then(|cfg| cfg.llama_top_n_sigma)
+        })
+        .or(settings.advanced_model_settings.llama_top_n_sigma)
+}
+
 pub(super) fn resolve_llama_adaptive_target(
     session: &Session,
     model: &Model,

src-tauri/src/chat_manager/execution/provider_fields.rs

@@ -24,7 +24,8 @@ use super::{
     resolve_llama_sampler_order, resolve_llama_sampler_profile, resolve_llama_seed,
     resolve_llama_single_gpu_device_id_leveled, resolve_llama_streaming_enabled,
     resolve_llama_strict_mode, resolve_llama_swa_full, resolve_llama_threads,
-    resolve_llama_threads_batch, resolve_llama_ubatch_size, resolve_llama_xtc_probability,
+    resolve_llama_threads_batch, resolve_llama_top_n_sigma, resolve_llama_ubatch_size,
+    resolve_llama_xtc_probability,
     resolve_llama_xtc_threshold, resolve_max_tokens, resolve_presence_penalty, resolve_temperature,
     resolve_top_k, resolve_top_p,
 };
@@ -191,6 +192,9 @@ fn build_llama_extra_fields(
     if let Some(v) = resolve_llama_xtc_threshold(session, model, settings) {
         extra.insert("llamaXtcThreshold".to_string(), json!(v));
     }
+    if let Some(v) = resolve_llama_top_n_sigma(session, model, settings) {
+        extra.insert("llamaTopNSigma".to_string(), json!(v));
+    }
     if let Some(v) = resolve_llama_adaptive_target(session, model, settings) {
         extra.insert("llamaAdaptiveTarget".to_string(), json!(v));
     }
@@ -261,6 +265,7 @@ mod tests {
             llama_dry_sequence_breakers: Some(vec![":".to_string()]),
             llama_xtc_probability: Some(0.5),
             llama_xtc_threshold: Some(0.1),
+            llama_top_n_sigma: Some(1.5),
             llama_adaptive_target: Some(0.3),
             llama_adaptive_decay: Some(0.95),
             ..Default::default()
@@ -307,7 +312,7 @@ mod tests {
         assert_eq!(extra.get("llamaNPenRange"), Some(&json!(64)));
         assert_eq!(
             extra.len(),
-            44,
+            45,
             "unexpected llama extra-body key count; a resolver stopped emitting or a new field was added without updating this fixture: {:?}",
             extra.keys().collect::<Vec<_>>()
         );

src-tauri/src/chat_manager/feature_generation.rs

@@ -96,6 +96,7 @@ impl FeatureGenerationSettings {
             llama_n_pen_range: self.llama_n_pen_range,
             llama_xtc_probability: self.llama_xtc_probability,
             llama_xtc_threshold: self.llama_xtc_threshold,
+            llama_top_n_sigma: self.llama_top_n_sigma,
             llama_dry_multiplier: self.llama_dry_multiplier,
             llama_dry_base: self.llama_dry_base,
             llama_dry_allowed_length: self.llama_dry_allowed_length,
@@ -124,6 +125,7 @@ impl FeatureGenerationSettings {
             || self.llama_n_pen_range.is_some()
             || self.llama_xtc_probability.is_some()
             || self.llama_xtc_threshold.is_some()
+            || self.llama_top_n_sigma.is_some()
             || self.llama_dry_multiplier.is_some()
             || self.llama_dry_base.is_some()
             || self.llama_dry_allowed_length.is_some()

src-tauri/src/chat_manager/types.rs

@@ -519,6 +519,8 @@ pub struct FeatureGenerationSettings {
     #[serde(default)]
     pub llama_xtc_threshold: Option<f64>,
     #[serde(default)]
+    pub llama_top_n_sigma: Option<f64>,
+    #[serde(default)]
     pub llama_adaptive_target: Option<f64>,
     #[serde(default)]
     pub llama_adaptive_decay: Option<f64>,
@@ -878,6 +880,8 @@ pub struct AdvancedModelSettings {
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub llama_xtc_threshold: Option<f64>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub llama_top_n_sigma: Option<f64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub llama_adaptive_target: Option<f64>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub llama_adaptive_decay: Option<f64>,
@@ -1032,6 +1036,7 @@ impl Default for AdvancedModelSettings {
             llama_dry_sequence_breakers: None,
             llama_xtc_probability: None,
             llama_xtc_threshold: None,
+            llama_top_n_sigma: None,
             llama_adaptive_target: None,
             llama_adaptive_decay: None,
             llama_last_runtime_report: None,

src-tauri/src/hf_browser/image_bundle.rs

@@ -294,7 +294,7 @@ fn to_search_result(detail: ApiModel) -> super::HfSearchResult {
 fn vae_family_term(profile: &sdcpp::HfBundleProfile) -> &'static str {
     if profile.id.starts_with("krea") {
         "wan"
-    } else if profile.id == "qwen-image-edit-2511" {
+    } else if profile.id.starts_with("qwen-image") {
         "qwen"
     } else {
         "flux"
@@ -453,10 +453,10 @@ fn role_compatible(profile: &sdcpp::HfBundleProfile, role: &str, model_id: &str,
     match role {
         "diffusion_model" => profile.diffusion_markers.iter().any(|marker| evidence.contains(marker)),
         "text_encoder" => profile.encoder_markers.iter().any(|marker| evidence.contains(marker)),
-        "vision_encoder" => profile.id == "qwen-image-edit-2511"
+        "vision_encoder" => profile.id.starts_with("qwen-image")
             && (evidence.contains("mmproj") || evidence.contains("vision")),
         "vae" if profile.id.starts_with("krea") => evidence.contains("wan") && evidence.contains("vae"),
-        "vae" if profile.id == "qwen-image-edit-2511" => evidence.contains("qwen") && evidence.contains("vae"),
+        "vae" if profile.id.starts_with("qwen-image") => evidence.contains("qwen") && evidence.contains("vae"),
         "vae" => (evidence.contains("flux") || evidence.contains("ae.")) && (evidence.contains("vae") || evidence.contains("ae.")),
         _ => false,
     }
@@ -482,6 +482,8 @@ async fn gguf_architecture_hint(
     [
         "qwen2.5-vl-7b",
         "qwen2.5vl-7b",
+        "qwen3-vl-8b",
+        "qwen3vl-8b",
         "qwen3-vl-4b",
         "qwen3vl-4b",
         "qwen3-8b",

src-tauri/src/image_generator/sdcpp.rs

@@ -542,6 +542,80 @@ const QWEN_VARIANTS: &[VariantSpec] = &[
     },
 ];
 
+const QWEN21_SHARED: &[ComponentSpec] = &[
+    ComponentSpec {
+        role: "text_encoder",
+        repo: "unsloth/Qwen3-VL-8B-Instruct-GGUF",
+        revision: "b93a7ee713758252c555be4210c00540df954dc2",
+        filename: "Qwen3-VL-8B-Instruct-Q4_K_M.gguf",
+        bytes: 5_027_785_568,
+        sha256: "108e7ff92b78eefd3db4741885104acba514255c11b617d3c7b197a5f46efe89",
+    },
+    ComponentSpec {
+        role: "vision_encoder",
+        repo: "unsloth/Qwen3-VL-8B-Instruct-GGUF",
+        revision: "b93a7ee713758252c555be4210c00540df954dc2",
+        filename: "mmproj-BF16.gguf",
+        bytes: 1_162_569_280,
+        sha256: "6516bb64bae1503a0fcd7ec9fa39655f8c481580be0a0a066397941d9761c9f4",
+    },
+    ComponentSpec {
+        role: "vae",
+        repo: "Comfy-Org/Qwen-Image-2.1",
+        revision: "0217efd1d9c9452e84f6448903d6e21bfa171ffd",
+        filename: "vae/qwen_image_2.1_vae_bf16.safetensors",
+        bytes: 675_509_688,
+        sha256: "bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9",
+    },
+];
+const QWEN21_VARIANTS: &[VariantSpec] = &[
+    VariantSpec {
+        id: "q3-k",
+        label: "Q3 K (smaller)",
+        description: "Lower memory use with a modest quality tradeoff.",
+        recommended: false,
+        smaller: true,
+        diffusion: ComponentSpec {
+            role: "diffusion_model",
+            repo: "leejet/Qwen-Image-2.1-GGUF",
+            revision: "cc11433936a06e9765f7c0c0b1f0436cfd2b9856",
+            filename: "qwen_image_2.1-Q3_K.gguf",
+            bytes: 3_270_553_632,
+            sha256: "5a6352cb450d334ba17a9d18c7fd8c3e774799245b9da0f963ce0c6624deb758",
+        },
+    },
+    VariantSpec {
+        id: "q4-k",
+        label: "Q4 K (recommended)",
+        description: "Best default balance of image quality, speed, and memory.",
+        recommended: true,
+        smaller: false,
+        diffusion: ComponentSpec {
+            role: "diffusion_model",
+            repo: "leejet/Qwen-Image-2.1-GGUF",
+            revision: "cc11433936a06e9765f7c0c0b1f0436cfd2b9856",
+            filename: "qwen_image_2.1-Q4_K.gguf",
+            bytes: 4_197_494_816,
+            sha256: "29f9c83c249ff0292fb2943fceddfa2319b446601866c82a4f8be062abea72c2",
+        },
+    },
+    VariantSpec {
+        id: "q6-k",
+        label: "Q6 K (higher quality)",
+        description: "Higher precision with greater memory and storage use.",
+        recommended: false,
+        smaller: false,
+        diffusion: ComponentSpec {
+            role: "diffusion_model",
+            repo: "leejet/Qwen-Image-2.1-GGUF",
+            revision: "cc11433936a06e9765f7c0c0b1f0436cfd2b9856",
+            filename: "qwen_image_2.1-Q6_K.gguf",
+            bytes: 5_996_851_232,
+            sha256: "1c51d1a8e6cf2b215719dea0c4f0f4cbfff9214a1c7571e57913f57d6a9c3c12",
+        },
+    },
+];
+
 const PROFILES: &[ProfileSpec] = &[
     ProfileSpec {
         id: "z-image-turbo",
@@ -703,6 +777,26 @@ const PROFILES: &[ProfileSpec] = &[
         variants: QWEN_VARIANTS,
         shared_components: QWEN_SHARED,
     },
+    ProfileSpec {
+        id: "qwen-image-2.1",
+        display_name: "Qwen Image 2.1",
+        family: "Qwen",
+        description: "Unified text-to-image and multi-reference editing with native transparency and strong in-image text rendering.",
+        license: "Qwen-Image Research License (non-commercial)",
+        source_url: "https://huggingface.co/Qwen/Qwen-Image-2.1",
+        supports_text_to_image: true,
+        supports_image_edit: true,
+        max_reference_images: Some(10),
+        requires_reference_image: false,
+        recommended_for_scenes: true,
+        default_width: 1024,
+        default_height: 1024,
+        default_steps: 35,
+        default_cfg: 6.0,
+        minimum_runtime_build: Some(883),
+        variants: QWEN21_VARIANTS,
+        shared_components: QWEN21_SHARED,
+    },
 ];
 
 #[derive(Debug, Clone, Serialize)]
@@ -763,6 +857,11 @@ fn hf_profile_rules(profile: &ProfileSpec) -> (Vec<&'static str>, Vec<&'static s
             vec!["qwen2.5-vl-7b", "qwen2.5vl-7b"],
             7.0,
         ),
+        "qwen-image-2.1" => (
+            vec!["qwen-image-2.1", "qwen_image_2.1", "qwen-image-2-1", "qwen_image_2_1"],
+            vec!["qwen3-vl-8b", "qwen3vl-8b"],
+            8.0,
+        ),
         _ => (Vec::new(), Vec::new(), 0.0),
     }
 }
@@ -3852,6 +3951,7 @@ fn profile_lora_architecture(profile_id: Option<&str>) -> Option<&'static str> {
         "flux-2-klein-9b" | "flux-2-klein-base-9b" => Some("flux2-klein-9b"),
         "krea-2-turbo" | "krea-2-raw" => Some("krea-2"),
         "qwen-image-edit-2511" => Some("qwen-image-edit-2511"),
+        "qwen-image-2.1" => Some("qwen-image-2.1"),
         _ => None,
     }
 }
@@ -4882,6 +4982,13 @@ async fn ensure_server(
         if let Some(params_backend_spec) = params_backend_spec {
             command.arg("--params-backend").arg(params_backend_spec);
         }
+    } else if runtime_build_number(&config.sdcpp_runtime_release)
+        .map_or(true, |build| build >= AUTO_FIT_VALUE_MIN_BUILD)
+    {
+        // stable-diffusion.cpp master-962 (PR #2132) turned --auto-fit into a
+        // value-taking option; older builds still expect the bare flag, and an
+        // unknown/non-numbered tag is treated as current.
+        command.arg("--auto-fit").arg("on");
     } else {
         command.arg("--auto-fit");
     }
@@ -5832,6 +5939,10 @@ fn find_profile_variant(
     Ok((profile, variant))
 }
 
+/// stable-diffusion.cpp build at which `--auto-fit` became a value-taking
+/// option (`on`/`off`) instead of a bare flag (PR #2132, commit d763dae).
+const AUTO_FIT_VALUE_MIN_BUILD: u32 = 962;
+
 fn runtime_build_number(release: &str) -> Option<u32> {
     release
         .strip_prefix("master-")?

src-tauri/src/llama_cpp/desktop/sampler.rs

@@ -1,7 +1,7 @@
 use super::*;
 
 pub(super) const DEFAULT_LLAMA_SAMPLER_PROFILE: &str = "balanced";
-pub(super) const DEFAULT_LLAMA_SAMPLER_ORDER: [&str; 10] = [
+pub(super) const DEFAULT_LLAMA_SAMPLER_ORDER: [&str; 11] = [
     "penalties",
     "grammar",
     "top_k",
@@ -10,6 +10,7 @@ pub(super) const DEFAULT_LLAMA_SAMPLER_ORDER: [&str; 10] = [
     "dry",
     "typical",
     "xtc",
+    "top_n_sigma",
     "temp",
     "adaptive_p",
 ];
@@ -44,6 +45,7 @@ pub(super) struct ResolvedSamplerConfig {
     pub(super) dry_sequence_breakers: Option<Vec<String>>,
     pub(super) xtc_probability: Option<f64>,
     pub(super) xtc_threshold: Option<f64>,
+    pub(super) top_n_sigma: Option<f64>,
     pub(super) frequency_penalty: Option<f64>,
     pub(super) presence_penalty: Option<f64>,
     pub(super) adaptive_target: Option<f64>,
@@ -176,6 +178,7 @@ fn normalize_sampler_stage(value: &str) -> Option<&'static str> {
         "dry" => Some("dry"),
         "typical" | "typ_p" | "typical_p" => Some("typical"),
         "xtc" => Some("xtc"),
+        "top_n_sigma" | "top_nsigma" | "topnsigma" | "nsigma" => Some("top_n_sigma"),
         "temp" | "temperature" => Some("temp"),
         "adaptive_p" | "adaptivep" | "adaptive" => Some("adaptive_p"),
         _ => None,
@@ -416,6 +419,10 @@ pub(super) fn build_sampler(
                 config.seed.unwrap_or_else(rand::random::<u32>),
             )
         });
+    let mut top_n_sigma_sampler = config.top_n_sigma.filter(|n| *n > 0.0).map(|n| {
+        active_params.insert("top_n_sigma".to_string(), json!(n));
+        LlamaSampler::top_n_sigma(n as f32)
+    });
 
     let mut adaptive_requested = false;
     for stage in requested_order {
@@ -468,6 +475,12 @@ pub(super) fn build_sampler(
                     samplers.push(sampler);
                 }
             }
+            "top_n_sigma" => {
+                if let Some(sampler) = top_n_sigma_sampler.take() {
+                    order.push("top_n_sigma");
+                    samplers.push(sampler);
+                }
+            }
             "temp" => {
                 if config.temperature > 0.0 {
                     order.push("temp");

src-tauri/src/llama_cpp/mod.rs

@@ -1298,6 +1298,13 @@ mod desktop {
             .or_else(|| body.get("llamaXtcThreshold"))
             .or_else(|| body.get("llama_xtc_threshold"))
             .and_then(|v| v.as_f64());
+        let top_n_sigma = body
+            .get("top_n_sigma")
+            .or_else(|| body.get("topNSigma"))
+            .or_else(|| body.get("llamaTopNSigma"))
+            .or_else(|| body.get("llama_top_n_sigma"))
+            .and_then(|v| v.as_f64())
+            .filter(|v| *v > 0.0);
         let adaptive_target = body
             .get("adaptive_target")
             .or_else(|| body.get("adaptiveTarget"))
@@ -3984,6 +3991,7 @@ mod desktop {
                 dry_sequence_breakers: dry_sequence_breakers.clone(),
                 xtc_probability,
                 xtc_threshold,
+                top_n_sigma,
                 frequency_penalty,
                 presence_penalty,
                 adaptive_target,

src-tauri/src/providers/config.rs

@@ -72,6 +72,7 @@ fn supported_extra_body_keys(provider_id: &str) -> &'static [&'static str] {
             "llamaDrySequenceBreakers",
             "llamaXtcProbability",
             "llamaXtcThreshold",
+            "llamaTopNSigma",
             "llamaAdaptiveTarget",
             "llamaAdaptiveDecay",
             "llamaDisableSamplerProfileDefaults",