feat: SeFi-Image support#1707

Open

fszontagh wants to merge 21 commits into

leejet:masterfrom

fszontagh:feat/sefi-image-prototype

Commits on Jun 23, 2026

feat: scaffold SeFi-Image prototype (VERSION_SEFI_IMAGE + dual-time embed block)
fszontagh
committed
fix: SeFi detection match raw checkpoint keys (no model.diffusion_model. prefix)
fszontagh
committed
feat: route SeFi tensors through Flux2 path and instantiate dual_time_embed block
fszontagh
committed
feat: SeFi-Image denoiser + model/conditioner dispatch (Qwen3-VL + Flux2 flow)
fszontagh
committed
feat: SeFi-Image VAE BN normalization on packed texture latents
fszontagh
committed
feat: SeFi Flux forward routes vec through dual_time_embed with delta_t-shifted semantic timestep
fszontagh
committed
feat: Python conversion script for SeFi-Image HF -> sd.cpp single-file format
fszontagh
committed
feat: SeFi-Image prototype loads end-to-end up to text encoder embedding mismatch
fszontagh
committed
feat: end-to-end SeFi-Image generation works (Qwen3-VL-2B head_count + VAE semantic slice)
fszontagh
committed
feat: SeFi dual-time schedule with α-shift + delta_t and 2-element timestep tensor
fszontagh
committed
feat: SeFi per-stream Euler sampler with proper dual-time integration
fszontagh
committed
fix: SEFIDualTimestepEmbeddings — timesteps already in [0,1000], drop 1000x rescale
fszontagh
committed
fix: SeFi LastLayer adaLN uses diffusers (scale,shift) order, not BFL (shift,scale)
fszontagh
committed
fix: SeFi dual-time slot 0 = texture timestep, slot 1 = semantic (empirical)
fszontagh
committed

Commits on Jun 25, 2026

fix(sefi): split conversion into transformer + VAE files (loads via --diffusion-model + --vae like krea2/flux2)
fszontagh
committed