feat: SeFi-Image support

[fszontagh]

Summary

Adds inference support for SeFi-Image, a dual-time flow-matching T2I family built on the Flux2 backbone with a Qwen3-VL text encoder. Tech report: arXiv:2606.22568. See docs/sefi_image.md.

What's in:

• VERSION_SEFI_IMAGE + version detection
• Dual-time embedding block (semantic_embedder + texture_embedder, concat)
• Per-stream Euler sampler with alpha-shift + delta_t
• SeFi-aware Qwen3-VL conditioning (chat template, layers 9/18/27)
• VAE BN normalization on packed texture latents
• script/convert_sefi.py for converting diffusers checkpoint to single sd.cpp safetensors
• --extra-sample-args sefi_alpha=0.3 / sefi_delta_t=0.1 overrides
• Filename heuristic: turbo in path => alpha=1.0, else alpha=0.3

Related Issue / Discussion

Closes #1702.

Additional Information

Example

./build/bin/sd-cli \
  --model /path/to/sefi_1b_turbo.safetensors \
  --llm   /path/to/qwen3_vl_2b.safetensors \
  -p "a photograph of an orange tabby cat sitting on a couch" \
  --cfg-scale 1.0 --steps 4 -W 1024 -H 1024 -s 42 \
  --diffusion-fa --offload-to-cpu \
  -o out.png

Tested variants (all 7 from huggingface.co/SeFi-Image)

Variant	Encoder	Baseline (12GB VRAM)	--max-vram 8 --stream-layers
1B-Base	qwen3_vl_2b	ok 109s	ok 172s
1B-turbo	qwen3_vl_2b	ok 14s	ok 17s
2B-Base	qwen3_vl_2b	ok 229s	ok 296s
2B-turbo	qwen3_vl_2b	ok 29s	ok 25s
5B-Base	qwen3_vl_4b	OOM	ok 563s
5B-turbo	qwen3_vl_4b	OOM	ok 170s
5B-RL	qwen3_vl_4b	OOM	ok 587s

5B variants use Qwen3-VL-4B-Instruct as the text encoder (1B/2B use 2B). 5B needs streaming on 12GB-class GPUs.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.

[GreenShadows]

The quality seems surprisingly good for such a small model.

[sz1kormar]

Speeds are amazing on turbo given that you only need 4 steps and 1 cfg to pull it off. @fszontagh could you attach some of your images here to show off SeFi-Image as an independent tester ?

[fszontagh]

Speeds are amazing on turbo given that you only need 4 steps and 1 cfg to pull it off. @fszontagh could you attach some of your images here to show off SeFi-Image as an independent tester ?

I started the changes which are required by leejet. After i will drop some images here.

[JohnLoveJoy]

It looks really good. These are images I got from Reddit.

[fszontagh]

./build/bin/sd-cli \
  --diffusion-model /data/SD_MODELS/diffusion_models/sefi/sefi_5b_turbo.safetensors \
  --vae /data/SD_MODELS/vae/flux2_ae_from_sefi.safetensors \
  --llm /data/SD_MODELS/Text-encoder/sefi/qwen3_vl_4b.safetensors \
  -p "a photograph of an orange tabby cat sitting on a couch" \
  --cfg-scale 1.0 --steps 4 -W 1024 -H 1024 -s 42 \
  --extra-sample-args sefi_alpha=1.0 \
  --diffusion-fa --max-vram 8 --stream-layers --offload-to-cpu \
  -o /tmp/sefi_refactored.png

[fszontagh]

Pushed the rework. Per-comment:

• auto_encoder_kl.hpp:516 (bn) - confirmed SeFi uses the standard Flux2 VAE. The bn.running_mean / bn.running_var weights match the hardcoded get_latents_mean_std constants within bf16. Dropped the SeFi-specific bn.* params, sefi_bn_apply, and the encode/decode branches. SeFi now uses the Flux2 VAE file directly (flux2_ae.safetensors); convert_sefi.py no longer emits a VAE file. diffusion_to_vae_latents still slices the 16 semantic channels before the standard Flux2 denormalize.
• stable-diffusion.cpp:1341 (denoiser fork) - added SEFI_FLOW_PRED to the prediction_t enum. SeFi version maps to it. FLUX2_FLOW_PRED case is now version-agnostic; SEFI_FLOW_PRED case constructs SefiFlowDenoiser.
• stable-diffusion.cpp:1348 (turbo knob) - removed the filename heuristic. Default timestep_shift_alpha = 1.0 (identity); base/RL pass --extra-sample-args sefi_alpha=0.3. No kAlphaTurbo / kAlphaBase constants left.
• stable-diffusion.cpp:2084 (dual-time override) - moved into process_timesteps. process_timesteps now takes a step arg; the SeFi branch returns {sem_timesteps[step], tex_timesteps[step]}. Sample loop is back to a single process_timesteps call.
• README.md:18 (Important news) - removed the SeFi-Image line.
• name_conversion.cpp:1210 (prefixes) - removed the backbone. / dual_time_embed. prefix injection. convert_sefi.py already emits canonical model.diffusion_model.* keys.

Smoke matrix after rework: 5B-turbo (82 s), 5B-Base (510 s), 5B-RL (567 s) all produce the same orange tabby as before. Within ~3% of pre-rework wallclock.

[fszontagh]

5b rl

5b base

a lovely cat holding a sign says 'SeFi.cpp' with 5b turbo: (~148s)

 ./build/bin/sd-cli \
    --diffusion-model /data/SD_MODELS/diffusion_models/sefi/sefi_5b_turbo.safetensors \
    --vae /data/SD_MODELS/vae/flux2_ae_from_sefi.safetensors \
    --llm /data/SD_MODELS/Text-encoder/sefi/qwen3_vl_4b.safetensors \
    -p "a lovely cat holding a sign says 'SeFi.cpp'" \
    --cfg-scale 1.0 --steps 4 -W 1024 -H 1024 -s 42 \
    --extra-sample-args sefi_alpha=1.0 \
    --diffusion-fa --max-vram 8 --stream-layers --offload-to-cpu \
    -o /tmp/sefi_sign.png

a lovely cat holding a sign says 'SeFi.cpp' with 5b base: (~540s)

./build/bin/sd-cli \
  --diffusion-model /data/SD_MODELS/diffusion_models/sefi/sefi_5b_base.safetensors \
  --vae /data/SD_MODELS/vae/flux2_ae_from_sefi.safetensors \
  --llm /data/SD_MODELS/Text-encoder/sefi/qwen3_vl_4b.safetensors \
  -p "a lovely cat holding a sign says 'SeFi.cpp'" \
  --cfg-scale 4.0 --steps 50 -W 1024 -H 1024 -s 42 \
  --extra-sample-args sefi_alpha=0.3 \
  --diffusion-fa --max-vram 8 --stream-layers --offload-to-cpu \
  -o /tmp/sefi_sign_base.png