SceneWorks/pulid-flux-mlx

Converted weights for the native MLX PuLID-FLUX face-identity path in the SceneWorks Rust GPU worker (mlx-gen-pulid, epic 3069 / cutover sc-3344). These are the PuLID-FLUX-specific inputs the worker fetches on first use; the FLUX.1-dev backbone (black-forest-labs/FLUX.1-dev) and the PuLID adapter (guozinan/PuLID/pulid_flux_v0.9.1.safetensors) come from their own repos, and the SCRFD detector + ArcFace embedder are reused from SceneWorks/instantid-mlx.

file	what	provenance
`eva02_clip_l_336.safetensors`	EVA02-CLIP-L-14-336 visual tower (f16, bare mlx-key names) — produces PuLID's `id_cond_vit` + hidden states	converted from `QuanSun/EVA-CLIP/EVA02_CLIP_L_336_psz14_s6B.pt` via `mlx-gen/tools/convert_eva_clip.py`
`bisenet_parsing.safetensors`	BiSeNet face-parsing net — drives `face_features_image` (background-whitened aligned crop) in `mlx-gen-face`	BiSeNet face-parsing weights

The worker loads these via the engine's env-var seam (PULID_EVA_WEIGHTS / PULID_FACE_WEIGHTS_DIR); mlx-gen-pulid casts the EVA tower to f32 at load (the conditioning path runs f32), so the f16 store is exact after the upcast.

Downloads last month: -; Downloads are not tracked for this model. How to track

MLX

Hardware compatibility

Quantized

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support