feat(ltx): canonical LTX-2.x Stage 2 recipe (LoRA + guidance + connectors)

Implement the Lightricks two-stage recipe (diffusers PR #13217) for the
LTX-2.x Dev family: Stage 1 at half-res with full four-way guidance,
2x latent upsample, Stage 2 with distilled LoRA + scheduler swap + identity
guidance on STAGE_2_DISTILLED_SIGMA_VALUES.

Extends to both LTX-2.0 and LTX-2.3 Dev via per-family distilled-LoRA
repos carried on the caps; Distilled variants take the same flow minus
the LoRA swap. Auto-couples Refine with a fixed 2x upsample on any Dev
variant with a known LoRA when the user enables Refine without Upsample.

- caps: is_ltx_2_3, use_cross_timestep, default_dynamic_shift,
  stage2_dev_lora_repo, supports_canonical_stage2, modality_default_scale,
  guidance_rescale_default; LTX-2.x defaults realigned to canonical
  cfg=3.0 / steps=30; per-variant STG block and four-way guidance wired
  for non-distilled 2.x
- process: canonical Stage 1/Stage 2 helpers, scheduler + opts snapshot
  under try/finally, per-family upsampler repo, audio latents threaded
  from Stage 1 into Stage 2, use_cross_timestep gated per caps
- overrides: skip the redundant unsharded LTX-2.3 connectors blob and
  share LTX2TextConnectors weights across 2.3 variants when te_shared_t5
- load: Gemma3 shared-TE path for LTX-2.3; gate use_dynamic_shifting=False
  override to 0.9.x only so LTX-2.x stays on its canonical token-count
  dynamic shift

modules/ltx/ltx_capabilities.py

@@ -10,23 +10,34 @@ class LTXCaps:
     repo_cls_name: str
     family: str  # '0.9' or '2.x'
     is_distilled: bool
+    is_ltx_2_3: bool
     is_i2v: bool
-    supports_input_media: bool  # accordion visible for any pipeline that accepts image/video input
-    supports_multi_condition: bool  # uses conditions=[LTX(2)VideoCondition(...)] kwarg; Condition classes only
+    supports_input_media: bool
+    supports_multi_condition: bool
     supports_image_cond_noise_scale: bool
     supports_decode_timestep: bool
     supports_stg: bool
     supports_audio: bool
     supports_frame_rate_kwarg: bool
+    # 2.3 transformer cross-attn reads the other modality's sigma; unset falls back to 2.0's
+    # independent-sigma path, which is a joint-distribution mismatch for 2.3 weights.
+    use_cross_timestep: bool
     default_cfg: float
     default_steps: int
     default_sampler_shift: float
+    default_dynamic_shift: bool
     default_width: int
     default_height: int
     default_frames: int
     default_frame_rate: int
     stg_default_scale: float = 0.0
     stg_default_blocks: list = field(default_factory=list)
+    # Dev 2.x trained under cfg + stg + modality + rescale four-way composition;
+    # distilled bakes these into its sigma schedule and stays at pipeline identity.
+    modality_default_scale: float = 1.0
+    guidance_rescale_default: float = 0.0
+    supports_canonical_stage2: bool = False
+    stage2_dev_lora_repo: Optional[str] = None
 
 
 CONDITION_CLASSES = {'LTXConditionPipeline', 'LTX2ConditionPipeline'}
@@ -69,12 +80,14 @@ def get_caps(model_name: str) -> Optional[LTXCaps]:
     is_i2v = 'I2V' in model_name or cls_name in ('LTXImageToVideoPipeline', 'LTX2ImageToVideoPipeline')
     is_condition_cls = cls_name in CONDITION_CLASSES
     supports_input_media = is_i2v or is_condition_cls
+    is_ltx_2_3 = is_ltx2 and '2.3' in model_name
 
     caps = LTXCaps(
         name=model_name,
         repo_cls_name=cls_name,
         family=family,
         is_distilled=is_distilled,
+        is_ltx_2_3=is_ltx_2_3,
         is_i2v=is_i2v,
         supports_input_media=supports_input_media,
         supports_multi_condition=is_condition_cls,
@@ -83,9 +96,11 @@ def get_caps(model_name: str) -> Optional[LTXCaps]:
         supports_stg=is_ltx2,
         supports_audio=is_ltx2,
         supports_frame_rate_kwarg=is_ltx2,
-        default_cfg=4.0 if is_ltx2 else 3.0,
-        default_steps=40 if is_ltx2 else 50,
+        use_cross_timestep=is_ltx_2_3,
+        default_cfg=3.0,
+        default_steps=30 if is_ltx2 else 50,
         default_sampler_shift=-1.0,
+        default_dynamic_shift=is_ltx2,
         default_width=768,
         default_height=512,
         default_frames=121 if is_ltx2 else 161,
@@ -96,6 +111,13 @@ def get_caps(model_name: str) -> Optional[LTXCaps]:
         caps.default_cfg = 1.0
         caps.default_steps = 8
 
+    if is_ltx2 and not is_distilled:
+        if is_ltx_2_3:
+            caps.stage2_dev_lora_repo = 'CalamitousFelicitousness/LTX-2.3-distilled-lora-384-Diffusers'
+        elif '2.0' in model_name:
+            caps.stage2_dev_lora_repo = 'CalamitousFelicitousness/LTX-2.0-distilled-lora-384-Diffusers'
+    caps.supports_canonical_stage2 = caps.stage2_dev_lora_repo is not None
+
     if is_ltx2:
         if '2.3' in model_name:
             caps.stg_default_blocks = [28]
@@ -103,6 +125,10 @@ def get_caps(model_name: str) -> Optional[LTXCaps]:
             caps.stg_default_blocks = [29]
         else:
             caps.stg_default_blocks = [28]
-        caps.stg_default_scale = 0.0
+        if not is_distilled:
+            # canonical T2V composition from huggingface/diffusers#13217
+            caps.stg_default_scale = 1.0
+            caps.modality_default_scale = 3.0
+            caps.guidance_rescale_default = 0.7
 
     return caps