ComfyUI Node
JR H3 VEDA Attention (Experimental)
Independent VEDA sparse attention for native MiniMax H3. Use a clean MODEL branch after the loader/LoRA, last before sampling. Requires official Veda-on-ComfyUI. Choose this branch or Sol-H3; existing Unified/Sol nodes are unchanged.
JR H3 VEDA Attention (Experimental)
- model
- model
- status
◄enabletrue►
◄predictorminimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors►
◄generated_sparsity90%►
◄reference_sparsity90%►
◄full_attention_layers►
◄full_attention_steps►
◄verbosefalse►
◄sage_attentiondisabled►
◄enable_low_vram_attentiontrue►
◄head_chunks4►
◄enable_low_vram_ffntrue►
◄ffn_chunks4►
◄ffn_seq_threshold4096►
CategoryJR MiniMax H3/Optimization
Inputs (14)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| enable | BOOLEAN | true | — |
| predictor | COMBO | minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors | 1 options: minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors |
| generated_sparsity | STRING | 90% | 90% skips 90% of visual key tiles. An integer such as 32 keeps that many tiles. |
| reference_sparsity | STRING | 90% | 0% keeps visual references dense. Text/audio stay dense in VEDA. |
| full_attention_layers | STRING | 0-based full-attention layers, e.g. 0,1,47-49. | |
| full_attention_steps | STRING | 0-based full-attention steps, e.g. 0. | |
| verbose | BOOLEAN | false | — |
| sage_attention | COMBO | disabled | Dense fallback only. Disabled uses ComfyUI's configured default attention. |
| enable_low_vram_attention | BOOLEAN | true | — |
| head_chunks | INT | 41–56 | — |
| enable_low_vram_ffn | BOOLEAN | true | — |
| ffn_chunks | INT | 41–64 | — |
| ffn_seq_threshold | INT | 4096256–262144 | — |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |
| status | STRING | — |