LTX2SamplingPreviewOverride is incompatible with ModelPatcherDynamic after ComfyUI #15063

[Kinasa0096]

Description

After ComfyUI PR #15063 changed the native LTX latent upscaler to use
ModelPatcherDynamic, LTX2SamplingPreviewOverride crashes before sampling
begins.

The native LatentUpscaleModelLoader now returns a model patcher for the LTX
spatial upscaler, but the KJNodes preview wrapper still treats the input as a
plain torch.nn.Module and calls .to(device) directly.

Error

AttributeError: 'ModelPatcherDynamic' object has no attribute 'to'

File "<COMFYUI>/custom_nodes/ComfyUI-KJNodes/nodes/ltxv_nodes.py",
line 899, in __call__
    self.latent_upscale_model.to(device)

The sampler shown as the failing workflow node is SamplerCustomAdvanced, but
the exception originates from OuterSampleCallbackWrapper registered by
LTX2SamplingPreviewOverride.

Steps to reproduce

1. Use a current ComfyUI version containing upscalers: convert latent_upsampler model to ModelPatcherDynamic Comfy-Org/ComfyUI#15063.
2. Load the LTX 2.3 spatial upscaler through the native
   LatentUpscaleModelLoader.
3. Connect it to the optional latent_upscale_model input of
   LTX2SamplingPreviewOverride, together with an LTX VAE and model.
4. Start sampling with SamplerCustomAdvanced.

The exception occurs before the first sampling step.

Disconnecting only the latent_upscale_model input from
LTX2SamplingPreviewOverride avoids the crash. The native
LTXVLatentUpsampler continues to work with the same upscaler.

Cause

ComfyUI PR #15063 converted the LTX latent upscaler to the managed dynamic-VRAM
path:

Comfy-Org/ComfyUI#15063

KJNodes currently performs direct module operations in the preview wrapper:

self.latent_upscale_model.to(device)
...
x0 = self.latent_upscale_model(x0.to(torch.bfloat16))
...
self.latent_upscale_model.to(mm.unet_offload_device())

With current ComfyUI, this object is a ModelPatcherDynamic, not the underlying
module.

The updated native implementation uses model_management.load_models_gpu(),
reads the underlying model from upscale_model.model, and obtains its device
and dtype from the patcher.

Expected behavior

LTX2SamplingPreviewOverride should support the current
ModelPatcherDynamic latent-upscaler representation, while ideally remaining
compatible with older plain-module inputs.

Environment

• OS: Windows
• GPU: NVIDIA GeForce RTX 5080
• DynamicVRAM: enabled
• ComfyUI: 0.28.0, commit 45ffd543
• ComfyUI-KJNodes: commit e27a505b
• Python: 3.12.10
• PyTorch: 2.10.0+cu130

Report prepared with assistance from OpenAI Codex and reviewed by the reporter.