ComfyUI Node
RAVEN Model Loader
Loads the official full, non-pruned BF16 MiniMax H3 diffusion model together with the mandatory RAVEN PEFT LoRA and returns a standard MODEL. The RAVEN residual strength is fixed at 1.0: the adapter is what the streaming sampler's 4-NFE consistency schedule was trained with, not a quality knob, so there is no 'off' setting. Weights land on the stock ModelPatcher partial CPU offload path (v0.1's only supported memory strategy). 'fp32' doubles the base weights to 132 GB+ on top of the FP32 residual - expect heavy offload traffic or OOM.
RAVEN Model Loader
- MODEL
◄unet_name▾►
◄lora_name▾►
◄weight_dtypedefault►
Categorymodel/loaders/raven
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| unet_name | COMBO | Official full, non-pruned BF16 MiniMax-H3 DiT. Pruned/adaln-curve checkpoints are rejected. The RAVEN adapter is trained against the full BF16 weights; the pruned/adaln-curve checkpoint has no time_embedder for its 266-module mapping to attach to. | |
| lora_name | COMBO | Mandatory RAVEN PEFT LoRA (532 tensors / 266 modules, FP32 A/B). Without it this node cannot produce a model. About 5 GB of FP32 A/B tensors, applied as an activation residual (never fused into the BF16 weights) and counted by Comfy's memory accounting from the first model_size() call. | |
| weight_dtype | COMBO | default | 'default' lets comfy.model_management pick (BF16 for H3). 'fp32' doubles the weights to 132 GB+ for the full non-pruned model, on top of the FP32 RAVEN residual - expect heavy offload traffic or OOM unless you know the box can hold it. There is no FP8/INT8 choice on purpose: comfy.ops fuses quantised linears and would silently skip the residual. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | A standard ComfyUI MODEL (stock static ModelPatcher) whose diffusion_model is the chunk-causal RAVEN DiT. Stock LoraLoaderModelOnly can be chained after it, and nothing here restricts it to one generation mode - other official H3 workflows can be explored with it, unverified and unsupported. The RAVEN Streaming Sampler in this pack is the narrower part: it refuses conditioning carrying condition/reference rows it has not implemented. |