Fibo 2.0 Edit Turbo — fp8 ComfyUI weights
fp8 (float8_e4m3fn) checkpoint of the Fibo 2.0 Edit Turbo
transformer for ComfyUI — 16.3 GB → 8.1 GB, half the download and half the resident
VRAM of the bf16 original. Cast from
theunlikely/Fibo-2.0-edit-turbo-ComfyUI's
lossless bf16 conversion: weight matrices (ndim ≥ 2) are e4m3fn, norms/biases/scalars stay
bf16 — the standard conservative FLUX-fp8 convention. No training, no fine-tuning.
| File | Destination |
|---|---|
| diffusion_models/fibo2_edit_turbo_fp8_e4m3fn.safetensors | ComfyUI/models/diffusion_models/ |
| Qwen3-VL-4B BF16 | ComfyUI/models/text_encoders/ |
| FLUX.2 VAE | ComfyUI/models/vae/ |
Runs with the native nodes in akhaliq/Fibo-2.0-ComfyUI
(fibo2_native/ — ModelPatcher-based, works with stock samplers/guiders/VAE, and loads
every FIBO checkpoint layout: this fp8 file, diffusers-layout bf16, and the official
Comfy-Org/Fibo-2.0 fused-key bf16 + int8_convrot).
The fp8 checkpoint loads directly; weights are upcast to bf16 on load in this version
(native fp8 compute is on the roadmap), so VRAM during sampling matches bf16 while the
download halves.
Note: Comfy-Org now ships an official int8_convrot quant (Hadamard-rotated, adaLN
kept bf16) — for most users that's the better-quantized download. This fp8 file remains
useful where e4m3 is preferred (Hopper-native fp8 kernels, diffusers runtimes).
Status: experimental. fp8 inference quality for this architecture is not yet validated by Bria or the community — if outputs look off, fall back to bf16 and please open a discussion. Report results either way; early datapoints are valuable.
License
CC BY-NC 4.0, inherited from the source model — non-commercial use, attribution to Bria AI.
Commercial use requires a separate agreement with Bria.
Weights provenance: briaai/Fibo-2.0-edit-turbo @ e976703, via theunlikely's lossless
conversion. Not affiliated with or endorsed by Bria or Comfy Org.