LTX-Video

Whilst we wait for Hunyuan Video.

nodes.py

@@ -62,7 +62,7 @@ def gguf_sd_loader(path, handle_prefix="model.diffusion_model."):
         if len(arch_field.types) != 1 or arch_field.types[0] != gguf.GGUFValueType.STRING:
             raise TypeError(f"Bad type for GGUF general.architecture key: expected string, got {arch_field.types!r}")
         arch_str = str(arch_field.parts[arch_field.data[-1]], encoding="utf-8")
-        if arch_str not in {"flux", "sd1", "sdxl", "sd3", "aura", "t5", "t5encoder"}:
+        if arch_str not in {"flux", "sd1", "sdxl", "sd3", "aura", "ltxv", "t5", "t5encoder"}:
             raise ValueError(f"Unexpected architecture type in GGUF file, expected one of flux, sd1, sdxl, t5encoder but got {arch_str!r}")
     else: # stable-diffusion.cpp
         # import here to avoid changes to convert.py breaking regular models

tools/convert.py

@@ -41,6 +41,16 @@ class ModelAura(ModelTemplate):
     ]
     keys_banned = ["joint_transformer_blocks.3.ff_context.out_projection.weight",]
 
+class ModelLTXV(ModelTemplate):
+    arch = "ltxv"
+    keys_detect = [
+        (
+            "adaln_single.emb.timestep_embedder.linear_2.weight",
+            "transformer_blocks.27.scale_shift_table",
+            "caption_projection.linear_2.weight",
+        )
+    ]
+
 class ModelSDXL(ModelTemplate):
     arch = "sdxl"
     shape_fix = True
@@ -65,7 +75,7 @@ class ModelSD1(ModelTemplate):
     ]
 
 # The architectures are checked in order and the first successful match terminates the search.
-arch_list = [ModelFlux, ModelSD3, ModelAura, ModelSDXL, ModelSD1]
+arch_list = [ModelFlux, ModelSD3, ModelAura, ModelLTXV,  ModelSDXL, ModelSD1]
 
 def is_model_arch(model, state_dict):
     # check if model is correct