| SeedVR2-1.4B — ComfyUI native support patch (MANUAL ALTERNATIVE — most people want the folder) |
| |
|
|
| >>> PREFER `ComfyUI-SeedVR2-1.4B/` INSTEAD OF THIS FILE. <<< |
| >>> |
| >>> Copy that folder into ComfyUI/custom_nodes/ and restart. It does everything this patch does |
| >>> WITHOUT editing any ComfyUI source file (so it survives updates), and it ALSO adds a |
| >>> single-frame VAE fast path worth 3-4.6x on decode. This file is kept only for people who |
| >>> would rather edit one line of ComfyUI than install a folder into custom_nodes/. |
| >>> |
| >>> Apply ONE of the two, never both. |
|
|
| WHAT THIS IS |
| |
| ComfyUI has native SeedVR2 support. It works out which architecture to build by inspecting the |
| checkpoint in `comfy/model_detection.py`, and every existing branch keys on `blocks.31` or |
| `blocks.35`. This model has six blocks (0-5), so no branch matches and the model will not load. |
|
|
| This patch adds one branch that recognises it. Nothing else changes. |
|
|
| The values are not guesses — they are the existing 7B "separate key" branch with the depth changed, |
| and each was cross-checked against the published weights: |
|
|
| vid_dim 3072 <- blocks.0.ada.vid.attn_gate has shape [3072] |
| heads 24 <- 24 x head_dim 128 = 3072 = vid_dim |
| num_layers 6 <- the checkpoint has blocks.0 .. blocks.5 and no others |
| mm_layers 6 <- every block carries BOTH .vid. and .txt. streams (31 tensors each) |
| mlp_type "normal" <- keys are mlp.vid.proj_in, NOT mlp.vid.proj_in_gate (which would be swiglu) |
| rope_type "rope3d", rope_dim 64, norm_eps 1e-5 <- unchanged from the 7B branch |
|
|
| The guard matters: it requires `blocks.5` to be PRESENT and `blocks.6` to be ABSENT, so it cannot |
| capture the real 7B (which has blocks 0-35) or the 3B (blocks 0-31). |
|
|
|
|
| HOW TO APPLY |
| |
| Edit `comfy/model_detection.py` in your ComfyUI install. Find this line (search for "seedvr2_7b_separate_key"): |
|
|
| seedvr2_7b_separate_key = "{}blocks.35.mlp.vid.proj_out.weight".format(key_prefix) |
|
|
| Insert the block below IMMEDIATELY BEFORE that line, at the same indentation (4 spaces). |
|
|
|
|
| THE PATCH |
| |
| # SeedVR2-1.4B — 6-layer distillation of the 7B. |
| # Same architecture as the 7B separate-key variant below (vid_dim 3072, heads 24, separate |
| # vid/txt MMModule keys in every block, non-gated MLP) but 6 blocks instead of 36. |
| seedvr2_1_4b_key = "{}blocks.5.mlp.vid.proj_out.weight".format(key_prefix) |
| if (seedvr2_1_4b_key in state_dict_keys |
| and "{}blocks.6.mlp.vid.proj_out.weight".format(key_prefix) not in state_dict_keys |
| and state_dict[seedvr2_1_4b_key].shape[0] == 3072): |
| dit_config = {} |
| dit_config["image_model"] = "seedvr2" |
| dit_config["vid_dim"] = 3072 |
| dit_config["heads"] = 24 |
| dit_config["num_layers"] = 6 |
| dit_config["mm_layers"] = 6 |
| dit_config["norm_eps"] = 1e-5 |
| dit_config["rope_type"] = "rope3d" |
| dit_config["rope_dim"] = 64 |
| dit_config["mlp_type"] = "normal" |
| return dit_config |
|
|
|
|
| WHERE TO PUT THE MODEL FILES |
| |
| seedvr2_distill_6L_1.4B_sharp_fp16.safetensors -> ComfyUI/models/diffusion_models/ |
| ema_vae_fp16.safetensors -> ComfyUI/models/vae/ |
|
|
| Load the transformer with **Load Diffusion Model** (not Load Checkpoint — this is a bare diffusion |
| model, not a packaged checkpoint with bundled VAE/text-encoder/conditioning), and the VAE with |
| **Load VAE**. The `SeedVR2 Preprocess` / `SeedVR2 Conditioning` / `SeedVR2 PostProcessing` nodes then |
| work exactly as they do for the stock SeedVR2 models. |
|
|
|
|
| STATUS |
| |
| VERIFIED. With this patch applied, ComfyUI's own detect_unet_config() reads the checkpoint and |
| returns: |
|
|
| image_model seedvr2 |
| num_layers 6 |
| mm_layers 6 |
| vid_dim 3072 |
| heads 24 |
| mlp_type normal |
| rope_type rope3d |
| rope_dim 64 |
|
|
| i.e. it builds the correct architecture. Tested against ComfyUI's real detection code, not a mock. |
|
|
| If ComfyUI upstream adds official support for this model, prefer that over this patch. |
|
|
