Fibo 2.0 Edit Turbo INT8-ConvRot
Native ComfyUI transformer checkpoint derived from Bria AI's Fibo 2.0 Edit Turbo, revision 177f2d4a06b6587144ef5c2b8452a4f1fe2ab4b7. Private repository.
Bria AI owns the original model. The original and adapted weights retain CC BY-NC 4.0, attribution and the original source reference. Noncommercial use only. Commercial use requires a separate agreement with Bria AI. See LICENSE for the complete terms. Source code has its separate licenses in the native pack.
| File | Bytes | SHA-256 |
|---|---|---|
fibo2-edit-turbo-int8-convrot.safetensors |
8778351546 | cb058c043a3b1597ed5765ffaea3883bedf936c692cc771b3e8d529e431e7817 |
The transformer uses 342 native int8_tensorwise W8A8 Linear matrices. ConvRot groups are 256 for 296 matrices and 64 for the 46 original 9,536-column down projections. Shapes are unchanged. Scales are FP32 per output row, and activations are dynamically quantized inside comfy-kitchen. All 441 protected tensors remain byte-identical at source BF16, including the Perceiver, input/output projections, normalization and timestep/modulation parameters.
Load locally
Download ComfyUI-Fibo2-Native.zip, extract its directory into ComfyUI/custom_nodes/, then follow INSTALL.md. Put the checkpoint in ComfyUI/models/diffusion_models/. Select it with Fibo 2 Native Model. Reuse a complete compatible Qwen3-VL-4B encoder and a FLUX.2 VAE, or obtain those components from the official upstream repository.
Three visual workflows and paired API graphs cover instruction edits, masks and ordered references. Use four Euler steps, CFG 1 and the supplied dimension-dependent scheduler. The native pack preserves the six intermediate Qwen bundles, Perceiver and five injection sites, VLM/VAE reference conditioning, grey-mask preprocessing and rendered-text conditioning. Runtime inference is local.
This single-file transformer is a native ComfyUI artifact, rather than a complete Diffusers pipeline directory. For original Diffusers inference, use the pinned official model and Diffusers implementation listed in the installation guide.
Validation
The native BF16 transformer reproduced all 16 saved upstream velocity predictions exactly in a four-task panel. For fixed upstream conditioning, the INT8 images reached SSIM 0.980453–0.998355, with observed cloud denoising latency 15–22% below native BF16. Actual signed-integer CUDA GEMMs and native activation quantization were traced. See BENCHMARKS.md, validation-metrics.json and quantization-report.json.
The full native encoder/VAE conditioning control reached SSIM 0.983341–0.998447 in the same panel. Encoder/VAE numerical differences and existing quantized-encoder effects are separate from denoiser-only quantization. This small panel does not establish universal near-lossless quality, identity, typography or every reference/mask configuration.
Reproduce
The source revisions, conversion command, resumable staging, precision policy and independent tests are in INSTALL.md and the archive. quantization-policy.json records the recipe. No calibration or lower-bit formats were used. The supplied model filename preserves the ordinary source namespaces and native per-layer .comfy_quant metadata.