ComfyUI Node
BSAI Prism Model Loader
A ComfyUI node in BSAI/Prism with 22 inputs and 2 outputs.
BSAI Prism Model Loader
- prism_model
- status
◄base_modelMOVA-360p►
◄resume_ckptpreview_alpha►
◄deviceN/A►
◄offloadcpu►
◄enable_bsatrue►
◄bsa_sparsity0.75►
◄bsa_chunk_3d_shape_q4 4 4►
◄bsa_chunk_3d_shape_k4 4 4►
◄bsa_cdf_threshold0.20►
◄enable_ivpq_dynamic_blocktrue►
◄enable_layer_adaptive_dynamic_blockfalse►
◄sparse_high_noise_onlyfalse►
◄enable_bsa_v2afalse►
◄bsa_v2a_sparsity0.88►
◄bsa_v2a_audio_chunk_size64►
◄bsa_v2a_chunk_3d_shape_k4 4 4►
◄enable_audio_guidancefalse►
◄audio_boost_gamma2.0►
◄enable_variance_guidancefalse►
◄variance_boost_gamma2.0►
◄enable_taylor_sparse_attnfalse►
◄enable_rectified_sparse_attnfalse►
CategoryBSAI/Prism
Inputs (22)
| Name | Type | Default | Description |
|---|---|---|---|
| base_model | COMBO | MOVA-360p | MOVA base model (architecture skeleton, no fine-tuned weights). Defines the overall structure, VAE, text_encoder and other base components. |
| resume_ckpt | COMBO | preview_alpha | Prism fine-tuned weights (preview_alpha / preview_beta). Loaded on top of the base; determines generation quality and style. Choose (None) to use the base only (not recommended, poor quality). |
| device | COMBO | N/A | Runtime device. Prism depends on flash-attn; NVIDIA CUDA only. |
| offload | COMBO | cpu | cpu = encoder/VAE moved in and out of GPU per stage, saves VRAM on a single GPU (official default); none = fully resident on GPU, higher peak usage. |
| enable_bsa | BOOLEAN | true | Enable Prism Block Sparse Attention (video self-attention top-k/top-p sparsity). |
| bsa_sparsity | FLOAT | 0.750.1–0.95 | top-k sparsity (official recommended 0.93 / 0.85 / 0.75; smaller = sparser/faster). |
| bsa_chunk_3d_shape_q | STRING | 4 4 4 | 3D macro-block shape (T H W) for video self-attention Q; official optimum 4 4 4. |
| bsa_chunk_3d_shape_k | STRING | 4 4 4 | 3D macro-block shape (T H W) for video self-attention K; official optimum 4 4 4. |
| bsa_cdf_threshold | FLOAT | 0.200–0.5 | CDF threshold for mixed Top-k + Top-p (official 0.16~0.4, optimum 0.2). 0 = pure Top-k. |
| enable_ivpq_dynamic_block | BOOLEAN | true | Prism core: anisotropic block shape adapted to local content variance and audio-video coupling. When enabled, it is mutually exclusive with audio/variance guidance and Taylor/Rectified (conflicts raise an error). |
| enable_layer_adaptive_dynamic_block | BOOLEAN | false | Shallow fixed blocks / deep dynamic blocks + shallow->deep audio caching (only effective when dynamic blocks are on). |
| sparse_high_noise_only | BOOLEAN | false | true = sparse attention only on the high-noise expert video_dit; the low-noise expert video_dit_2 keeps full attention. |
| enable_bsa_v2a | BOOLEAN | false | Bridge v2a cross-attention BSA (Q=audio, K=video 3D blocking). |
| bsa_v2a_sparsity | FLOAT | 0.880.1–0.95 | — |
| bsa_v2a_audio_chunk_size | INT | 6432–512 | audio Q block size (multiple of 64). |
| bsa_v2a_chunk_3d_shape_k | STRING | 4 4 4 | — |
| enable_audio_guidanceopt | BOOLEAN | false | Audio guidance BSA (Path A; mutually exclusive with dynamic blocks). |
| audio_boost_gammaopt | FLOAT | 2.00–4 | — |
| enable_variance_guidanceopt | BOOLEAN | false | Channel variance guidance BSA (mutually exclusive with dynamic blocks). |
| variance_boost_gammaopt | FLOAT | 2.00–4 | — |
| enable_taylor_sparse_attnopt | BOOLEAN | false | Taylor sparse-attention bias correction (mutually exclusive with Rectified and with dynamic blocks). |
| enable_rectified_sparse_attnopt | BOOLEAN | false | Rectified sparse-attention bias correction (mutually exclusive with Taylor and with dynamic blocks). |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| prism_model | BSAI_PRISM_MODEL | — |
| status | STRING | — |