Add audio reference + per-lora audio strength via MultiLoRALoader

Rewrite rgthree Power Lora Loader to phazei MultiLoRALoader in LTX mode so each lora has separate video / audio strength multipliers. Add LTXVReferenceAudio node injection with optional MelBandRoFormer stem separation for voice ID transfer. Exclude KV node from POS/NEG consumer redirect to avoid cycle when KV + audio_ref both enabled. Add rotary_embedding_torch dependency.

app.py

@@ -91,6 +91,8 @@ CUSTOM_NODES = [
     ("ComfyUI-RMBG", "https://github.com/1038lab/ComfyUI-RMBG.git"),
     ("ComfyUI-PromptRelay", "https://github.com/kijai/ComfyUI-PromptRelay.git"),
     ("ComfyUI-FunPack", "https://github.com/digital-garbage/ComfyUI-FunPack.git"),
+    ("ComfyUI-MelBandRoFormer", "https://github.com/kijai/ComfyUI-MelBandRoFormer.git"),
+    ("ComfyUI-MultiLoRALoader", "https://github.com/phazei/ComfyUI-MultiLoRALoader.git"),
 ]
 
 # FunPackKVApply wrapper, written into comfy's custom_nodes at startup.
@@ -456,6 +458,12 @@ DOWNLOADS = [
         "dest": MODELS / "loras" / "ltx23" / "Singularity-LTX-2.3_OmniCine_V1nsf.safetensors",
         "label": "singularity lora",
     },
+    {
+        "repo": "Kijai/MelBandRoFormer_comfy",
+        "file": "MelBandRoformer_fp16.safetensors",
+        "dest": MODELS / "diffusion_models" / "MelBandRoformer_fp16.safetensors",
+        "label": "mel band roformer (stem separation)",
+    },
 ]
 
 SULPHUR_LORA_FILENAME = "ltx23/LTX_SulphurEXP_LoRA_fro99-avgrank105.safetensors"
@@ -524,6 +532,11 @@ NODE_LTXV_CONDITIONING = "523"  # consumes positive from CLIPTextEncode 536
 # LTXVImgToVideoInplaceKJ pass 1 (node 772) slot 0. Disabled when MSR
 # mode is on (model chain already rewired).
 KV_NEW_NODE = "kv_apply"
+NODE_AUDIO_VAE_LOADER = "617"
+AUDIO_REF_NEW_LOAD = "audio_ref_load"
+AUDIO_REF_NEW_MEL_LOADER = "audio_ref_mel_loader"
+AUDIO_REF_NEW_MEL_SAMPLER = "audio_ref_mel_sampler"
+AUDIO_REF_NEW_NODE = "audio_ref"
 NODE_I2V_REF_LATENT = "772"  # LTXVImgToVideoInplaceKJ pass 1, slot 0
 
 NODE_OUTPUT = "597"
@@ -1020,7 +1033,11 @@ def _ensure_comfy() -> None:
     if _comfy_ready:
         return
 
-    _ensure_repo(COMFY, "https://github.com/comfyanonymous/ComfyUI.git")
+    _ensure_repo(
+        COMFY,
+        "https://github.com/comfyanonymous/ComfyUI.git",
+        commit="4e1f7cb1db1c26bb9ee61cf1875776517e2abae8",
+    )
     _install_filtered_requirements(COMFY / "requirements.txt")
 
     custom_root = COMFY / "custom_nodes"
@@ -1040,6 +1057,7 @@ def _ensure_comfy() -> None:
         "upscale_models",
         "latent_upscale_models",
         "vae",
+        "diffusion_models",
     ):
         (MODELS / folder).mkdir(parents=True, exist_ok=True)
     INPUT.mkdir(parents=True, exist_ok=True)
@@ -1267,30 +1285,17 @@ def _convert_workflow(visual_path: str) -> dict[str, Any]:
 
         widgets = node.get("widgets_values") or []
         if class_type == "Power Lora Loader (rgthree)":
-            # rgthree stores loras as dict entries in a list; convert each into
-            # a lora_N keyed input which is what the node reads in API form.
-            # Normalize Windows backslashes in lora paths so workflows authored
-            # on Windows resolve to the right file on Linux (otherwise rgthree
-            # treats the backslash as a literal character in the filename and
-            # silently skips the lora).
-            # Skip OmniNFT entries from the template - we expose them via
-            # OPTIONAL_LORAS so users can tune strength / pick between the
-            # converted and the bf16 RL variant from the UI.
-            lora_idx = 0
-            for item in widgets if isinstance(widgets, list) else []:
-                if isinstance(item, dict) and "lora" in item and "strength" in item:
-                    lora_path = item["lora"]
-                    if isinstance(lora_path, str):
-                        lora_path = lora_path.replace("\\", "/")
-                    if isinstance(lora_path, str) and "omninft" in lora_path.lower():
-                        continue
-                    lora_idx += 1
-                    inputs[f"lora_{lora_idx}"] = {
-                        "on": bool(item.get("on", True)),
-                        "lora": lora_path,
-                        "strength": item.get("strength", 1.0),
-                        "strengthTwo": item.get("strengthTwo"),
-                    }
+            # We rewrite rgthree's Power Lora Loader to phazei's MultiLoRALoader
+            # in LTX mode so each lora has separate video/audio strength control
+            # (Vid, V2A, Aud, A2V, Other per-tensor-pattern multipliers on top of
+            # the global STR). Same output signature (model, clip), so downstream
+            # connections work unchanged. Lora list lives in the lora_data JSON
+            # string; _inject_optional_loras populates it later. OmniNFT entries
+            # from the template are dropped here (exposed separately via the
+            # OPTIONAL_LORAS sliders).
+            class_type = "MultiLoRALoader"
+            inputs["lora_data"] = "[]"
+            inputs["ltx_mode"] = True
         elif isinstance(widgets, dict):
             for key, value in widgets.items():
                 if key != "videopreview":
@@ -1515,23 +1520,11 @@ def _convert_runexx_workflow(visual_path: str) -> dict[str, Any]:
 
         widgets = n.get("widgets_values") or []
         if class_type == "Power Lora Loader (rgthree)":
-            # Same OmniNFT-skip rule as the primary converter so user-side
-            # OPTIONAL_LORAS sliders are the only authority on those entries.
-            lora_idx = 0
-            for item in widgets if isinstance(widgets, list) else []:
-                if isinstance(item, dict) and "lora" in item and "strength" in item:
-                    lora_path = item["lora"]
-                    if isinstance(lora_path, str):
-                        lora_path = lora_path.replace("\\", "/")
-                    if isinstance(lora_path, str) and "omninft" in lora_path.lower():
-                        continue
-                    lora_idx += 1
-                    inputs[f"lora_{lora_idx}"] = {
-                        "on": bool(item.get("on", True)),
-                        "lora": lora_path,
-                        "strength": item.get("strength", 1.0),
-                        "strengthTwo": item.get("strengthTwo"),
-                    }
+            # Same rewrite as the primary converter: rgthree -> MultiLoRALoader
+            # in LTX mode for per-modality strength control.
+            class_type = "MultiLoRALoader"
+            inputs["lora_data"] = "[]"
+            inputs["ltx_mode"] = True
         elif isinstance(widgets, dict):
             for key, value in widgets.items():
                 if key != "videopreview":
@@ -1655,6 +1648,17 @@ def _inject_params(
     omninft_bf16_lora_strength: float = 0.0,
     better_motion_lora_strength: float = 0.0,
     physics_v2_lora_strength: float = 0.0,
+    sulphur_audio_strength: float = 0.15,
+    sulphur_v1_audio_strength: float = 0.15,
+    vbvr_audio_strength: float = 0.5,
+    dreamly_audio_strength: float = 0.6,
+    synth_audio_strength: float = 0.0,
+    plora_audio_strength: float = 0.0,
+    singularity_audio_strength: float = 0.3,
+    omninft_audio_strength: float = 0.8,
+    omninft_bf16_audio_strength: float = 0.0,
+    better_motion_audio_strength: float = 0.0,
+    physics_v2_audio_strength: float = 0.0,
     cache_at_step: int = 0,
     cache_warmup: int = 50,
     energy_threshold: float = 0.3,
@@ -1673,6 +1677,10 @@ def _inject_params(
     prompt_segments: str = "",
     kv_enabled: bool = False,
     kv_strength: float = 1.0,
+    audio_ref_enabled: bool = False,
+    audio_ref_filename: str | None = None,
+    audio_ref_guidance_scale: float = 3.0,
+    audio_ref_stem_sep: bool = False,
 ) -> dict[str, Any]:
     # MSR (multi-reference) mode patches the workflow heavily - bypasses the
     # likeness/anchor system, inserts IC-LoRA conditioning, adds crop guides
@@ -1711,6 +1719,16 @@ def _inject_params(
     # which may be the relay node's output.
     if kv_enabled and not msr_enabled:
         _inject_kv_conditioning(workflow, strength=float(kv_strength))
+    # Audio reference: voice ID transfer. Splices LTXVReferenceAudio between
+    # PowerLora and downstream, also patching conditioning. Disabled in MSR
+    # mode (heavily-rewired chain) and skipped if no audio uploaded.
+    if (audio_ref_enabled and audio_ref_filename and not msr_enabled):
+        _inject_audio_reference(
+            workflow,
+            audio_filename=audio_ref_filename,
+            guidance_scale=float(audio_ref_guidance_scale),
+            stem_sep=bool(audio_ref_stem_sep),
+        )
     # Refine-pass sigmas. original=workflow default. tuned=drops the 0.715
     # high-sigma step. custom=validated upstream string.
     _inject_refine_sigmas(workflow, _resolve_sigmas(sigma_preset, sigma_custom))
@@ -1723,19 +1741,35 @@ def _inject_params(
     else:
         resolved_cache_step = int(cache_at_step)
     workflow[NODE_LOAD_IMAGE]["inputs"]["image"] = image_name
-    _inject_optional_loras(workflow, {
-        "lora_sulphur": sulphur_lora_strength,
-        "lora_sulphur_v1": sulphur_v1_lora_strength,
-        "lora_vbvr": vbvr_lora_strength,
-        "lora_dreamly": dreamly_lora_strength,
-        "lora_synth": synth_lora_strength,
-        "lora_plora": plora_lora_strength,
-        "lora_singularity": singularity_lora_strength,
-        "lora_omninft": omninft_lora_strength,
-        "lora_omninft_bf16": omninft_bf16_lora_strength,
-        "lora_better_motion": better_motion_lora_strength,
-        "lora_physics_v2": physics_v2_lora_strength,
-    })
+    _inject_optional_loras(
+        workflow,
+        video_strengths={
+            "lora_sulphur": sulphur_lora_strength,
+            "lora_sulphur_v1": sulphur_v1_lora_strength,
+            "lora_vbvr": vbvr_lora_strength,
+            "lora_dreamly": dreamly_lora_strength,
+            "lora_synth": synth_lora_strength,
+            "lora_plora": plora_lora_strength,
+            "lora_singularity": singularity_lora_strength,
+            "lora_omninft": omninft_lora_strength,
+            "lora_omninft_bf16": omninft_bf16_lora_strength,
+            "lora_better_motion": better_motion_lora_strength,
+            "lora_physics_v2": physics_v2_lora_strength,
+        },
+        audio_strengths={
+            "lora_sulphur": sulphur_audio_strength,
+            "lora_sulphur_v1": sulphur_v1_audio_strength,
+            "lora_vbvr": vbvr_audio_strength,
+            "lora_dreamly": dreamly_audio_strength,
+            "lora_synth": synth_audio_strength,
+            "lora_plora": plora_audio_strength,
+            "lora_singularity": singularity_audio_strength,
+            "lora_omninft": omninft_audio_strength,
+            "lora_omninft_bf16": omninft_bf16_audio_strength,
+            "lora_better_motion": better_motion_audio_strength,
+            "lora_physics_v2": physics_v2_audio_strength,
+        },
+    )
     workflow[NODE_POSITIVE]["inputs"]["text"] = prompt
     workflow[NODE_NEGATIVE]["inputs"]["text"] = negative_prompt
     workflow[NODE_SEED]["inputs"]["seed"] = seed
@@ -1873,26 +1907,43 @@ OPTIONAL_LORAS = {
 }
 
 
-def _inject_optional_loras(workflow: dict[str, Any], strengths: dict[str, float]) -> None:
-    """Add the optional loras to the rgthree power loader.
-
-    Each optional lora is disabled (strength 0) by default and appended as an
-    extra lora_N entry only when a positive strength is requested. Idempotent.
+def _inject_optional_loras(
+    workflow: dict[str, Any],
+    video_strengths: dict[str, float],
+    audio_strengths: dict[str, float] | None = None,
+) -> None:
+    """Populate the MultiLoRALoader's lora_data JSON string.
+
+    LTX-mode entry format (per phazei's dispatch): per-key alpha is multiplied
+    by the modality factor matching the tensor name pattern, then the global
+    `str` applies on top. vid covers main video attn/ff.net tensors, aud covers
+    audio_attn / audio_ff.net, v2a / a2v cover cross-modal attn. Setting aud
+    independent of vid lets a non-audio-trained lora influence video without
+    distorting the audio stream. Disabled (skipped) when video_strength <= 0
+    and audio_strength <= 0. Idempotent.
     """
     node = workflow.get(NODE_POWER_LORA)
     if node is None:
         return
-    inputs = node["inputs"]
+    audio_strengths = audio_strengths or {}
+    entries: list[dict[str, Any]] = []
     for key, filename in OPTIONAL_LORAS.items():
-        inputs.pop(key, None)
-        strength = strengths.get(key, 0.0)
-        if strength and float(strength) > 0:
-            inputs[key] = {
-                "on": True,
-                "lora": filename,
-                "strength": float(strength),
-                "strengthTwo": None,
-            }
+        vid = float(video_strengths.get(key, 0.0) or 0.0)
+        aud = float(audio_strengths.get(key, vid) or 0.0)
+        if vid <= 0 and aud <= 0:
+            continue
+        entries.append({
+            "lora": filename,
+            "on": True,
+            "str": 1.0,
+            "vid": vid,
+            "v2a": vid,
+            "aud": aud,
+            "a2v": vid,
+            "other": vid,
+        })
+    node["inputs"]["lora_data"] = json.dumps(entries)
+    node["inputs"]["ltx_mode"] = True
 
 
 def _validate_sigmas(s: str) -> str:
@@ -2340,6 +2391,92 @@ def _inject_kv_conditioning(workflow: dict[str, Any], strength: float = 1.0) ->
     return True
 
 
+def _inject_audio_reference(
+    workflow: dict[str, Any],
+    audio_filename: str,
+    guidance_scale: float = 3.0,
+    stem_sep: bool = False,
+) -> bool:
+    """Splice an LTXVReferenceAudio node between Power Lora Loader and its
+    downstream model consumers, also patching the positive/negative
+    conditioning chain. The node encodes the ref audio via the existing
+    LTXVAudioVAELoader (617), patches model with identity guidance, and
+    routes through patched conditioning.
+
+    When stem_sep=True we run the ref through MelBandRoFormer first to
+    isolate vocals from background music/noise. Vocals (slot 0) feed
+    LTXVReferenceAudio.reference_audio.
+
+    Returns True on success, False if required upstream nodes are absent.
+    """
+    required = (NODE_POWER_LORA, NODE_POSITIVE, NODE_NEGATIVE, NODE_AUDIO_VAE_LOADER)
+    if not all(nid in workflow for nid in required):
+        return False
+    power_loader = workflow[NODE_POWER_LORA]
+    upstream_model = power_loader["inputs"].get("model")
+    if upstream_model is None:
+        return False
+
+    # LoadAudio reads from comfy's INPUT dir by filename.
+    workflow[AUDIO_REF_NEW_LOAD] = {
+        "class_type": "LoadAudio",
+        "inputs": {"audio": audio_filename},
+    }
+    ref_audio_source: list = [AUDIO_REF_NEW_LOAD, 0]
+
+    if stem_sep:
+        # MelBandRoFormer separates vocals from instruments.
+        # Model loaded from models/diffusion_models/.
+        workflow[AUDIO_REF_NEW_MEL_LOADER] = {
+            "class_type": "MelBandRoFormerModelLoader",
+            "inputs": {"model_name": "MelBandRoformer_fp16.safetensors"},
+        }
+        workflow[AUDIO_REF_NEW_MEL_SAMPLER] = {
+            "class_type": "MelBandRoFormerSampler",
+            "inputs": {
+                "model": [AUDIO_REF_NEW_MEL_LOADER, 0],
+                "audio": [AUDIO_REF_NEW_LOAD, 0],
+            },
+        }
+        ref_audio_source = [AUDIO_REF_NEW_MEL_SAMPLER, 0]  # vocals
+
+    # LTXVReferenceAudio patches model + conditioning.
+    workflow[AUDIO_REF_NEW_NODE] = {
+        "class_type": "LTXVReferenceAudio",
+        "inputs": {
+            "model": list(upstream_model) if isinstance(upstream_model, list) else upstream_model,
+            "positive": [NODE_POSITIVE, 0],
+            "negative": [NODE_NEGATIVE, 0],
+            "reference_audio": ref_audio_source,
+            "audio_vae": [NODE_AUDIO_VAE_LOADER, 0],
+            "identity_guidance_scale": float(guidance_scale),
+            "start_percent": 0.0,
+            "end_percent": 1.0,
+        },
+    }
+
+    # Route Power Lora's model through the audio-ref-patched model.
+    power_loader["inputs"]["model"] = [AUDIO_REF_NEW_NODE, 0]
+
+    # Reroute downstream conditioning consumers through patched outputs.
+    # Slot 1 = patched positive, slot 2 = patched negative.
+    # Exclude AUDIO_REF_NEW_NODE itself (self-reference) and KV_NEW_NODE
+    # (KV reads raw POSITIVE as context-only signal; redirecting would
+    # create a cycle since AUDIO_REF.model = [KV_NEW, 0]).
+    exclude = {AUDIO_REF_NEW_NODE}
+    if KV_NEW_NODE in workflow:
+        exclude.add(KV_NEW_NODE)
+    _redirect_consumers(
+        workflow, [NODE_POSITIVE, 0], [AUDIO_REF_NEW_NODE, 1],
+        exclude_node_ids=exclude,
+    )
+    _redirect_consumers(
+        workflow, [NODE_NEGATIVE, 0], [AUDIO_REF_NEW_NODE, 2],
+        exclude_node_ids=exclude,
+    )
+    return True
+
+
 def _safe_frames(seconds: float, fps: int = 24) -> int:
     frames = max(9, int(seconds * fps) + 1)
     return ((frames - 1 + 7) // 8) * 8 + 1
@@ -2426,6 +2563,17 @@ def get_gpu_duration(
     omninft_bf16_lora_strength: float = 0.0,
     better_motion_lora_strength: float = 0.0,
     physics_v2_lora_strength: float = 0.0,
+    sulphur_audio_strength: float = 0.15,
+    sulphur_v1_audio_strength: float = 0.15,
+    vbvr_audio_strength: float = 0.5,
+    dreamly_audio_strength: float = 0.6,
+    synth_audio_strength: float = 0.0,
+    plora_audio_strength: float = 0.0,
+    singularity_audio_strength: float = 0.3,
+    omninft_audio_strength: float = 0.8,
+    omninft_bf16_audio_strength: float = 0.0,
+    better_motion_audio_strength: float = 0.0,
+    physics_v2_audio_strength: float = 0.0,
     cache_at_step: int = 0,
     cache_warmup: int = 50,
     energy_threshold: float = 0.3,
@@ -2444,6 +2592,10 @@ def get_gpu_duration(
     prompt_segments: str = "",
     kv_enabled: bool = False,
     kv_strength: float = 1.0,
+    audio_ref_enabled: bool = False,
+    audio_ref_file: str | None = None,
+    audio_ref_guidance_scale: float = 3.0,
+    audio_ref_stem_sep: bool = False,
     progress: gr.Progress | None = None,
 ) -> int:
     # Manual override: gen_budget > 0 forces an exact GPU budget.
@@ -2456,6 +2608,13 @@ def get_gpu_duration(
     mode_cost = 1.10 if mode != "anchor only" else 1.0
     if input_mode == "multi-reference (MSR)":
         mode_cost *= 1.10
+    if audio_ref_enabled and audio_ref_file and audio_ref_guidance_scale > 0:
+        # Identity guidance does an extra forward pass per step during the
+        # active sigma range - roughly 60% slowdown at default scale=3.0.
+        mode_cost *= 1.60
+    elif audio_ref_enabled and audio_ref_file:
+        # ref tokens still encoded + attended in conditioning, ~10% bump.
+        mode_cost *= 1.10
     # Two regimes: tight (30+5*work) lets default 4s fit the 120s/day free
     # ZeroGPU allowance; anything longer falls back to the older wider
     # formula (45+8*work) that's proven to complete on long gens.
@@ -2496,6 +2655,17 @@ def generate(
     omninft_bf16_lora_strength: float = 0.0,
     better_motion_lora_strength: float = 0.0,
     physics_v2_lora_strength: float = 0.0,
+    sulphur_audio_strength: float = 0.15,
+    sulphur_v1_audio_strength: float = 0.15,
+    vbvr_audio_strength: float = 0.5,
+    dreamly_audio_strength: float = 0.6,
+    synth_audio_strength: float = 0.0,
+    plora_audio_strength: float = 0.0,
+    singularity_audio_strength: float = 0.3,
+    omninft_audio_strength: float = 0.8,
+    omninft_bf16_audio_strength: float = 0.0,
+    better_motion_audio_strength: float = 0.0,
+    physics_v2_audio_strength: float = 0.0,
     cache_at_step: int = 0,
     cache_warmup: int = 50,
     energy_threshold: float = 0.3,
@@ -2514,6 +2684,10 @@ def generate(
     prompt_segments: str = "",
     kv_enabled: bool = False,
     kv_strength: float = 1.0,
+    audio_ref_enabled: bool = False,
+    audio_ref_file: str | None = None,
+    audio_ref_guidance_scale: float = 3.0,
+    audio_ref_stem_sep: bool = False,
     progress: gr.Progress = gr.Progress(track_tqdm=True),
 ) -> tuple[str, str, int]:
     seed_value = random.randint(0, 2**32 - 1) if randomize_seed or seed < 0 else int(seed)
@@ -2563,6 +2737,19 @@ def generate(
         msr_ref4_name = _save_ref(msr_ref4, "ref4") if msr_enabled else None
         msr_bg_name = _save_ref(msr_background, "bg") if any_msr else None
 
+        # Copy audio reference into comfy's INPUT dir so LoadAudio can find it.
+        audio_ref_name: str | None = None
+        if audio_ref_enabled and audio_ref_file:
+            try:
+                src = pathlib.Path(audio_ref_file)
+                if src.exists():
+                    ext = src.suffix.lower() or ".wav"
+                    audio_ref_name = f"input_audio_{uuid.uuid4().hex[:10]}{ext}"
+                    shutil.copy2(src, INPUT / audio_ref_name)
+            except Exception as e:
+                print(f"[audio_ref] failed to copy: {e}", flush=True)
+                audio_ref_name = None
+
         if msr_original:
             workflow = _inject_runexx_params(
                 _runexx_workflow_template(),
@@ -2605,6 +2792,17 @@ def generate(
                 omninft_bf16_lora_strength=omninft_bf16_lora_strength,
                 better_motion_lora_strength=better_motion_lora_strength,
                 physics_v2_lora_strength=physics_v2_lora_strength,
+                sulphur_audio_strength=sulphur_audio_strength,
+                sulphur_v1_audio_strength=sulphur_v1_audio_strength,
+                vbvr_audio_strength=vbvr_audio_strength,
+                dreamly_audio_strength=dreamly_audio_strength,
+                synth_audio_strength=synth_audio_strength,
+                plora_audio_strength=plora_audio_strength,
+                singularity_audio_strength=singularity_audio_strength,
+                omninft_audio_strength=omninft_audio_strength,
+                omninft_bf16_audio_strength=omninft_bf16_audio_strength,
+                better_motion_audio_strength=better_motion_audio_strength,
+                physics_v2_audio_strength=physics_v2_audio_strength,
                 cache_at_step=int(cache_at_step),
                 cache_warmup=int(cache_warmup),
                 energy_threshold=float(energy_threshold),
@@ -2623,6 +2821,10 @@ def generate(
                 prompt_segments=str(prompt_segments or ""),
                 kv_enabled=bool(kv_enabled),
                 kv_strength=float(kv_strength),
+                audio_ref_enabled=bool(audio_ref_enabled),
+                audio_ref_filename=audio_ref_name,
+                audio_ref_guidance_scale=float(audio_ref_guidance_scale),
+                audio_ref_stem_sep=bool(audio_ref_stem_sep),
             )
 
         mode_label = " (MSR-original)" if msr_original else (" (MSR)" if msr_enabled else "")
@@ -2631,17 +2833,17 @@ def generate(
             f"[gen] {width}x{height} {frames}f seed={seed_value} mode={mode} "
             f"preset={preset} sigmas={sigma_preset} face={mode} "
             f"kv={kv_enabled}@{kv_strength:.2f} "
-            f"sulphur_fro99={sulphur_lora_strength:.2f} "
-            f"sulphur_v1={sulphur_v1_lora_strength:.2f} "
-            f"vbvr={vbvr_lora_strength:.2f} "
-            f"dreamly={dreamly_lora_strength:.2f} "
-            f"synth={synth_lora_strength:.2f} "
-            f"plora={plora_lora_strength:.2f} "
-            f"singularity={singularity_lora_strength:.2f} "
-            f"omninft={omninft_lora_strength:.2f} "
-            f"omninft_bf16={omninft_bf16_lora_strength:.2f} "
-            f"better_motion={better_motion_lora_strength:.2f} "
-            f"physics_v2={physics_v2_lora_strength:.2f} "
+            f"sulphur_fro99={sulphur_lora_strength:.2f}/{sulphur_audio_strength:.2f} "
+            f"sulphur_v1={sulphur_v1_lora_strength:.2f}/{sulphur_v1_audio_strength:.2f} "
+            f"vbvr={vbvr_lora_strength:.2f}/{vbvr_audio_strength:.2f} "
+            f"dreamly={dreamly_lora_strength:.2f}/{dreamly_audio_strength:.2f} "
+            f"synth={synth_lora_strength:.2f}/{synth_audio_strength:.2f} "
+            f"plora={plora_lora_strength:.2f}/{plora_audio_strength:.2f} "
+            f"singularity={singularity_lora_strength:.2f}/{singularity_audio_strength:.2f} "
+            f"omninft={omninft_lora_strength:.2f}/{omninft_audio_strength:.2f} "
+            f"omninft_bf16={omninft_bf16_lora_strength:.2f}/{omninft_bf16_audio_strength:.2f} "
+            f"better_motion={better_motion_lora_strength:.2f}/{better_motion_audio_strength:.2f} "
+            f"physics_v2={physics_v2_lora_strength:.2f}/{physics_v2_audio_strength:.2f} "
             f"likeness={likeness_strength:.2f} "
             f"like_anchor={likeness_anchor_strength:.2f} "
             f"lat_anchor={latent_anchor_strength:.2f} "
@@ -2649,7 +2851,10 @@ def generate(
             f"anchor_sim={anchor_similarity_threshold:.2f} "
             f"energy={energy_threshold:.2f} "
             f"cache_step={cache_at_step} cache_warm={cache_warmup} "
-            f"relay={prompt_relay_enabled} input_mode={input_mode!r}",
+            f"relay={prompt_relay_enabled} input_mode={input_mode!r} "
+            f"audio_ref={audio_ref_enabled}@{audio_ref_guidance_scale:.1f} "
+            f"audio_stem_sep={audio_ref_stem_sep} "
+            f"audio_file={bool(audio_ref_name)}",
             flush=True,
         )
         result = _execute_workflow(workflow)
@@ -2855,6 +3060,74 @@ with gr.Blocks(title="10Eros LTX 2.3 image-to-video") as demo:
                     0.0, 2.0, value=1.0, step=0.05,
                     label="K/V strength (0 = off, 1 = funpack default, >1 = stronger identity)",
                 )
+            with gr.Accordion("audio", open=False):
+                audio_ref_enabled = gr.Checkbox(
+                    value=False,
+                    label="audio reference (voice ID transfer)",
+                )
+                audio_ref_guidance_scale = gr.Slider(
+                    0.0, 10.0, value=3.0, step=0.1,
+                    label="identity guidance scale (0 = no extra pass, 3 = default, higher = stronger voice ID)",
+                )
+                audio_ref_stem_sep = gr.Checkbox(
+                    value=False,
+                    label="isolate voice from background (stem separation, slower)",
+                )
+                audio_ref_file = gr.Audio(
+                    type="filepath",
+                    label="audio reference (~5s clip recommended)",
+                )
+                with gr.Accordion("per-lora audio strength (advanced)", open=False):
+                    gr.Markdown(
+                        "controls how each lora affects the **audio** stream "
+                        "(loras default to applying equally to video + audio). "
+                        "set to 0 to stop a lora from influencing audio while "
+                        "keeping its video effect."
+                    )
+                    sulphur_audio_strength = gr.Slider(
+                        0.0, 1.0, value=0.15, step=0.05,
+                        label="sulphur fro99 (audio)",
+                    )
+                    sulphur_v1_audio_strength = gr.Slider(
+                        0.0, 1.0, value=0.15, step=0.05,
+                        label="sulphur v1 (audio)",
+                    )
+                    vbvr_audio_strength = gr.Slider(
+                        0.0, 1.0, value=0.5, step=0.05,
+                        label="vbvr (audio)",
+                    )
+                    dreamly_audio_strength = gr.Slider(
+                        0.0, 1.0, value=0.6, step=0.05,
+                        label="dreamly (audio)",
+                    )
+                    synth_audio_strength = gr.Slider(
+                        0.0, 1.0, value=0.0, step=0.05,
+                        label="synth (audio)",
+                    )
+                    plora_audio_strength = gr.Slider(
+                        0.0, 1.0, value=0.0, step=0.05,
+                        label="plora (audio)",
+                    )
+                    singularity_audio_strength = gr.Slider(
+                        0.0, 1.0, value=0.3, step=0.05,
+                        label="singularity (audio)",
+                    )
+                    omninft_audio_strength = gr.Slider(
+                        0.0, 2.0, value=0.8, step=0.05,
+                        label="omninft converted (audio)",
+                    )
+                    omninft_bf16_audio_strength = gr.Slider(
+                        0.0, 2.0, value=0.0, step=0.05,
+                        label="omninft RL bf16 / kijai (audio)",
+                    )
+                    better_motion_audio_strength = gr.Slider(
+                        0.0, 1.0, value=0.0, step=0.05,
+                        label="better motion / mistic (audio)",
+                    )
+                    physics_v2_audio_strength = gr.Slider(
+                        0.0, 1.0, value=0.0, step=0.05,
+                        label="physics v2 / mistic (audio)",
+                    )
             with gr.Accordion("multi-reference settings (MSR)", open=False, visible=False) as msr_settings_acc:
                 msr_frame_count = gr.Dropdown(
                     [17, 25, 33, 41], value=41,
@@ -2943,6 +3216,17 @@ with gr.Blocks(title="10Eros LTX 2.3 image-to-video") as demo:
             omninft_bf16_lora_strength,
             better_motion_lora_strength,
             physics_v2_lora_strength,
+            sulphur_audio_strength,
+            sulphur_v1_audio_strength,
+            vbvr_audio_strength,
+            dreamly_audio_strength,
+            synth_audio_strength,
+            plora_audio_strength,
+            singularity_audio_strength,
+            omninft_audio_strength,
+            omninft_bf16_audio_strength,
+            better_motion_audio_strength,
+            physics_v2_audio_strength,
             cache_at_step,
             cache_warmup,
             energy_threshold,
@@ -2961,6 +3245,10 @@ with gr.Blocks(title="10Eros LTX 2.3 image-to-video") as demo:
             prompt_segments,
             kv_enabled,
             kv_strength,
+            audio_ref_enabled,
+            audio_ref_file,
+            audio_ref_guidance_scale,
+            audio_ref_stem_sep,
         ],
         outputs=[video, status, used_seed],
     )

requirements.txt

@@ -25,3 +25,4 @@ protobuf
 spandrel
 soundfile
 imageio-ffmpeg
+rotary_embedding_torch