Bernini V2 ComfyUI weights

Native ComfyUI packages for ByteDance/Bernini-Diffusers-v2. They are used by T8mars/comfyui-bernini-v2-T8 and do not require a Diffusers pipeline at runtime. Native ComfyUI Core support is under review in Comfy-Org/ComfyUI#16019. The obsolete sharded proposal in #16001 is closed. The current custom-node release is 0.3.3; it is reviewed and Active in the Comfy Registry. Core PR #16019 is mergeable with all CI checks passing and no unresolved review threads, and is awaiting maintainer review.

Bernini v2 first plans the requested scene and motion with Qwen2.5-VL, then renders the result with two Wan2.2 experts. It supports text-to-image, image-to-image, text-to-video, video-to-video, reference-to-video, and reference-guided video editing.

Available models

The native runtime uses ComfyUI's standard folders. Every listed component is one file; there is no Diffusers directory, shard index, or runtime manifest.

Published native-file checksums:

The Balanced INT8 package quantizes 1,300 linear layers across the planner, T5, and both Wan experts. Every converted layer was reconstructed and checked; no layer required a quality fallback at cosine 0.99 and relative-error 2% thresholds.

Both native profiles come from immutable upstream revision 399cf6a18a4c523b367b2b1ac25a2a61009e7df3. The Planner embeds the exact Qwen config and tokenizer; UMT5 embeds the released SentencePiece model. The two Wan experts each contain all 40 blocks and have matching normalized key contracts.

GGUF and NVFP4 are external renderer-only paths and do not replace the native planner package. The published GGUF pair has matching 1,095-tensor high/low contracts, all 40 Wan blocks, and a repaired F32 5-D patch embedding. Each file is 8,756,353,664 bytes. Balanced INT8 remains the recommended quality profile; GGUF Q4_K_S trades some facial and limb detail for lower renderer storage.

Installation

Install Bernini v2 (Native) from ComfyUI-Manager, or clone the node repository into ComfyUI/custom_nodes.

The current Registry release is 0.3.3 with status Active.

Download the recommended standalone INT8 files directly into the standard ComfyUI model folders:

hf download t8star/Bernini-V2-Comfy \
  --include "text_encoders/*_int8.safetensors" "diffusion_models/*_int8.safetensors" "vae/wan_2.1_vae.safetensors" \
  --local-dir /path/to/ComfyUI/models

For BF16 instead:

hf download t8star/Bernini-V2-Comfy \
  --include "text_encoders/*_bf16.safetensors" "diffusion_models/*_bf16.safetensors" "vae/wan_2.1_vae.safetensors" \
  --local-dir /path/to/ComfyUI/models

For the experimental GGUF renderer pair, first install ComfyUI-GGUF, then download the files to its diffusion-model directory:

hf download t8star/Bernini-V2-Comfy \
  --include "Bernini-v2-GGUF-Q4_K_S/*" \
  --local-dir /path/to/ComfyUI/models/diffusion_models

• High SHA-256: b72df0b32d305b7acade0a7245edde37716f7202f5fbc99fda43aedf5d1ebc87
• Low SHA-256: d396d3dcf935deb5bb1d8e6c6735e644adfa23d277efe0c2865bba03d6b7c92b
• High/low contract SHA-256: e779b82a06707b08f85d03e812293e4a14a96eccf00f189f239443e366351b76

The same repository includes the verified standard Wan 2.1 VAE. Download it directly into the ComfyUI model root:

hf download t8star/Bernini-V2-Comfy \
  --include "vae/wan_2.1_vae.safetensors" \
  --local-dir /path/to/ComfyUI/models

The resulting path is:

ComfyUI/models/vae/wan_2.1_vae.safetensors

The node repository also includes a pinned downloader:

python tools/download_vae.py --output /path/to/ComfyUI/models/vae

Open one of the six workflows under the node repository's examples/workflows/ directory after the files are in place.

Validated runtime

The current ComfyUI Core candidate completed the official T2V preset at 640×368, 33 frames, and 16 fps: 50 MaskGIT planning steps, one VIT denoising step, 50 flow-UniPC renderer steps, and the released T2V guidance values. A clean-reboot rerun with no competing GPU compute process finished in 1,024.578 seconds (about 17 minutes 5 seconds). All 33 decoded frames are unique. Peak ComfyUI-visible VRAM was 16.585 GiB on the 24 GB RTX 5090 Laptop GPU; peak process-tree RSS was 49.686 GiB and peak VMS/commit was 82.573 GiB.

An earlier accepted run of the same Core preset took 1,756.39 seconds and peaked at 23.255 GiB. The clean rerun is 41.7% faster, showing that a four-hour unfinished run is abnormal while PyTorch 2.7 + CUDA 12.8 eager/offload speed is still sensitive to host process and memory state. The new and preceding videos have all-frame SSIM 0.93627 and the same coherent fox-walk composition. Minor leg-gait and tail-root merging remains model output quality rather than a loader or quantization failure.

• Clean-reboot 2.063-second H.264 rerun
• Clean-reboot nine-frame contact sheet
• Clean-reboot full-resolution middle frame

The preceding accepted result remains available for comparison:

• 2.063-second H.264 result
• Five-frame contact sheet
• Full-resolution middle frame

The Balanced INT8 tensors passed a production-step 640×368, 33-frame T2V run on a 24 GB RTX 5090 Laptop GPU. ComfyUI-visible VRAM peaked at 16.51 GiB. Two consecutive uncached jobs in one server process produced identical decoded frames and returned memory to baseline. Reduced two-second, long-edge-640 T2V, V2V, R2V, and RV2V runs also passed.

The final standalone INT8 files additionally passed strict header/key/tokenizer validation, real construction of Planner, Core Wan UMT5, and both 14.288B Wan experts, plus a 640×368, 33-frame end-to-end 5/2/10-step format gate. It completed in 342.31 seconds with 33 unique frames and recognizable coherent fox motion. That reduced-step result validates packaging and execution, not production-preset detail quality. A 1/1/2-step result was intentionally rejected as a quality sample because it remained a blurred color mass.

The external NVFP4 renderer pair from rzgar/Bernini-v2-ComfyUI also passed a production-step 640×368, 33-frame T2V visual gate with the native Balanced-INT8 planner. All 33 frame hashes are unique and no quantization collapse is visible. The 1,518-second run used PyTorch 2.11/CUDA 12.8 eager operations because driver 576.28 cannot initialize CUDA 13, and another process occupied about 8.5 GiB of VRAM. It is quality evidence, not a valid optimized speed or isolated-memory comparison.

• NVFP4 2.063-second H.264 result
• NVFP4 nine-frame inspection grid
• NVFP4 full-resolution middle frame

The GGUF Q4_K_S renderer pair passed the same 640×368, 33-frame production-step gate with seed 42 and 25/5/40 planner/VIT/renderer steps. All 33 decoded frames are unique, motion is coherent, and there is no block-noise or frozen-frame collapse. Compared with Balanced INT8, the face is softer and isolated frames show an elongated lifted paw or mild rear-leg/tail-root merging. This is a usable experimental low-storage lane, not a quality-parity claim. The run took 1,319.907 seconds on the older PyTorch 2.7/CUDA 12.8 compatibility path; its memory samples are not an isolated benchmark.

• GGUF Q4_K_S 2.063-second H.264 result
• GGUF Q4_K_S nine-frame inspection grid
• GGUF Q4_K_S full-resolution middle frame

Detailed settings, hashes, and visual evidence are available in the quality report and low-memory guide.

License and attribution

The original Bernini v2 weights and code are released under Apache License 2.0. These dtype-converted and quantized packages remain under Apache-2.0. The GGUF conversion policy and 5-D repair workflow are adapted from the Apache-2.0 city96/ComfyUI-GGUF tooling; the runtime loader is not bundled. The Wan 2.1 VAE companion is an unchanged mirror of Comfy-Org's Apache-2.0 repackage at revision c4f60d30c55a624e35427060fdd217579a6c1d77; its SHA-256 is 2fc39d31359a4b0a64f55876d8ff7fa8d780956ae2cb13463b0223e15148976b.

• Original model: https://huggingface.co/ByteDance/Bernini-Diffusers-v2
• Original source: https://github.com/bytedance/Bernini
• Native ComfyUI nodes: https://github.com/T8mars/comfyui-bernini-v2-T8

Links

• Bilibili: https://space.bilibili.com/385085361
• YouTube: https://www.youtube.com/@T8star-Aix/
• API: https://api.seedance.nz/sign-up?aff=5f4w
• Online AI apps: https://www.runninghub.ai/zh-cn/user-center/1907375370302308353/userPost?inviteCode=rh-v1121
• Hugging Face: https://huggingface.co/t8star