MetaView Model Loader

If MetaViewDA3Loader supplies the geometry and MetaView3DConditioning bakes it into the conditioning, MetaView Model Loader (class MetaViewModelLoader) supplies the thing that actually renders the new view: the MetaView DiT. It's a grafted Qwen-Image-Edit transformer - 60 double-stream blocks - with an extra attention branch that attends the 3D features extracted by Depth Anything 3, plus a PRoPE positional scheme that encodes the camera. This node loads that model as a standard ComfyUI MODEL so it drops into a normal SamplerCustom graph.

The short version of the setup: one input, one output, one model file. There's not much to tune, which is the point.

How it works

The loader grabs the model file from models/diffusion_models and loads it onto ComfyUI's native fp8 quantized op path. The weights are fp8-scaled: they stay fp8 on disk and in VRAM, and are dequantized to bf16 for the actual matmuls (full_precision_matrix_mult on). That matches the MetaView reference's --no_fp8_mm path - fp8 storage with full-precision compute, so you get the memory savings without the precision gamble. Compute dtype is set to bf16.

The loader also registers the model's sampling config: Qwen-Image / flow-matching, with a latent format matching the WAN 2.1 VAE (16 channels, /8 spatial). That's what tells the sampler how to treat the latents it's handed. It's all standard ComfyUI machinery, just configured for this one specific model.

Inputs and output

• model_name - a dropdown over everything in models/diffusion_models. Pick metaview_dit_fp8scaled.safetensors. If your file doesn't show up, refresh node definitions.
• model - the patched MODEL. Wire it into SamplerCustom (which also gets MetaViewSigmas sigmas and the latent from MetaView3DConditioning).

The text encoder and VAE for this pipeline come from stock ComfyUI nodes - a CLIPLoader with type qwen_image for the Qwen2.5-VL encoder and a VAELoader for the Qwen-Image VAE. This node is deliberately just the renderer.

Model file and install

Install the Enviral Design Node Pack (Manager → "Enviral Design Node Pack", or clone into ComfyUI/custom_nodes and restart). Then the actual model download:

ComfyUI/models/diffusion_models/metaview_dit_fp8scaled.safetensors

That file is the MetaView DiT artifact, and it's a real download - this is the heavy model in the pipeline. Place it in models/diffusion_models, refresh, and it appears in the dropdown. The pack creates no junctions or symlinks, so a model manager or a manual copy is on you.

Troubleshooting

• Not in the dropdown. Standard ComfyUI list caching - press R on the canvas or Refresh Node Definitions after dropping the file in.
• "Leftover state dict keys" warning in the console. The loader logs a warning if the checkpoint has keys it didn't consume. Usually harmless, but if your renders come out wrong, it's a sign the file doesn't match what the loader expects - double-check you have the right artifact and not a mismatched reupload.
• Weird sampler results. Don't improvise the sampler settings. The weights are fused Lightning: 8 steps, CFG 1.0, euler, with the MetaViewSigmas schedule and the latent from MetaView3DConditioning. Stray from that and the model won't show what it can do.

This is the quiet node in the MetaView trio: no knobs, one file, done. Get it wired to SamplerCustom and let the conditioning node and sigmas node handle the drama.